Make PromptKit profile handling safer and more consistent

.gitignore

@@ -1,3 +1,9 @@
+# build and testing artifacts
+notarius
+notarius-output
+workspace/
+.codebase-memory/
+
 # ---> Go
 # If you prefer the allow list template instead of the deny list, see community template:
 # https://github.com/github/gitignore/blob/main/community/Golang/Go.AllowList.gitignore
@@ -47,7 +53,8 @@ go.work.sum
 .LSOverride
 
 # Icon must end with two \r
-Icon
+Icon
+
 
 # Thumbnails
 ._*
@@ -67,4 +74,3 @@ Icon
 Network Trash Folder
 Temporary Items
 .apdisk
-.apdisk

AGENTS.md

@@ -1,3 +1,2 @@
-Please carefully review the documents in `docs/policy` before making any changes to this repository.
+Please review `docs/development.md` for initial orientation in this repository
- - `architecture.md` provides the canonical high-level architecture policy for this repository.
+and follow its task-specific reading guide.
- - `documentation.md` provides the canonical documentation policy for this repository.

README.md

@@ -1,36 +1,45 @@
 # Notarius
 
-Notarius is a Go CLI for extracting structured artifacts from source material
+Notarius is a Go CLI for turning source material into structured artifacts with
-with explicit, configurable pipeline modules.
+configured extraction pipelines. The implemented D&D workflow reads Seriatim
+transcript JSON and can produce scene descriptions, item and currency events,
+NPC identities, combat turns, NPC interactions, and spell casts.
 
-The current implementation reads Seriatim transcript JSON, chunks the source
+## Quickstart
-units, extracts D&D spell-cast artifacts with an OpenAI-compatible LLM, and
-writes JSON output plus diagnostics for each run.
 
-```sh
+Provide an OpenRouter API key through the environment, then run the maintained
-NOTARIUS_LLM_DEFAULT_BASE_URL=http://127.0.0.1:8080/v1 \
+minimal example:
-NOTARIUS_LLM_DEFAULT_MODEL=your-model \
-go run ./cmd/notarius run dnd-session \
-  --config examples/dnd-spells.config.yml \
-  --input examples/seriatim-minimal-transcript.json
-```
 
-If the provider requires authentication, set
+~~~
-`NOTARIUS_LLM_DEFAULT_API_KEY` in the environment before running the command.
+OPENROUTER_API_KEY=your-api-key \
-Outputs are written under `./notarius-output/<run-id>/` unless `--output-dir`
+  go run ./cmd/notarius run dnd-session \
-is provided.
+    --config examples/dnd-minimal.config.yml \
+    --input examples/seriatim-minimal-transcript.json
+~~~
 
-Useful references:
+The command publishes a JSON output bundle. Its command syntax and exit
+behavior are documented in the [CLI reference](docs/cli.md); configuration,
+credentials, and module selection are owned by the
+[configuration reference](docs/config.md).
 
-- [CLI reference](docs/cli.md)
+For the complete ordered D&D workflow, use
-- [Configuration reference](docs/config.md)
+[the complete configuration](examples/dnd-complete.config.yml) with
-- [Operations](docs/operations.md)
+[its synthetic transcript](examples/dnd-complete-transcript.json). It
-- [Troubleshooting](docs/troubleshooting.md)
+demonstrates all implemented D&D lanes and the supporting campaign references.
-- [Seriatim input contract](docs/integrations/seriatim.md)
+
-- [OpenAI-compatible provider contract](docs/integrations/openai-compatible.md)
+## Documentation
-- [JSON output contract](docs/integrations/json-output.md)
+
-- [D&D spell artifact contract](docs/integrations/dnd-spell-artifacts.md)
+- [CLI reference](docs/cli.md) — commands, flags, output streams, and exits.
-- [Developer workflow](docs/policy/development.md)
+- [Configuration reference](docs/config.md) — configuration files, profiles,
-- [Internal architecture docs](docs/internal/overview.md)
+  validation, and module selection.
-- [Maintained example config](examples/dnd-spells.config.yml)
+- [Operations](docs/operations.md) — output, state, recovery, and debug
-- [Maintained example input](examples/seriatim-minimal-transcript.json)
+  handling.
+- [Integration contracts](docs/integrations/) — Seriatim input and published
+  artifact formats.
+- [Subprocess consumer guide](docs/consumers/subprocess.md) — invoke Notarius
+  from an orchestrator and consume a published result.
+- [Internal overview](docs/internal/overview.md) — implemented component map
+  for maintainers.
+- [Developer guide](docs/development.md) — contributor orientation and
+  validation guidance.
+- [Future work](docs/roadmap/future.md) — unimplemented ideas and priorities.

docs/adr/0001-record-architecture-decisions.md

@@ -0,0 +1,23 @@
+# ADR-0001: Record architecture decisions as ADRs
+
+**Status:** Accepted
+**Date:** 2026-07-13
+
+## Context
+Architectural reasoning made during design (pattern choices, rejected
+alternatives, trigger conditions for revisiting) is lost if only the final
+state is documented.
+
+## Decision
+We keep a living overview in docs/policy/architecture.md describing current
+intended state, and immutable, numbered ADRs (Nygard format) in docs/adr/
+recording each significant decision, its alternatives, and its consequences.
+Changed decisions get a new ADR that marks the old one Superseded.
+
+## Alternatives considered
+- Overview doc only: loses the "why" and the rejected options.
+- arc42 / RFC-style design docs: heavier than warranted for a solo repo.
+
+## Consequences
+Small ongoing writing cost; durable reasoning trail; cheap onboarding for
+future contributors (including future-us).

docs/adr/0002-linear-pipes-and-filters-pipeline.md

@@ -0,0 +1,51 @@
+# ADR-0002: Linear pipes-and-filters pipeline, not a general DAG
+
+**Status:** Accepted
+**Date:** 2026-07-13
+
+## Context
+
+Notarius processes source material through one known workflow:
+
+```text
+input -> chunk -> extract -> merge -> normalize -> output
+```
+
+Input and chunking apply to the source as a whole. Each selected artifact lane
+then performs extract, merge, and normalize, after which output aggregates the
+lane outcomes. Chunk extraction has a natural scatter-gather shape, but no
+current use case requires arbitrary branches, joins, or user-defined stage
+topology.
+
+## Decision
+
+Notarius implements a fixed six-stage pipes-and-filters pipeline. Configuration
+selects implementations for these stages but cannot add stages, reorder them,
+or define an arbitrary graph.
+
+The framework owns stage sequencing and the scatter-gather boundary between
+chunk, extract, and merge. Extract results are handed to merge in deterministic
+source-chunk order regardless of execution strategy. Each artifact lane remains
+logically linear. Output runs after every selected lane has either produced an
+accepted normalized artifact or reached a recorded rejection. A framework
+execution failure aborts the pipeline.
+
+The runner's concrete internal representation and stage-specific scheduling
+policies are implementation details. Concurrency must preserve the pipeline's
+deterministic handoffs, validation behavior, and provenance, and all execution
+strategies must continue to honor context cancellation.
+
+## Alternatives considered
+
+- Build a general DAG engine now. This would support hypothetical branching
+  topologies, but would add scheduling, topology validation, configuration, and
+  state-management complexity without a current consumer. Revisit this choice
+  only when a concrete workflow requires a topology the fixed pipeline cannot
+  express.
+
+## Consequences
+
+The runner, configuration model, and operator mental model remain small. Stage
+ownership stays visible, and general chunking, merging, or normalization cannot
+be hidden inside extractors. A future DAG requirement will require an explicit
+architectural change rather than incremental exceptions to the fixed pipeline.

docs/adr/0003-typed-interfaces-with-two-zone-data-model.md

@@ -0,0 +1,119 @@
+# ADR-0003: Strongly typed stage interfaces with a two-zone data model
+
+**Status:** Accepted
+**Date:** 2026-07-13
+
+## Context
+
+Pipeline stages must exchange source data and extracted artifacts. Universal
+source data has one engine-wide meaning, while extracted artifacts have
+domain-specific shapes. Passing opaque bytes or `any` between all stages would
+make invalid wiring and merge behavior runtime concerns. Requiring JSON at
+every handoff would preserve interoperability but discard useful Go type safety
+while all modules are in-process.
+
+The framework must also support multiple configured artifact domains, durable
+checkpoints, diagnostics, and output encoders without making those consumers
+depend on every domain's Go types.
+
+## Decision
+
+Notarius uses two typed data zones followed by one serialized boundary.
+
+### Source zone
+
+Input and chunk stages use conservative, engine-owned document, segment, chunk,
+and source-reference types. Their exact Go names are implementation details.
+Every segment carries engine-owned source provenance identifying the source
+location from which it was produced. Chunks preserve the ordered provenance of
+their segments.
+
+Source-format-specific fields remain in input modules or explicitly namespaced
+metadata; they do not become framework contracts.
+
+### Domain artifact zone
+
+Each artifact lane has one domain-owned Go artifact type `T`. Its extract,
+merge, normalize, and domain-aware validation implementations use generic,
+strongly typed contracts over the same `T`. Raw JSON, opaque bytes, and `any`
+are not stage-handoff contracts within a lane.
+
+Each registered domain artifact type supplies a codec for `T`. The codec owns:
+
+- stable schema identity and an explicit schema version;
+- JSON serialization and deserialization;
+- the media type and schema metadata required at serialized boundaries; and
+- rejection of data that cannot be represented by the declared artifact
+  schema.
+
+An artifact type's JSON representation is a maintained domain contract.
+Changing it incompatibly requires a new schema version.
+
+Extract, merge, and normalize may change the contents of `T`, but they do not
+change the lane's canonical Go artifact type or artifact schema identity. An
+extractor maps any provider- or prompt-specific response type into `T` before
+returning. A future lane that requires different artifact types at different
+stages requires a new architectural decision.
+
+### Serialized boundary
+
+After normalization, each typed artifact is converted into an engine-owned
+serialized artifact containing bytes, media type, and schema metadata. Output
+aggregation and output encoders consume this type-erased form. Intermediate
+checkpoint and debug encodings do not become stage-handoff contracts.
+
+LLM transport, checkpoints, and opt-in debug recording are also explicit
+serialization boundaries. They may encode or decode a typed artifact through
+its domain codec, but they do not change the in-memory type used between
+extract, merge, normalize, and typed validators. Checkpoint reuse requires a
+compatible schema identity and version.
+
+An LLM structured-response schema is a module transport contract and may differ
+from the domain artifact schema. The calling module owns the response type and
+maps it into the canonical `T`; the artifact codec remains authoritative for
+artifact checkpoints and output serialization.
+
+The framework may use private type-erased adapters to store heterogeneous lane
+registrations and execute configured domains. Such an adapter must assemble a
+type-consistent lane before execution and must not expose `any` or raw payloads
+as module-facing handoffs inside the domain artifact zone.
+
+### Construction and dependencies
+
+Every module operation accepts `context.Context`. Modules receive stable runtime
+collaborators through an injected dependency set at construction time. In
+particular, LLM-using modules receive the application-provided structured LLM
+client and do not construct provider clients or bypass shared scheduling.
+
+The application boundary enforces one configurable global upper bound on
+in-flight LLM calls across all stages, lanes, retries, and validators.
+
+Configuration options are parsed and validated while a module is constructed,
+before that module executes. Per-run data such as source material, references,
+session identity, and lane identity remains operation input rather than a
+construction dependency.
+
+## Alternatives considered
+
+- Pass raw bytes between stages. This maximizes decoupling but moves wiring,
+  parsing, and merge errors to runtime and prevents domain types from being the
+  canonical in-process contract.
+- Require JSON plus schemas at every stage boundary. This is appropriate for an
+  out-of-process boundary, but adds serialization and parsing inside the current
+  in-process pipeline. The stable codec contract preserves this upgrade path if
+  remote plugins are introduced.
+- Use a uniform `Process(any) (any, error)` contract. This simplifies a fully
+  dynamic engine but turns incompatible module composition into type assertions
+  and runtime failures. The fixed topology does not require that tradeoff.
+
+## Consequences
+
+Domain pipelines gain compile-time handoff safety and explicit merge semantics.
+Serialization, schema compatibility, checkpoint decoding, and output erasure
+have named owners. Dynamic registration requires a small erased adapter around
+each typed lane, and generic stage implementations must be instantiated for a
+specific artifact type or behavior rather than manipulating arbitrary JSON.
+
+The engine-owned source model becomes a long-lived contract and must evolve
+conservatively. Domain authors must maintain a codec and versioned schema in
+addition to their Go artifact type.

docs/adr/0004-package-modules-by-domain.md

@@ -0,0 +1,81 @@
+# ADR-0004: Package modules by domain, not by stage
+
+**Status:** Accepted
+**Date:** 2026-07-13
+
+## Context
+
+Module packages can be grouped first by pipeline stage, such as
+`modules/chunk/dnd/scenes`, or first by domain, such as
+`modules/dnd/chunk/scenes`. A domain's extract, merge, normalize, validation,
+schema, prompt, and artifact-codec implementations collaborate around the same
+artifact types and are likely to evolve together.
+
+Go package dependencies also constrain registration. If shared types live in a
+domain root package, that package cannot import child implementation packages
+to register them because the children already import the root types.
+
+## Decision
+
+Production extensions are grouped by domain under:
+
+```text
+internal/modules/<domain>/<stage>/<name>
+```
+
+Shared artifact types live at the domain root, for example
+`internal/modules/dnd/types.go`. Domain-specific validators, prompt fragments,
+schemas, reference helpers, and codecs also live within that domain tree.
+
+Each domain exposes one production registration entry point from a sibling
+registrar package, for example `internal/modules/dnd/register`. The registrar
+may import the domain root and its child implementations; the domain root does
+not import its registrar or child packages. This keeps shared types available
+as `dnd.SpellList` without creating a Go import cycle.
+
+The `generic` tree is a peer extension family for reusable implementations that
+contain no concrete source-format or artifact-domain knowledge. Source-format
+and output-format families, such as Seriatim and JSON output, follow the same
+domain-first organization even when they do not define a type in the
+[domain artifact zone](0003-typed-interfaces-with-two-zone-data-model.md#domain-artifact-zone).
+
+Concrete domain implementation packages do not import another concrete domain.
+Generic extension packages never import concrete domains. A domain registrar
+may import domain-neutral generic extension packages to instantiate a reusable
+strategy for that domain's artifact type; the generic implementation remains
+unaware of the concrete type's domain semantics. Reuse needed directly by a
+domain implementation lives in a domain-neutral framework or helper package,
+not in a peer extension package.
+
+The application composition root may import multiple registrar packages, and
+black-box integration tests may compose multiple domains. Other cross-domain
+reuse occurs through engine contracts and composition-time registration rather
+than concrete peer-domain imports.
+
+A domain registrar owns registration of that domain's modules, validators,
+default validator chains, artifact codecs, schemas, and prompt assets. It does
+not take ownership of application execution or process behavior.
+
+## Alternatives considered
+
+- Group modules by stage. This keeps interchangeable strategies side by side,
+  but scatters a domain's shared artifact model and collaborating extensions
+  across the repository. It is preferable when generic strategy libraries
+  dominate or when the project is primarily a stage-extension framework rather
+  than an application composed from domain suites.
+- Put both shared types and `Register` in the domain root. This gives the
+  shortest import path but creates an import cycle once child implementations
+  import the root artifact types.
+
+## Consequences
+
+The repository layout makes supported domains immediately visible, and adding
+or extracting a domain affects one cohesive subtree. The CLI composition root
+depends on a small set of domain registrars instead of every leaf package.
+
+Package moves must preserve user-visible module and validator keys unless a
+separate compatibility decision changes them. Shared behavior that cannot be
+expressed through framework contracts may need to move into a domain-neutral
+framework package rather than creating a concrete peer-domain import. Registrar
+packages become explicit composition points for instantiating generic typed
+strategies, in addition to registering domain-owned implementations.

docs/adr/0005-cache-canonical-chunk-plans-by-source.md

@@ -0,0 +1,141 @@
+# ADR-0005: Cache one canonical chunk plan per source
+
+**Status:** Accepted
+**Date:** 2026-07-17
+
+## Context
+
+Notarius may run several extraction passes over the same source. A D&D
+transcript, for example, may first produce NPC artifacts and later produce
+spell or combat artifacts, with output from an earlier pass supplied as a
+reference to a later pass.
+
+An LLM-backed chunker may process an entire, potentially large source in one
+expensive request. Recomputing boundaries for every pipeline or pass repeats
+that cost and can make otherwise comparable extraction runs use different
+source partitions. Stable chunk material also gives later extraction requests
+a better opportunity to benefit from provider-side prompt caching.
+
+Chunk boundaries can affect extraction quality. Evidence may span a boundary,
+overlap may produce duplicates, and different partitions may change the context
+available to a model. Merge and normalization should remove structural signs
+of chunking from durable output, but they cannot guarantee recovery of evidence
+that an extractor did not receive.
+
+Notarius therefore needs an explicit policy for choosing between automatically
+applying the latest chunking configuration and preserving one stable partition
+for repeated work on the same source.
+
+## Decision
+
+Notarius assigns one active canonical chunk plan to a source and reuses that
+plan by default across pipelines and invocations.
+
+The canonical source identity is derived from the validated generic source
+document and covers the source-unit identity, order, and content needed to
+interpret plan boundaries. Input-adapter and chunk-producer identities are
+recorded as provenance, but the active-plan lookup does not vary with:
+
+- pipeline identity or selected artifact lanes;
+- the configured chunk module or its options;
+- references;
+- LLM provider, model, profile, prompt, or response schema; or
+- configuration for later pipeline stages.
+
+When an active plan exists, Notarius uses it even if the current pipeline
+configures a different chunk module or different chunk-module settings. The
+configured chunk module generates a plan only when none exists or when the
+operator explicitly requests recomputation.
+
+The framework-owned minimum plan contract is an ordered, non-empty set of
+source-unit ranges. Each range identifies the inclusive start and end unit for
+one chunk. A chunk module may also provide namespaced, domain-specific
+annotations at plan or range scope. Those annotations are stored with the plan
+and passed through the pipeline when present, but they remain optional.
+Downstream stages must not assume that annotations associated with the
+currently configured chunk module are present on a reused plan produced by a
+different module.
+
+The cache stores the plan rather than fully materialized chunks. The framework
+validates a reused plan against the current source and deterministically
+materializes its ranges into chunks. The same source and plan must produce
+byte-stable chunk input for later stages.
+
+Canonical plan storage is a distinct cache surface with an independently
+configurable location. It is not coupled to the roots or lifecycles of
+invocation checkpoints, diagnostics, debug artifacts, or durable output. This
+allows per-user and system-service deployments to apply cache-specific
+ownership, permissions, placement, and cleanup policy without relocating other
+Notarius state.
+
+One mutable active plan is stored under the canonical source identity and
+retains provenance for the module and relevant runtime inputs that produced it.
+Refreshing the active plan atomically replaces that one mutable record; readers
+must observe either the previous complete plan or the replacement complete
+plan, never a partial update.
+The effective plan producer is reported separately from the chunk module
+requested by the current pipeline; reuse must not attribute cached boundaries
+or annotations to a module that did not produce them.
+
+Reuse is enabled by default. Operators can explicitly:
+
+- bypass cached plans for an invocation without changing the active plan; or
+- recompute a plan with the configured chunk module and make it active for
+  later work.
+
+Exact storage layout, configuration fields, CLI syntax, publication mechanics,
+recovery behavior, and diagnostics are implementation and operational
+contracts rather than part of this decision.
+
+## Alternatives considered
+
+- Recompute chunks on every invocation. This always applies the current
+  chunking configuration, but repeats the most expensive stage and weakens
+  provider-side caching and cross-pass comparability.
+- Cache every distinct chunking request by including module options,
+  references, prompts, profiles, and other runtime inputs in its identity. This
+  closely associates a cached result with its producing request, but reduces
+  reuse and permits boundary drift across operationally different passes.
+- Key plans by source plus chunk module and options. This shares plans across
+  pipelines using the same strategy, but changing the configured strategy
+  silently selects a different partition rather than preserving one canonical
+  partition for the source.
+- Require operators to name or supply a plan for every run. Explicit selection
+  is reproducible and may be useful as an advanced operation, but adds friction
+  to the default workflow and does not provide automatic reuse.
+- Store fully materialized chunks. This simplifies loading, but duplicates
+  source content and couples durable state to the current chunk representation
+  rather than the stable boundary decision.
+- Store canonical plans beneath the general workspace root. This would reuse an
+  existing location setting, but it couples a reusable application cache to
+  checkpoint, diagnostic, and debug state that have different ownership,
+  sensitivity, retention, and deployment requirements.
+
+## Consequences
+
+Independent pipelines and passes over the same source use stable boundaries by
+default. This reduces repeated LLM work, improves cross-pass comparability, and
+increases the opportunity for cached provider reads.
+
+The configured chunk module may not execute, and its settings may have no
+effect, when an active plan already exists. Domain-specific annotations reflect
+the plan's original producer and may be absent or differ from those the current
+module would produce. User-visible provenance must make the effective plan
+clear.
+
+A poor or outdated partition remains active until an operator replaces it.
+This can preserve suboptimal context boundaries and affect extraction recall or
+duplication even when merge and normalization hide the partition structure in
+durable output. Stable reuse is an intentional priority over automatically
+incorporating later chunk-strategy changes.
+
+The framework gains a durable minimal chunk-plan contract and deterministic
+materialization responsibility. Chunk modules must separate required boundary
+output from optional annotations, and downstream modules may rely only on the
+minimal boundary contract unless a future decision introduces an explicit plan
+compatibility mechanism.
+
+Operators must configure and secure canonical plan storage independently from
+other workspace state when the per-user default is not appropriate. Removing
+that cache remains recoverable because Notarius can regenerate it from the
+source, but doing so may repeat an expensive LLM operation.

docs/adr/0006-separate-output-cache-and-debug-state.md

@@ -0,0 +1,114 @@
+# ADR-0006: Separate output, cache, and debug state
+
+**Status:** Superseded by [ADR-0007](0007-separate-checkpoint-recording-from-reuse.md)
+**Date:** 2026-07-17
+
+## Context
+
+Notarius currently exposes a workspace as a shared parent for checkpoints,
+debug artifacts, and preferred diagnostics settings. Diagnostics are a second
+inspection surface with their own enablement, directory, retention, and legacy
+configuration. Durable output uses a separate CLI-selected root, while the
+canonical chunk-plan cache introduced by ADR-0005 correctly uses an independent
+cache root.
+
+These concepts reflect implementation history more than operator intent. A user
+must understand differences among workspace state, diagnostics, debug artifacts,
+checkpoints, and chunk plans before deciding where Notarius may write. Some of
+those distinctions are important internally: a redacted run summary has a
+different sensitivity from a trace containing source material, prompts, and
+model responses. They do not require separate public filesystem categories.
+
+Notarius needs a smaller state model that communicates why data exists, how it
+may be treated, and whether it is reconstructible.
+
+## Decision
+
+Notarius exposes three filesystem surfaces: output, cache, and debug. The
+public workspace concept and diagnostics as a separate output surface are
+removed.
+
+### Output
+
+Output is the durable result of a run and the only surface intended for normal
+consumption. It contains the logical files produced by the output stage,
+including the maintained result, manifest, warning, and rejection contracts.
+Output is not cache or inspection state.
+
+### Cache
+
+Cache contains reconstructible state used to avoid repeated work or resume an
+interrupted workflow. Canonical chunk plans and invocation checkpoints are
+distinct cache families with independent identities, compatibility rules,
+enablement policies, locations, and cleanup lifecycles.
+
+ADR-0005 continues to govern canonical chunk-plan selection and reuse. Grouping
+chunk plans and checkpoints under the public cache category does not permit a
+checkpoint to compete with canonical plan reuse or couple their storage roots.
+
+Checkpointing is an invocation policy rather than a prerequisite hidden in
+persistent workspace configuration. An explicit resume invocation may read
+compatible checkpoints and record replacement checkpoint state for work it
+executes. Runs that do not request resume perform no checkpoint I/O.
+
+### Debug
+
+Debug is an explicitly requested per-run inspection bundle intended for
+developers and troubleshooting. It is off by default. When enabled, one bundle
+contains both redacted run summaries and detailed stage and LLM traces. The
+internal distinction between a safe summary and a sensitive trace remains, but
+there is one public enablement and location model.
+
+Debug data is never a cache input and has no automatic retention policy.
+Notarius does not create a debug directory unless debug is requested, and it
+does not automatically delete a requested bundle. Credentials remain redacted
+at every level, while the bundle as a whole is treated as potentially sensitive 
+because traces may contain source, reference, prompt, model-response, and 
+intermediate artifact content.
+
+Concise progress, warnings, and failures continue to use stdout or stderr. A
+run without debug may fail without producing a filesystem inspection record.
+
+Exact configuration fields, CLI flags, default paths, layouts, compatibility
+handling, and migration mechanics are configuration and operational contracts
+rather than part of this decision.
+
+## Alternatives considered
+
+- Keep workspace, diagnostics, checkpoints, debug, and chunk-plan cache as
+  separate public concepts. This preserves compatibility and the current safe
+  default-on failure records, but retains overlapping configuration and asks
+  operators to reason about implementation-specific categories.
+- Keep diagnostics as an always-available redacted operational surface and use
+  debug only for sensitive traces. This distinction is useful for a daemon or
+  managed service with an operational logging contract, but the current CLI can
+  report concise failures on stderr and provide inspection data when explicitly
+  requested.
+- Put all non-output state beneath one physical root. This minimizes path
+  configuration, but couples reconstructible caches to per-run inspection data
+  and couples cache families whose identity, sensitivity, and cleanup policies
+  differ.
+- Treat checkpoints as durable run state rather than cache. This emphasizes
+  resumability, but checkpoints are derived, compatibility-checked data that may
+  be deleted and recomputed. Cache more accurately describes their lifecycle.
+
+## Consequences
+
+The operator model becomes smaller: normal runs produce output and may use
+cache; developers explicitly request debug. Public configuration no longer
+exposes a workspace or overlapping diagnostics and debug systems.
+
+The implementation retains separate collaborators and serializers where their
+security or lifecycle boundaries differ. Redacted summaries remain useful as
+the index to a debug bundle, and chunk plans and checkpoints retain separate
+stores even though both are cache.
+
+Existing configuration, environment variables, flags, examples, and
+documentation require a deliberate compatibility transition. Default-on
+diagnostic directories disappear, so failures without debug are inspectable
+only through stderr and any durable output completed before the failure.
+
+Debug becomes easier to request and substantially more complete, but enabling
+it creates sensitive files that the operator must protect and remove. Cache
+cleanup is recoverable but may repeat expensive work, while deleting output is
+data loss from the user's perspective.

docs/adr/0007-separate-checkpoint-recording-from-reuse.md

@@ -0,0 +1,50 @@
+# ADR-0007: Separate checkpoint recording from reuse
+
+**Status:** Accepted
+**Date:** 2026-07-19
+
+## Context
+
+ADR-0006 made checkpoint I/O conditional on an explicit `--resume` invocation.
+That policy requires an operator to anticipate the need for recovery before a
+run begins. A failed ordinary run cannot reuse completed work because it did not
+record checkpoints.
+
+Recording reconstructible state and authorizing reuse are separate operational
+decisions. Recording consumes storage and retains sensitive derived application
+data, while reuse may change which module operations execute during a run.
+
+## Decision
+
+ADR-0006's separation of output, cache, and debug surfaces remains in effect;
+this decision supersedes only its checkpoint invocation policy.
+
+Checkpoint recording is controlled by an explicit persistent Boolean
+configuration setting and remains disabled by default. When recording is
+enabled, every run records checkpoint transitions and reusable approved stage
+results.
+
+Checkpoint loading remains an invocation policy. Only a run with `--resume`
+loads and reuses compatible completed work. A recording-enabled run without
+`--resume` executes every stage normally and never loads checkpoints. A resume
+request while recording is disabled is rejected.
+
+The existing checkpoint identities, compatibility rules, payload format,
+filesystem root behavior, and pipeline collaborator contracts remain unchanged.
+
+## Alternatives considered
+
+- Continue coupling reads and writes to `--resume`. This is safe by default but
+  prevents recovery unless resume was anticipated on the earlier run.
+- Always record checkpoints. This maximizes recovery but creates potentially
+  sensitive state without explicit operator consent.
+- Add a multi-value recording policy. This preserves the old behavior as an
+  option but adds configuration complexity without a current need.
+
+## Consequences
+
+Operators can opt into recovery-ready runs while keeping checkpoint reuse
+explicit. Enabled successful, rejected, and failed runs may all leave sensitive
+checkpoint state, so operators remain responsible for access and retention.
+Disabled configurations perform no checkpoint I/O, and `--resume` requires the
+operator to enable recording first.

docs/adr/0008-ordered-pipeline-steps.md

@@ -0,0 +1,50 @@
+# ADR-0008: Bounded ordered pipeline steps and explicit artifact references
+
+**Status:** Accepted
+**Date:** 2026-07-21
+
+## Context
+
+Notarius currently models one pipeline-wide input, chunking plan, artifact
+lanes, and output boundary. Some workflows need a deterministic handoff from
+one set of normalized artifacts to a later set of artifacts, such as using
+extracted NPC records while grounding later combat events. The workflow needs
+an explicit topology without turning the pipeline into a general-purpose
+workflow engine.
+
+## Decision
+
+Add an ordered collection of pipeline steps. Each step owns one or more
+artifact lanes, and lanes within a step retain the existing independent
+execution model. The pipeline continues to have one input, chunk plan, output,
+and failure boundary. Steps are barriers: a later step may consume only
+normalized artifacts from an earlier step.
+
+Generated references use an explicit step-and-lane selector. Reference slots
+declare the generated artifact kinds and media types they accept. The resolver
+validates the topology, ordering, lane identity, artifact kind, schema, and
+codec compatibility before execution. External references remain supported as
+path sources, and the legacy top-level artifact map is interpreted as an
+implicit `default` step.
+
+Pipeline-level references may not select generated artifacts. General DAGs,
+branches, loops, conditional execution, joins, and inferred dependencies are
+not part of this model.
+
+## Alternatives considered
+
+- A general DAG would provide more flexibility but would also require a new
+  scheduler, lifecycle model, failure semantics, and provenance model.
+- Separate pipeline runs connected through filesystem paths would lose the
+  static topology and typed compatibility checks.
+- Inferring dependencies from module or lane names would make ordering and
+  configuration errors difficult to detect reliably.
+
+## Consequences
+
+The resolved pipeline has a deterministic, inspectable topology and can
+include it in its identity digest. Configuration validation can reject invalid
+generated bindings before any work begins. Existing single-step profiles keep
+their behavior through the implicit `default` step. Execution handoff and
+multi-step scheduling require follow-up work in the runner and checkpoint
+layers.

docs/adr/0009-minimal-evidence-grounded-extraction-artifacts.md

@@ -0,0 +1,98 @@
+# ADR-0009: Prefer minimal evidence-grounded extraction artifacts
+
+**Status:** Accepted
+**Date:** 2026-07-22
+
+## Context
+
+Notarius is intended to extract structured facts from source material. Several
+early D&D artifacts grew to include descriptive prose, inferred relationships,
+immediate outcomes, summaries, and other enrichment alongside the facts that
+identify an event or entity. Those fields make one model call responsible for
+both extraction and synthesis.
+
+In practice, the richer contracts have produced overlapping or weakly grounded
+fields and have made structurally valid, semantically coherent output harder for
+cost-effective smaller models. They also increase prompt size, validation and
+normalization policy, durable schema surface, downstream coupling, and the
+number of claims whose provenance must be evaluated.
+
+The application needs a consistent rule for deciding what belongs in an
+extractor before redesigning the current D&D spell, NPC, and combat-turn
+contracts or adding new artifact families.
+
+## Decision
+
+An extraction module answers one narrowly stated question and returns the
+smallest durable structured artifact that usefully answers it.
+
+Every model-produced field in an extraction artifact must:
+
+- be necessary to answer the extractor's stated question or serve a known
+  downstream consumer;
+- represent a fact or bounded classification that can be supported directly by
+  cited source ranges;
+- remain independently meaningful without model-generated explanatory prose;
+  and
+- justify the additional prompt, schema, validation, normalization, and
+  compatibility surface it creates.
+
+Source references are required provenance for extracted records. Auxiliary
+references may disambiguate identities or canonical names, but they do not
+establish source facts and are not copied into evidence.
+
+Extraction artifacts do not include narrative summaries, general analysis,
+speculative enrichment, inferred biography or relationships, or redundant
+free-text descriptions by default. When such output has a demonstrated use, it
+belongs in an explicitly named extraction, classification, enrichment, or
+analysis module with its own contract and evidence policy.
+
+Occurrence-level facts are not forced into entity-level attributes. A fact
+that can change between encounters, such as an NPC's role in a scene, belongs
+on an occurrence artifact rather than as one scalar property of a normalized
+NPC registry entry.
+
+Deterministic mapping and normalization may assign application-owned
+identifiers, canonicalize known catalog values, order and deduplicate evidence,
+and collapse records under an explicit identity rule. They must not manufacture
+removed descriptive fields or synthesize missing claims to satisfy an older
+contract.
+
+This is a default design rule, not a prohibition on rich artifacts. A richer
+field is appropriate when its consumer, evidence semantics, and ownership are
+explicit.
+
+## Alternatives considered
+
+- Keep rich schemas and improve prompts or use larger models. This retains
+  potentially convenient prose but does not resolve overlapping field
+  responsibilities, weak provenance, higher cost, or unnecessary downstream
+  coupling.
+- Make enrichment fields optional. This reduces rejection pressure but leaves
+  ambiguous artifact semantics and inconsistent records, and many strict
+  structured-output providers still require nullable placeholders.
+- Keep minimal private LLM schemas while preserving rich durable artifacts.
+  Deterministic code would have to invent, default, or separately derive the
+  missing fields, hiding synthesis behind the extraction boundary.
+- Use one broad session-analysis module. This reduces the number of lanes but
+  couples unrelated facts, schemas, retries, evaluation, and downstream
+  consumers into one model call.
+
+## Consequences
+
+Extraction prompts and response schemas become smaller, more focused, and more
+suitable for lower-cost models. Artifacts carry fewer unsupported claims, and
+their evidence and validation policies become easier to explain and evaluate.
+Independent extractors can evolve, retry, and be consumed without requiring
+unrelated enrichment.
+
+Some descriptive convenience fields will disappear from primary artifacts.
+Consumers that genuinely need them may require a separate module and explicit
+pipeline step. Entity registries may no longer resolve aliases or relationships
+unless a dedicated, evidence-grounded capability supplies them.
+
+Removing durable fields is a schema compatibility change. Each affected
+artifact requires an explicit version and reference policy; private prompt
+changes alone are insufficient. Current-behavior integration and internal
+documentation must change with implementation, while the roadmap owns the
+proposed contract until then.

docs/adr/0010-workload-oriented-llm-profile-defaults.md

@@ -0,0 +1,49 @@
+# ADR-0010: Use workload-oriented LLM profile defaults
+
+**Status:** Accepted
+**Date:** 2026-08-03
+
+## Context
+
+LLM-backed D&D operations share an execution-policy choice, but repeating a
+provider or model-named profile on every module binding ties pipeline structure
+to a deployment decision. Different environments may require different model,
+backend, timeout, or reasoning settings while retaining the same workload.
+
+Notarius also needs a usable default for maintained D&D prompts without making
+an operator profile mandatory. That default must remain owned by the D&D
+family, while generic LLM infrastructure stays unaware of domain-specific
+policy.
+
+## Decision
+
+Pipelines may name one workload-oriented default profile, inherited only by
+selected LLM-backed bindings and validators. Binding-level profile IDs remain
+intentional exceptions, and the run-wide CLI profile override has highest
+precedence.
+
+The D&D family owns an embedded fallback profile named `dnd-extraction`.
+Operators may provide a complete profile with the same ID through a PromptKit
+filesystem source. PromptKit selects the higher-precedence matching definition;
+Notarius does not merge profile documents. Production, development, and local
+deployments can therefore use different execution policy behind one unchanged
+pipeline ID.
+
+## Alternatives considered
+
+- Repeat a model-named profile on every binding. This makes routine deployment
+  policy changes noisy and obscures the shared workload intent.
+- Require every deployment to install a profile file. This adds configuration
+  friction and leaves maintained D&D prompts without an application-owned
+  fallback.
+- Put D&D profile policy in generic LLM infrastructure. This breaks domain
+  ownership and makes generic code depend on one workload.
+
+## Consequences
+
+Pipeline configuration expresses workload intent rather than a specific
+provider or model. Operators can replace the complete execution policy without
+editing bindings, while binding-level and run-wide exceptions remain available.
+Profile changes affect resolved pipeline and checkpoint identity, so they may
+intentionally cause work to be recomputed. The D&D fallback becomes a
+maintained application execution-policy asset.

docs/cli.md

@@ -1,129 +1,165 @@
 # CLI Reference
 
 This is the canonical reference for the implemented Notarius command-line
-interface.
+interface. For the shortest successful run, see the [README](../README.md).
+Configuration fields, discovery rules, and selectable module keys are defined
+in [Configuration](config.md); runtime state and recovery procedures are
+defined in [Operations](operations.md).
 
-## Quick Run
+## Command Summary
 
-```sh
+~~~
-NOTARIUS_LLM_DEFAULT_BASE_URL=http://127.0.0.1:8080/v1 \
-NOTARIUS_LLM_DEFAULT_MODEL=your-model \
-go run ./cmd/notarius run dnd-session \
-  --config examples/dnd-spells.config.yml \
-  --input examples/seriatim-minimal-transcript.json
-```
-
-Set `NOTARIUS_LLM_DEFAULT_API_KEY` if the OpenAI-compatible provider requires
-a bearer token.
-
-## Commands
-
-```text
 notarius help
-notarius run <pipeline-id> --input path/to/source.json [--config path/to/config.yml] [--only lane-a,lane-b]
+notarius run <pipeline-id> --input path/to/source.json [--json] [flags]
-notarius config validate --config path/to/config.yml [--pipeline pipeline-id] [--only lane-a,lane-b]
+notarius config validate [--config path/to/config.yml] [--pipeline pipeline-id] [--only lane-a,lane-b]
-notarius pipelines list --config path/to/config.yml [--json]
+notarius pipelines list [--config path/to/config.yml] [--json]
-```
+~~~
 
-Running `notarius` with no arguments, `notarius help`, `notarius --help`, or
+Running Notarius without arguments, or with **help**, **--help**, or **-h**,
-`notarius -h` prints usage and exits successfully.
+writes the command summary to standard output and exits with status 0.
 
-## `run`
+## run
 
-`notarius run <pipeline-id>` executes a configured pipeline against one input
+~~~
-file.
+notarius run <pipeline-id> --input path/to/source.json [--json] [flags]
+~~~
 
-Flags:
+The **run** command executes the named pipeline for one input file. The
+pipeline ID and **--input** are required.
 
-- `--input path`: required source input file.
+| Flag | Meaning |
-- `--config path`: config file path. If omitted, Notarius checks
+| --- | --- |
-  `NOTARIUS_CONFIG`, then `/usr/local/etc/notarius/config.yml`.
+| **--config path** | Use this configuration file. When omitted, configuration discovery applies; see [Configuration](config.md). |
-- `--only lane-a,lane-b`: run only the named artifact lanes. Values are
+| **--input path** | Source input file to process. Required. |
-  comma-separated and must be non-empty.
+| **--output-dir path** | Override the configured output root for this run. |
-- `--output-dir path`: output root. The run writes to `<path>/<run-id>/`.
+| **--json** | Write the successful run-result receipt as JSON to standard output. |
-  Defaults to `./notarius-output`.
+| **--chunk_cache auto\|bypass\|refresh** | Override chunk-plan cache handling for this run. |
-- `--diagnostics-dir path`: diagnostics work directory override for this
+| **--resume** | Reuse compatible recorded checkpoints when checkpoint recording is enabled. |
-  invocation.
+| **--recompute-step step-id** | With **--resume**, recompute the selected ordered step and its dependent lanes. It cannot be combined with **--only**. |
-- `--llm-profile id`: override every effective module binding to use one LLM
+| **--debug** | Retain a debug bundle for this run. |
-  profile.
+| **--debug-dir path** | Override the debug-bundle root. Requires **--debug**. |
+| **--only lane-a,lane-b** | Run only the selected comma-separated artifact lanes when that selection is valid for the configured pipeline. |
+| **--llm-profile id** | Highest-precedence configured profile for selected LLM-backed bindings and validators; it replaces binding and [pipeline](config.md#pipelines) defaults. |
+| **--session-id id** | Supply a non-empty prompt session identifier to LLM-backed module calls. |
+| **--reasoning-effort value** | Replace the selected PromptKit profile's reasoning effort for every LLM-backed call in this run. The value must be non-empty and the flag may be specified only once. |
+| **--clear-reasoning-effort** | Clear reasoning effort inherited from the selected PromptKit profile for every LLM-backed call in this run. |
+| **--reference selector=path** | Add or replace a file reference binding. Repeatable. |
+| **--without-reference selector** | Remove a configured optional reference binding. Repeatable. |
 
-On success, the command prints the completed pipeline ID, approved and rejected
+**--chunk_cache** accepts only **auto**, **bypass**, or **refresh**.
-artifact counts, and the output directory. If the run completes with warnings,
+**--debug-dir**, **--output-dir**, **--session-id**, and
-the warning count is printed to stderr.
+**--reasoning-effort**, and **--recompute-step** reject explicit empty values.
+**--reasoning-effort** and **--clear-reasoning-effort** are mutually exclusive.
+When neither is present, reasoning effort comes from the selected PromptKit
+profile. These controls apply to the shared run client, including retries and
+LLM-backed validators, and do not modify configuration or profile files.
+Persistent reasoning settings remain a PromptKit profile concern.
+**--recompute-step** requires **--resume**; checkpoint requirements and reuse
+behavior are documented in [Operations](operations.md).
 
-For durable output, diagnostics, retention, and failure inspection, see
+### Reference selectors
-[Operations](operations.md).
 
-The current `run` command requires the resolved pipeline to use exactly one
+Use **--reference** only for a reference slot declared by the selected
-distinct LLM profile after defaults and overrides are applied.
+configured target. The accepted selector forms are:
 
-## `config validate`
+| Form | Target |
+| --- | --- |
+| slot=path | The unique selected target that declares slot. |
+| chunk.slot=path | The chunker. |
+| merge.slot=path | The unique selected merger that declares slot. |
+| lane.slot=path | The unique extractor, merger, or normalizer in lane that declares slot. |
+| lane.extract.slot=path | The extractor in lane. |
+| lane.merge.slot=path | The merger in lane. |
+| lane.normalize.slot=path | The normalizer in lane. |
 
-`notarius config validate` loads and validates configuration.
+**--without-reference** uses the same selector forms without =path. Slot
+names, requiredness, and configured bindings are part of the
+[configuration contract](config.md).
 
-Flags:
+### Run output
 
-- `--config path`: config file path. If omitted, discovery uses
+Without **--json**, standard output contains the completed pipeline ID, counts
-  `NOTARIUS_CONFIG`, then `/usr/local/etc/notarius/config.yml`.
+of normalized and rejected outputs, and the output directory. A debug-enabled
-- `--pipeline pipeline-id`: additionally resolve one configured pipeline against
+run also prints its debug-bundle path to standard output. A successful run with
-  the production module catalog.
+warnings reports the warning count to standard error. The published JSON bundle
-- `--only lane-a,lane-b`: validate resolution for selected artifact lanes. This
+is defined by the [JSON output contract](integrations/json-output.md).
-  flag requires `--pipeline`.
+
+With **--json**, successful standard output is exactly one
+`notarius.run-result.v1` JSON document followed by a newline, with no
+human-oriented status or debug-path line. Its fields and compatibility policy
+are defined by the [run-result contract](integrations/run-result.md). A caller
+must check for exit status 0 before decoding this output; a failed write can
+leave incomplete standard-output bytes that are not a result document.
+
+Example:
+
+~~~
+OPENROUTER_API_KEY=your-api-key \
+  go run ./cmd/notarius run dnd-session \
+    --config examples/dnd-minimal.config.yml \
+    --input examples/seriatim-minimal-transcript.json
+~~~
+
+## config validate
+
+~~~
+notarius config validate [--config path/to/config.yml] [--pipeline pipeline-id] [--only lane-a,lane-b]
+~~~
+
+This command loads and validates a configuration. With **--pipeline**, it also
+resolves that pipeline against the production module catalog. **--only** selects
+lanes during that resolution and requires **--pipeline**.
+
+Success is written to standard output as either config "<path>" is valid or
+config "<path>" is valid for pipeline "<pipeline-id>".
 
 Examples:
 
-```sh
+~~~
 go run ./cmd/notarius config validate \
-  --config examples/dnd-spells.config.yml
+  --config examples/dnd-minimal.config.yml \
+  --pipeline dnd-session
 
-go run ./cmd/notarius config validate \
+OPENROUTER_API_KEY=validation-placeholder \
-  --config examples/dnd-spells.config.yml \
+  go run ./cmd/notarius config validate \
-  --pipeline dnd-session \
+    --config examples/dnd-complete.config.yml \
-  --only spells
+    --pipeline dnd-session
-```
+~~~
 
-## `pipelines list`
+The placeholder in the second command is sufficient only for offline
+validation; it cannot run a provider-backed pipeline.
 
-`notarius pipelines list` prints configured pipeline IDs in sorted order.
+## pipelines list
 
-Flags:
+~~~
+notarius pipelines list [--config path/to/config.yml] [--json]
+~~~
 
-- `--config path`: config file path. If omitted, discovery uses
+This command lists configured pipeline IDs in sorted order. By default, it
-  `NOTARIUS_CONFIG`, then `/usr/local/etc/notarius/config.yml`.
+writes one ID per line to standard output. **--json** writes an object shaped as
-- `--json`: print `{"pipelines":[...]}` instead of one ID per line.
+{"pipelines":[...]} instead.
 
-Examples:
+~~~
-
-```sh
 go run ./cmd/notarius pipelines list \
-  --config examples/dnd-spells.config.yml
+  --config examples/dnd-minimal.config.yml
+~~~
 
-go run ./cmd/notarius pipelines list \
+## Output Streams And Exit Statuses
-  --config examples/dnd-spells.config.yml \
-  --json
-```
 
-## Exit Codes
+Successful commands write their primary result to standard output. Warnings and
+errors are written to standard error.
 
-- `0`: command succeeded.
+For **run --json**, warnings remain on standard error and standard output is a
-- `1`: command syntax was valid, but loading config, resolving modules, running
+machine-readable success result only. Syntax and runtime diagnostics remain on
-  the pipeline, calling the provider, writing output, or writing diagnostics
+standard error. Parse the result only after the process exits with status 0.
-  failed.
-- `2`: command syntax was invalid, a command was unknown, a required argument
-  was missing, or a flag value was malformed.
 
-## Implemented Production Pipeline Modules
+| Status | Meaning |
+| --- | --- |
+| 0 | The command completed successfully, including root help. |
+| 1 | Command syntax was valid but configuration loading or validation, pipeline resolution or execution, provider use, output, or requested debug handling failed. |
+| 2 | The command or flag syntax was invalid, including unknown commands, missing required arguments, invalid flag values, or invalid flag combinations. |
 
-The production CLI currently registers these module keys:
+The root help spellings are the supported help path. Invoking **--help** on
-
+**run**, **config validate**, or **pipelines list** is handled by the flag
-- input: `seriatim`
+parser as a usage error: it writes an error to standard error and exits with
-- chunk: `generic`
+status 2.
-- extract: `dnd/spells`
-- merge: `appendorder`
-- normalize: `noop`
-- output: `json`
-
-The production CLI does not currently register validator modules.
-
-For YAML structure, defaults, environment overrides, and module binding syntax,
-see [Configuration](config.md).

docs/config.md

@@ -1,213 +1,446 @@
 # Configuration
 
-This is the canonical reference for implemented Notarius configuration.
+This is the canonical reference for Notarius configuration. Configuration files
+are YAML and must declare version 4. They select pipelines and their modules;
+the [CLI reference](cli.md) owns invocation syntax, and
+[Operations](operations.md) owns run-state procedures.
 
-Notarius reads YAML config files with `version: 1`. File config is applied over
+## Configuration Discovery And Precedence
-built-in defaults, then environment overrides are applied.
 
-## Discovery
+Commands that load configuration choose a file in this order:
 
-Commands that accept `--config` load configuration in this order:
+1. a non-empty **--config** CLI value;
+2. a non-empty **NOTARIUS_CONFIG** environment value;
+3. the installed default file at **/usr/local/etc/notarius/config.yml**, when
+   it exists.
 
-1. the `--config` path, when provided;
+The command fails if none of these paths provides a configuration file.
-2. `NOTARIUS_CONFIG`, when set to a non-empty path;
-3. `/usr/local/etc/notarius/config.yml`.
 
-If none is available, the command fails with a config file not found error.
+For configuration values, precedence is:
 
-## Minimal Example
+1. built-in defaults;
+2. the selected YAML file;
+3. supported operational environment variables; and
+4. the CLI run overrides that apply to a command.
 
-```yaml
+Environment variables do not provide a second configuration schema. They only
-version: 1
+override the fields listed below.
-llm_profiles:
-  default:
-    provider: openai-compatible
-    base_url: http://127.0.0.1:8080/v1
-    model: your-model
-pipelines:
-  dnd-session:
-    input: seriatim
-    chunk:
-      module: generic
-      options:
-        max_units: 50
-    artifacts:
-      spells:
-        extract: dnd/spells
-```
 
-The maintained fixture is [examples/dnd-spells.config.yml](../examples/dnd-spells.config.yml).
+## Maintained Examples
 
-## Top-Level Fields
+- [Minimal D&D configuration](../examples/dnd-minimal.config.yml) is a
+  single-lane Seriatim-to-spell pipeline.
+- [Complete D&D configuration](../examples/dnd-complete.config.yml) uses
+  ordered steps, all implemented D&D lanes, generated references, state
+  settings, bounded LLM concurrency, and the maintained
+  [operator profile](../examples/profiles/dnd-extraction.yml).
 
-- `version`: required. The only supported value is `1`.
+Use these complete files as starting points rather than combining the
-- `llm_profiles`: optional map of LLM profile IDs to profile settings.
+illustrative fragments in this reference.
-- `pipelines`: optional map of pipeline IDs to pipeline definitions.
-- `concurrency`: optional global concurrency settings.
-- `diagnostics`: optional diagnostics settings.
 
-Unknown YAML fields are rejected.
+## File Shape And Defaults
 
-## Defaults
+Unknown fields, duplicate mapping keys, empty identifiers, and identifiers that
+become duplicates after trimming whitespace are rejected. Every top-level field
+other than **version** is optional.
 
-Built-in defaults:
+| Field | Type | Default | Rules |
+| --- | --- | --- | --- |
+| **version** | integer | none | Required; must be 4. |
+| **promptkit** | object | none | Profile source and optional local-backend configuration. |
+| **pipelines** | map | empty | Maps pipeline IDs to pipeline definitions. |
+| **concurrency** | object | see below | Global LLM and extraction limits. |
+| **output** | object | see below | Published output settings. |
+| **cache** | object | see below | Chunk-plan and checkpoint settings. |
+| **debug** | object | see below | Debug-bundle root only; it does not enable capture. |
 
-```yaml
+Built-in defaults are:
-llm_profiles:
+
-  default:
+| Field | Default |
-    provider: openai-compatible
+| --- | --- |
-    timeout: 600
+| **concurrency.total_llm** | 1 |
-    max_retries: 3
+| **concurrency.stage_workers.extract** | Effective **total_llm** |
-    max_concurrency: 1
+| **output.directory** | **./notarius-output** |
+| **cache.chunk_plans.mode** | **auto** |
+| **cache.chunk_plans.directory** | Empty, selecting the per-user chunk-plan root |
+| **cache.checkpoints.enabled** | false |
+| **cache.checkpoints.directory** | Empty, selecting the per-user checkpoint root |
+| **debug.directory** | **./notarius-debug** |
+
+An empty cache directory in YAML deliberately selects the corresponding
+per-user root. An explicit empty output or debug directory is invalid.
+
+## PromptKit Profiles
+
+The optional **promptkit** object selects one source of profile definitions and
+may register one conventional local OpenAI-compatible backend:
+
+~~~yaml
+version: 4
+
+promptkit:
+  profile_dir: ./profiles
+  # profile_file: ./profiles.yml
+  local_backend:
+    endpoint: http://localhost:8000/v1
+    concurrency_limit: 2
+~~~
+
+| Field | Type | Rules |
+| --- | --- | --- |
+| **profile_dir** | string | Non-empty directory containing profile files. |
+| **profile_file** | string | Non-empty profile file. |
+| **local_backend** | object | Optional registration for the conventional PromptKit backend ID **local**. |
+| **local_backend.endpoint** | string | Required when **local_backend** is present; absolute HTTP or HTTPS URL with a host. |
+| **local_backend.concurrency_limit** | integer | Optional non-negative limit; defaults to 0. |
+
+Set at most one of **profile_dir** and **profile_file**. Relative values use
+the process working directory, not the configuration file's directory. The
+complete example's `./examples/profiles/dnd-extraction.yml` value is therefore
+valid when Notarius is launched from the repository root; use an absolute path
+for services and containers.
+
+An operator source is optional. For a requested ID, PromptKit checks the
+configured operator source first, then Notarius's embedded fallback profiles,
+then its own built-in catalog. A matching profile is complete: it replaces a
+lower-precedence definition rather than merging with it. The maintained
+[`dnd-extraction` operator profile](../examples/profiles/dnd-extraction.yml)
+is a secret-free deployment artifact; production, development, and local
+deployments can each provide a complete definition with that same workload ID.
+Use workload-oriented IDs for new profiles instead of model names.
+[Operations](operations.md#promptkit-profile-deployment) owns the deployment
+workflow and credential-handling guidance.
+
+When **local_backend** is present, its endpoint is trimmed and must use HTTP or
+HTTPS case-insensitively, be absolute, and have a non-empty host. URL paths are
+allowed. User information, queries, and fragments are rejected. A zero
+**concurrency_limit** leaves the local backend unrestricted inside PromptKit;
+a positive value limits simultaneous local generations. The application-wide
+**concurrency.total_llm** limit still applies in both cases. Neither local
+backend field has an environment override. Omitting **local_backend** registers
+nothing and preserves existing built-in and endpoint-only profile behavior.
+
+A file-backed PromptKit profile selects the registration by its case-sensitive
+backend ID:
+
+~~~yaml
+id: local-summary
+backend: local
+model: example-model
+~~~
+
+Keep credentials out of the local-backend object. A PromptKit profile may name
+its credential environment variable through `api_key_env`; set that variable
+only in the run environment. PromptKit owns the
+[pinned profile-file format](https://gitea.maximumdirect.net/eric/promptkit/src/tag/v0.5.0/docs/formats.md).
+The [PromptKit upstream boundary](integrations/pkg-promptkit.md) identifies the
+supported package API, and [Operations](operations.md#operational-limits)
+describes the effective concurrency layers.
+
+`notarius config validate --pipeline <id>` resolves the selected pipeline and
+inspects every explicit effective profile without contacting a provider or
+requiring credential values. It rejects absent, malformed, or incompatible
+profiles before a run prepares modules. Credential availability is checked only
+when a generation is prepared.
+
+## Migrating Version 3 Configuration
+
+Version 3 files are not decoded or rewritten. Change **version: 3** to
+**version: 4** and rename the top-level **scriptorium:** section to
+**promptkit:**. Version 4 decoding is strict, so a remaining **scriptorium**
+field is rejected as unknown.
+
+## Operational Environment Variables
+
+These variables are applied after YAML values:
+
+| Variable | Overrides | Rules |
+| --- | --- | --- |
+| **NOTARIUS_TOTAL_LLM_CONCURRENCY** | **concurrency.total_llm** | Integer. |
+| **NOTARIUS_STAGE_WORKERS_EXTRACT** | **concurrency.stage_workers.extract** | Integer. |
+| **NOTARIUS_OUTPUT_DIR** | **output.directory** | Non-empty path. |
+| **NOTARIUS_CACHE_CHUNK_PLANS_MODE** | **cache.chunk_plans.mode** | **auto**, **bypass**, or **refresh**. |
+| **NOTARIUS_CACHE_CHUNK_PLANS_DIR** | **cache.chunk_plans.directory** | Non-empty path. |
+| **NOTARIUS_CACHE_CHECKPOINTS_DIR** | **cache.checkpoints.directory** | Non-empty path. |
+| **NOTARIUS_DEBUG_DIR** | **debug.directory** | Non-empty path. |
+
+Integer values are trimmed then parsed as base-10 integers. Directory and
+output values reject NUL characters. **NOTARIUS_CONFIG** participates only in
+configuration discovery.
+
+## Concurrency, Output, Cache, And Debug
+
+~~~yaml
 concurrency:
-  total_llm: 1
+  total_llm: 2
-diagnostics:
+  stage_workers:
-  work_dir: /tmp/notarius
+    extract: 2
-  retention: auto
+output:
-```
+  directory: ./notarius-output
+cache:
+  chunk_plans:
+    mode: auto
+    directory: ./notarius-cache/chunk-plans
+  checkpoints:
+    enabled: true
+    directory: ./notarius-cache/checkpoints
+debug:
+  directory: ./notarius-debug
+~~~
 
-No pipelines are built in. A run requires a configured pipeline.
+**concurrency.total_llm** must be greater than zero. The only supported
+**concurrency.stage_workers** key is **extract**; its value must be from 1
+through **total_llm**. When omitted, it is recalculated from the effective
+**total_llm** after YAML and environment precedence.
 
-## LLM Profiles
+**cache.chunk_plans.mode** accepts **auto**, **bypass**, or **refresh**.
-
+**cache.checkpoints.enabled** is a boolean. The CLI can override the output
-Each `llm_profiles` entry may contain:
+directory and chunk-plan mode for one run; see [CLI reference](cli.md#run).
-
-- `provider`: optional provider key. Empty means `openai-compatible`; any other
-  non-empty value must be `openai-compatible`.
-- `base_url`: provider base URL. Required for actual LLM calls.
-- `model`: provider model name. Required for actual LLM calls.
-- `api_key_env`: environment variable name to read for the API key.
-- `timeout`: request timeout as whole seconds or a Go-style duration string such
-  as `10m`.
-- `max_retries`: retry count for provider calls. Must be zero or greater.
-- `max_concurrency`: per-profile LLM concurrency. Must be zero or greater; when
-  zero, Notarius uses `concurrency.total_llm`.
-
-Raw API keys are not accepted as file config fields. Use `api_key_env` or an
-environment override.
-
-## Environment Overrides
-
-These environment variables are applied after the config file:
-
-- `NOTARIUS_CONFIG`: config discovery path.
-- `NOTARIUS_LLM_DEFAULT_API_KEY`: API key for the `default` LLM profile.
-- `NOTARIUS_LLM_DEFAULT_BASE_URL`: base URL for the `default` LLM profile.
-- `NOTARIUS_LLM_DEFAULT_MODEL`: model for the `default` LLM profile.
-- `NOTARIUS_LLM_DEFAULT_TIMEOUT_SECONDS`: integer timeout seconds for the
-  `default` LLM profile.
-- `NOTARIUS_LLM_DEFAULT_MAX_RETRIES`: integer retry count for the `default` LLM
-  profile.
-- `NOTARIUS_LLM_DEFAULT_MAX_CONCURRENCY`: integer max concurrency for the
-  `default` LLM profile.
-- `NOTARIUS_TOTAL_LLM_CONCURRENCY`: integer global LLM concurrency.
-- `NOTARIUS_WORK_DIR`: diagnostics work directory.
-- `NOTARIUS_DIAGNOSTICS_RETENTION`: diagnostics retention mode.
-
-Integer environment values must parse as base-10 integers.
 
 ## Pipelines
 
-A pipeline defines the fixed Notarius workflow:
+Each **pipelines** entry has a unique, non-empty ID and the following shape:
 
-```text
+~~~yaml
-input -> chunk -> extract -> merge -> normalize -> output
+pipelines:
-```
+  dnd-session:
+    llm_profile: dnd-extraction
+    input: seriatim
+    chunk: generic
+    output: json
+    artifacts:
+      spells:
+        extract: dnd/spells
+        merge: appendorder
+        normalize: dnd/spells
+~~~
 
-Pipeline fields:
+| Field | Type | Default | Rules |
+| --- | --- | --- | --- |
+| **llm_profile** | string | none | Optional non-empty default PromptKit profile ID for selected LLM-backed bindings and validators. An explicitly present blank value is invalid. |
+| **input** | module binding | none | Required. |
+| **chunk** | module binding | **generic** | Optional. |
+| **output** | module binding | **json** | Optional. |
+| **artifacts** | map | none | Compact single-step lane map. |
+| **steps** | list | none | Ordered lane definitions. Mutually exclusive with **artifacts**. |
+| **references** | map | none | External reference defaults for eligible targets. |
 
-- `input`: required module binding.
+Use either **artifacts** or **steps**. The compact **artifacts** form is an
-- `chunk`: optional module binding. Default module is `generic`.
+implicit single step. An explicit **steps** list must be non-empty; every step
-- `artifacts`: required for pipeline resolution. It maps artifact lane IDs to
+needs a unique non-empty **id**, an **artifacts** map, and may have
-  lane definitions.
+**references**. A lane ID must not appear more than once in a pipeline,
-- `output`: optional module binding. Default module is `json`.
+including across explicit steps.
 
-Artifact lane fields:
+For each selected LLM-backed binding or validator, profile selection occurs
+after module, validator, and `--only` lane selection. It uses the
+run-level **--llm-profile** value first, then the binding's **llm_profile**,
+then the pipeline's **llm_profile**, and finally the PromptKit default.
+Deterministic bindings do not receive these defaults or run overrides.
 
-- `extract`: required module binding.
+A lane has these fields:
-- `merge`: optional module binding. Default module is `appendorder`.
-- `normalize`: optional module binding. Default module is `noop`.
-- `validators`: optional list of module bindings. The production CLI currently
-  does not register validator modules.
 
-`notarius run` and `notarius config validate --pipeline` resolve the pipeline
+| Field | Type | Default | Rules |
-against the production module catalog and fail fast for unknown or incompatible
+| --- | --- | --- | --- |
-module keys.
+| **extract** | module binding | none | Required. |
+| **merge** | module binding | **appendorder** | Optional. |
+| **normalize** | module binding | **noop** | Optional. |
+| **references** | map | none | Supported compatibility alias for **extract.references**. |
+| **validators** | list | none | Non-empty lane-level lists are rejected. Set validator overrides on a binding instead. |
 
-## Module Bindings
+The lane-level **references** alias remains accepted. When the alias and
+**extract.references** bind the same slot, **extract.references** wins. Use
+the binding-local form in new configurations.
 
-Every module binding may use shorthand:
+## Module Bindings And Validators
 
-```yaml
+Use a module key directly when no other binding fields are needed:
+
+~~~yaml
 input: seriatim
-```
+~~~
 
-or object form:
+Use an object for fields:
 
-```yaml
+~~~yaml
-chunk:
+extract:
-  module: generic
+  module: dnd/spells
-  llm_profile: default
+  llm_profile: dnd-extraction
+  retries: 2
+  references:
+    spell_catalog: ./dnd-spell-catalog.json
+~~~
+
+| Binding field | Type | Default | Rules |
+| --- | --- | --- | --- |
+| **module** | string | none | Required for an object binding. Must be a registered compatible key. |
+| **llm_profile** | string | none | Optional non-empty PromptKit profile ID for an LLM-backed binding. It overrides the pipeline default unless the run supplies **--llm-profile**. |
+| **retries** | integer | 0 | Non-negative additional attempts for chunk, extract, merge, and normalize bindings. |
+| **options** | object | none | Must satisfy the selected module. |
+| **references** | map | none | Valid only on chunk, extract, merge, and normalize bindings. |
+| **validators** | list | production chain | Valid only on chunk, extract, merge, and normalize bindings. |
+
+Omitting **validators** uses the registered chain. **validators: []** selects
+an empty chain; a non-empty list replaces the chain in the listed order.
+Validator bindings accept only **module**, **llm_profile**, and **options**.
+They reject **references**, **retries**, and nested **validators**. Deterministic
+validators reject an explicit **llm_profile**. Deterministic module bindings
+also reject an explicit **llm_profile**.
+
+The **json** output module accepts optional **include_chunk_map** and
+**evidence_context** settings:
+
+~~~yaml
+output:
+  module: json
   options:
-    max_units: 50
+    include_chunk_map: true
-```
+    evidence_context:
+      enabled: true
+      window_units: 3
+      lanes:
+        - npcs
+        - spells
+~~~
 
-Binding fields:
+**include_chunk_map** is a boolean and defaults to false. It adds the accepted
+chunk map when one exists; its wire format is defined in the
+[chunk-map contract](integrations/chunk-map.md).
 
-- `module`: module key.
+Omitting **evidence_context** disables evidence publication. When present, it
-- `llm_profile`: optional LLM profile ID. Empty means `default`.
+is an object with these strict fields:
-- `options`: optional module-specific settings.
 
-The `--llm-profile` run flag overrides every effective module binding to use
+| Field | Type | Rules |
-one configured profile.
-
-## Implemented Production Modules
-
-| Slot | Key | Notes |
 | --- | --- | --- |
-| input | `seriatim` | Reads Seriatim transcript JSON. |
+| **enabled** | boolean | Required. `false` permits no other evidence fields. |
-| chunk | `generic` | Splits source units into ordered chunks. |
+| **lanes** | array of strings | Required and non-empty when enabled. Each value is trimmed and must be unique; every value must name a configured pipeline lane. |
-| extract | `dnd/spells` | Extracts `dnd.spell_cast` artifacts. |
+| **window_units** | non-negative integer | Optional when enabled; defaults to 3. Zero retains only directly cited units. |
-| merge | `appendorder` | Keeps candidates in append order. |
-| normalize | `noop` | Passes merged artifacts through unchanged. |
-| output | `json` | Produces JSON output files. |
 
-The `generic` chunker accepts:
+Unknown outer or nested option fields are rejected, as are incompatible YAML
+types. The allowlist remains valid when a run uses lane filtering: a configured
+lane that is not active for that invocation simply contributes no evidence.
+Evidence publication is opt-in because it can persist source text and metadata.
+Its payload contract is [Published Evidence Context](integrations/evidence-context.md).
 
-- `max_units`: positive integer, default `50`;
+## References And Ordered Handoffs
-- `overlap_units`: non-negative integer, default `0`, and must be less than
-  `max_units`.
 
-## Diagnostics
+Reference maps bind named slots that the selected target declares. A scalar is
+an external path. Pipeline-level maps accept only external paths; step-local
+and binding-local maps may also select a normalized artifact from an earlier
+step:
 
-`diagnostics` fields:
+~~~yaml
+steps:
+  - id: describe-session
+    artifacts:
+      npcs:
+        extract: dnd/npcs
+        normalize: dnd/npcs
+  - id: extract-events
+    references:
+      npcs:
+        artifact:
+          step: describe-session
+          lane: npcs
+    artifacts:
+      spells:
+        extract: dnd/spells
+        normalize: dnd/spells
+~~~
 
-- `work_dir`: directory for per-run diagnostics. Default: `/tmp/notarius`.
+An artifact selector contains only **step** and **lane**. The producer must be
-- `retention`: `auto`, `always`, or `never`. Empty uses `auto`.
+an earlier step and the selected artifact must be compatible with the consumer
+slot. A generated binding supplies one accepted normalized artifact; it does
+not name a file. A configured generated dependency remains required even when
+that consumer slot is otherwise optional.
 
-`auto` retains diagnostics for failed runs and successful runs with warnings.
+Pipeline references are defaults. A matching step-local or binding-local
-`always` retains diagnostics for every run. `never` removes diagnostics for
+external path overrides a pipeline default. Required slots must be bound after
-successful runs without regard to warnings; failed runs are retained.
+these configuration values and any CLI reference overrides are applied.
+Reference paths in YAML are resolved relative to the configuration file.
 
-The `--diagnostics-dir` run flag overrides `diagnostics.work_dir` for that
+### D&D Reference Slots
-invocation.
+
+The following slot names are accepted by the implemented D&D modules when the
+selected target declares them:
+
+| Slot | Source and use |
+| --- | --- |
+| **party** | Optional text campaign context. This is the canonical party-roster spelling. |
+| **roster** | Accepted compatibility alias for **party**. |
+| **players** | Optional text player context. |
+| **glossary** | Optional text campaign glossary. |
+| **spell_catalog** | Optional JSON spell-catalog overlay for spell extraction and normalization. See [spell-catalog overlays](integrations/dnd-spell-catalog-overlays.md). |
+| **npcs** | Normalized NPC registry. Optional for spells and combat turns; required for NPC interactions. |
+| **scene_descriptions** | Required normalized scene-description artifact for combat-turn extraction. |
+
+Scene descriptions accept **party**, **players**, and **glossary**, but not
+**roster**. NPC interactions require **npcs** for both extraction and
+normalization. Combat turns require **scene_descriptions** for extraction; the
+normalized combat-turn module may use optional **npcs**. The complete example
+shows generated **npcs** and **scene_descriptions** bindings.
+
+## Production Module Keys
+
+| Kind | Keys |
+| --- | --- |
+| Input | **seriatim** |
+| Chunk | **generic**, **dnd/scenes** |
+| Extract | **dnd/spells**, **dnd/npcs**, **dnd/combat-turns**, **dnd/item-events**, **dnd/npc-interactions**, **dnd/scene-descriptions** |
+| Merge | **appendorder** |
+| Normalize | **noop**, **dnd/spells**, **dnd/npcs**, **dnd/combat-turns**, **dnd/item-events**, **dnd/npc-interactions**, **dnd/scene-descriptions** |
+| Output | **json** |
+
+The D&D artifact contracts define each emitted schema:
+[spells](integrations/dnd-spell-artifacts.md),
+[NPCs](integrations/dnd-npc-artifacts.md),
+[NPC interactions](integrations/dnd-npc-interaction-artifacts.md),
+[combat turns](integrations/dnd-combat-turn-artifacts.md),
+[item events](integrations/dnd-item-event-artifacts.md), and
+[scene descriptions](integrations/dnd-scene-description-artifacts.md).
+
+## Production Validator Keys And Default Chains
+
+Available validator keys are:
+
+| Family | Keys |
+| --- | --- |
+| Generic | **generic/always_accept**, **generic/always_reject**, **generic/valid_json**, **generic/valid_json_schema** |
+| Spells | **extract/dnd/spells/shape**, **extract/dnd/spells/catalog**, **extract/dnd/spells/source_refs**, **extract/dnd/spells/source_relatedness** |
+| NPCs | **extract/dnd/npcs/shape**, **extract/dnd/npcs/source_refs**, **extract/dnd/npcs/source_relatedness**, **normalize/dnd/npcs/identity** |
+| Combat turns | **extract/dnd/combat-turns/shape**, **extract/dnd/combat-turns/source_refs**, **extract/dnd/combat-turns/source_relatedness**, **normalize/dnd/combat-turns/invariants** |
+| Item events | **extract/dnd/item-events/shape**, **extract/dnd/item-events/source_refs**, **extract/dnd/item-events/source_relatedness**, **normalize/dnd/item-events/invariants** |
+| NPC interactions | **extract/dnd/npc-interactions/shape**, **extract/dnd/npc-interactions/registry**, **extract/dnd/npc-interactions/source_refs**, **extract/dnd/npc-interactions/source_relatedness**, **normalize/dnd/npc-interactions/invariants** |
+| Scene descriptions | **extract/dnd/scene-descriptions/shape**, **extract/dnd/scene-descriptions/source_refs**, **extract/dnd/scene-descriptions/source_relatedness**, **normalize/dnd/scene-descriptions/invariants** |
+
+When no override is configured, production D&D bindings use the following
+ordered chains. Each row lists extract then normalize; spell chains are the
+same at both stages.
+
+| Lane | Extract | Normalize |
+| --- | --- | --- |
+| Spells | generic/valid_json, extract/dnd/spells/shape, extract/dnd/spells/catalog, extract/dnd/spells/source_refs, generic/valid_json_schema, extract/dnd/spells/source_relatedness | Same as extract |
+| NPCs | generic/valid_json, extract/dnd/npcs/shape, extract/dnd/npcs/source_refs, generic/valid_json_schema, extract/dnd/npcs/source_relatedness | generic/valid_json, extract/dnd/npcs/shape, normalize/dnd/npcs/identity, extract/dnd/npcs/source_refs, generic/valid_json_schema, extract/dnd/npcs/source_relatedness |
+| Combat turns | generic/valid_json, extract/dnd/combat-turns/shape, extract/dnd/combat-turns/source_refs, generic/valid_json_schema, extract/dnd/combat-turns/source_relatedness | generic/valid_json, extract/dnd/combat-turns/shape, normalize/dnd/combat-turns/invariants, extract/dnd/combat-turns/source_refs, generic/valid_json_schema, extract/dnd/combat-turns/source_relatedness |
+| Item events | generic/valid_json, extract/dnd/item-events/shape, extract/dnd/item-events/source_refs, generic/valid_json_schema, extract/dnd/item-events/source_relatedness | generic/valid_json, extract/dnd/item-events/shape, normalize/dnd/item-events/invariants, extract/dnd/item-events/source_refs, generic/valid_json_schema, extract/dnd/item-events/source_relatedness |
+| NPC interactions | generic/valid_json, extract/dnd/npc-interactions/shape, extract/dnd/npc-interactions/registry, extract/dnd/npc-interactions/source_refs, generic/valid_json_schema, extract/dnd/npc-interactions/source_relatedness | generic/valid_json, extract/dnd/npc-interactions/shape, extract/dnd/npc-interactions/registry, normalize/dnd/npc-interactions/invariants, extract/dnd/npc-interactions/source_refs, generic/valid_json_schema, extract/dnd/npc-interactions/source_relatedness |
+| Scene descriptions | generic/valid_json, extract/dnd/scene-descriptions/shape, extract/dnd/scene-descriptions/source_refs, generic/valid_json_schema, extract/dnd/scene-descriptions/source_relatedness | generic/valid_json, extract/dnd/scene-descriptions/shape, normalize/dnd/scene-descriptions/invariants, extract/dnd/scene-descriptions/source_refs, generic/valid_json_schema, extract/dnd/scene-descriptions/source_relatedness |
+
+Chains are only registered for the D&D extract and normalize modules shown
+above; select an explicit override when a different compatible chain is
+required.
 
 ## Validation
 
-Configuration validation checks:
+Validate a file and one pipeline before running it:
 
-- supported config version and known YAML fields;
+~~~sh
-- non-empty, non-duplicated IDs after trimming;
+go run ./cmd/notarius config validate \
-- supported LLM provider and non-negative profile limits;
+  --config examples/dnd-minimal.config.yml \
-- positive global LLM concurrency;
+  --pipeline dnd-session
-- supported diagnostics retention and non-empty work directory;
+~~~
-- module binding LLM profiles refer to configured profiles.
 
-Pipeline resolution additionally checks:
+Configuration validation rejects invalid YAML, unsupported fields, invalid
-
+defaults or environment overrides, incompatible module keys, unknown options,
-- the pipeline ID exists;
+invalid reference bindings, missing required reference slots, invalid validator
-- at least one artifact lane is declared and selected;
+overrides, and incompatible generated artifact handoffs. Use
-- selected lanes exist when `--only` is used;
+[pipelines list](cli.md#pipelines-list) to inspect configured IDs.
-- required module keys are present;
-- module keys are registered for the expected slot;
-- module capability requirements are satisfied.

docs/consumers/subprocess.md

@@ -0,0 +1,74 @@
+# Using Notarius As A Subprocess
+
+Use this workflow when an orchestrator runs Notarius and consumes its published
+artifacts. The [CLI reference](../cli.md) owns invocation syntax and exit
+statuses, while the [run-result receipt](../integrations/run-result.md) and
+[Published JSON Output contract](../integrations/json-output.md) own the
+durable result formats.
+
+## Run And Check The Process
+
+Optionally preflight a selected configuration and pipeline before work starts:
+
+```sh
+notarius config validate --config /path/to/notarius.yml --pipeline pipeline-id
+```
+
+Invoke the run with explicit paths and machine-readable output. Capture
+standard output and standard error separately; do not combine them before
+processing the result.
+
+```sh
+notarius run pipeline-id \
+  --config /path/to/notarius.yml \
+  --input /path/to/source.json \
+  --output-dir /path/to/output-root \
+  --json
+```
+
+Use absolute paths for supplied input, configuration, output-root, and
+reference files. When a stable prompt session identifier or references are
+needed, pass the supported CLI flags. Supply credentials through Notarius's
+documented configuration and environment mechanisms, never as command-line
+arguments or generated secret-bearing configuration.
+
+Wait for the process before interpreting standard output. Only an exit status
+of 0 permits decoding the receipt. On a nonzero exit, retain standard error for
+diagnosis and ignore all standard-output bytes: a failed receipt write may have
+left a partial document.
+
+## Discover Required Artifacts
+
+Decode the successful receipt and accept the schema versions supported by the
+caller. Use its `output_directory` as the bundle root. For the production JSON
+output, resolve `index_file` under that root with a confinement check and reject
+an absolute path or a result that escapes the root.
+
+Read the resulting `index.json` and locate each artifact by `lane_id`, not by a
+guessed filename. Before decoding a selected payload, verify its descriptor's
+media type and schema identity against the relevant published artifact
+contract. The JSON bundle contract links to the available lane contracts.
+
+If `index.json` has an `evidence_context` descriptor, treat it as a
+pipeline-wide artifact rather than a lane entry. Verify its six descriptor
+fields before decoding the linked file according to the [Published Evidence
+Context contract](../integrations/evidence-context.md). Use each
+`evidence_refs` entry as the citation to source material. Its surrounding
+context range and included units explain the citation, but do not widen or
+replace the cited source reference.
+
+A zero exit status may still report rejected outputs, warnings, or absent
+lanes. The caller decides which lane IDs are required for its own work and
+which are optional; it should make that decision explicitly rather than infer
+failure from the receipt counts alone.
+
+## Preserve Provenance And Handle Data Carefully
+
+Keep the receipt with the published `manifest.json`, and retain
+`rejected.json` and `warnings.json` when review or later provenance requires
+them. Treat the input, output bundle, cache, debug bundle, and captured process
+logs as potentially sensitive data. Apply the caller's access controls and
+retention policy, and avoid copying secrets into arguments, logs, or
+provenance records. An evidence-context artifact contains source-unit text and
+metadata, and selected lanes can cover most of an input; preserve and share it
+only when that source content is authorized for the recipient.

docs/development.md

@@ -0,0 +1,43 @@
+# Development
+
+This is the first-read landing page for people and LLM coding agents working on
+Notarius. It provides a concise repository orientation and routes each kind of
+change to its canonical documentation.
+
+Notarius is a Go CLI for configured structured extraction workflows. Start with
+the [README](../README.md) for product context, [Architecture](policy/architecture.md)
+for system boundaries, and [Internal Overview](internal/overview.md) for the
+implemented component map.
+
+## What To Read
+
+| When working on | Read | Why |
+| --- | --- | --- |
+| Finding the package or component that owns current behavior | [Internal Overview](internal/overview.md) | It is the implemented component inventory and routes to focused internals. |
+| Application shape, package boundaries, contracts, dependency direction, runtime guarantees, or safety properties | [Architecture](policy/architecture.md) and relevant [ADRs](adr/) | Architecture defines the intended system and its invariants; ADRs preserve significant decision rationale. |
+| Any documentation addition or revision | [Documentation Policy](policy/documentation.md) | It defines canonical homes, audiences, current-behavior rules, and maintenance requirements. |
+| Adding, changing, reviewing, or deleting tests | [Testing Policy](policy/testing.md) | It defines risk-based sufficiency, durable test boundaries, test-double guidance, and criteria for retaining tests. |
+| CLI composition or command behavior | [CLI Internals](internal/cli.md) and [CLI Reference](cli.md) | The internal guide owns composition and command flow; the reference owns public syntax. |
+| Building a subprocess caller or changing its result protocol | [Subprocess Consumer Guide](consumers/subprocess.md), [Run Result Receipt](integrations/run-result.md), and [CLI Internals](internal/cli.md) | These separate caller workflow, durable receipt contract, and CLI implementation behavior. |
+| Configuration loading, resolution, or user-visible configuration behavior | [Configuration Internals](internal/configuration.md) and [Configuration](config.md) | The internal guide owns loading and resolution mechanics; the reference owns the configuration contract. |
+| Pipeline resolution or execution | [Pipeline Internals](internal/pipeline.md) | It documents profiles, references, validation, retries, checkpoints, and runner behavior. |
+| Production modules or validators | [Module Internals](internal/modules.md), [D&D Module Internals](internal/dnd.md), and [D&D integration contracts](integrations/) | The generic guide owns extension mechanics, the D&D guide owns shared family conventions, and the contracts own durable output shapes. |
+| LLM clients, prompts, schemas, profiles, or scheduling | [LLM Runtime](internal/llm.md) | It documents the transport boundary and PromptKit integration. |
+| Output, cache, resume, or debug artifacts | [Run State Internals](internal/state.md), [Operations](operations.md), and [Configuration](config.md) | These separate implementation details, operator behavior, and configuration contracts. |
+| External input formats, artifact schemas, or durable output files | [Integration Contracts](integrations/) | Integration documents define external and durable data contracts. |
+| Proposed or unimplemented behavior | [Roadmap](roadmap/) | Future work belongs only in roadmap documentation until implemented. |
+
+For an existing subsystem, also inspect its focused tests and the package-local
+types and contracts before changing behavior.
+
+## Validation
+
+Use focused package tests while iterating. Run the repository-wide checks when
+a change affects shared contracts, application behavior, or maintained
+documentation examples:
+
+```sh
+go test ./...
+go vet ./...
+go build ./cmd/notarius
+```

docs/integrations/chunk-map.md

@@ -0,0 +1,81 @@
+# Accepted Chunk Map
+
+This document defines the optional durable `chunk-map.json` artifact in a
+[published JSON bundle](json-output.md). It describes the accepted,
+materialized chunk plan used by one run. It is not a lane payload and is never
+an input to a later pipeline step.
+
+## Contract Identity
+
+| Property | Value |
+| --- | --- |
+| Artifact kind | `source/chunk-map` |
+| Logical file | `chunk-map.json` |
+| Media type | `application/json` |
+| Schema ID | `notarius.source.chunk_map` |
+| Schema name | `notarius_source_chunk_map_v1` |
+| Schema version | `v1` |
+
+The optional `chunk_map` descriptor in `index.json` identifies this artifact.
+Export is controlled by the JSON output binding described in
+[Configuration](../config.md#module-bindings-and-validators).
+
+## Wire Shape
+
+Every payload has these required fields:
+
+| Field | Meaning |
+| --- | --- |
+| `source_id` | Accepted source-document identity. |
+| `source_digest` | Lower-case `sha256:` digest of that source document. |
+| `plan_digest` | Lower-case `sha256:` digest of the logical chunk plan. |
+| `requested_chunker` | Chunk module selected by the resolved pipeline. |
+| `producer` | Original accepted-plan producer. `input_module` and `chunk_module` are required; `llm_profile` is optional. |
+| `plan_annotations` | Plan-level annotation namespace map; `{}` when none are present. |
+| `chunks` | Non-empty execution-order chunk collection. |
+
+Each `chunks` entry contains non-empty `id`, zero-based `index`, `source_ref`,
+positive `unit_count`, and an explicit `annotations` map. `source_ref` contains
+the same `source_id` as the top-level value plus positive inclusive
+`start_unit_id` and `end_unit_id` values. Endpoints identify source units; their
+numeric values do not by themselves establish source-document order.
+
+Annotation namespaces are non-empty trimmed strings. Their values are arbitrary
+valid JSON and are retained without interpreting a module-specific namespace.
+
+## Ordering And Validation
+
+`chunks` are in execution order. Their indexes are contiguous, start at zero,
+and equal their array positions; chunk IDs are unique. The emitted map is built
+only after the selected plan has been accepted and materialized against the
+source document, so its ranges, unit counts, annotations, and digests describe
+that exact plan.
+
+The codec rejects malformed JSON, trailing content, unknown fixed-object
+fields, invalid identities or digests, invalid annotations, duplicate chunk
+IDs, non-contiguous indexes, and a `plan_digest` that does not match the
+reconstructed logical plan. The checked-in
+[schema](../../internal/framework/chunkmap/assets/schemas/source_chunk_map.v1.json)
+defines the strict JSON shape.
+
+## Valid Example
+
+The compact
+[source chunk-map fixture](../../internal/framework/chunkmap/testdata/source_chunk_map.v1.json)
+is decoded by the production codec and demonstrates an accepted map with
+annotations, producer identity, and ordered chunks.
+
+## Publication And Compatibility
+
+The map is present only when a chunk plan was accepted and its export is
+enabled. It remains publishable if a later lane is rejected, but is absent when
+chunk-plan validation rejects the plan. `requested_chunker` identifies the
+current pipeline selection, while `producer` identifies the component that
+originally produced the accepted plan; they may differ when an accepted plan is
+reused.
+
+The map contains structure rather than source content: it excludes transcript
+bytes, source-unit metadata, chunk text, private model output, reference
+content, debug data, and filesystem paths. Treat the exported map with the
+same care as other published output. Publication location and retention are
+defined in [Operations](../operations.md#output-bundles).

docs/integrations/dnd-combat-turn-artifacts.md

@@ -0,0 +1,69 @@
+# D&D Combat-Turn Artifact
+
+This contract defines the durable combat-action occurrence list produced by
+`dnd/combat-turns`. It records source-grounded turns and actions; it is not a
+complete initiative tracker, combat summary, or state model.
+
+## Identity and compatibility
+
+| Property | Value |
+| --- | --- |
+| Artifact kind | `dnd/combat-turn-list` |
+| Schema ID | `notarius.dnd.combat_turns` |
+| Schema name | `notarius_dnd_combat_turns_v1` |
+| Schema version | `v1` |
+| Media type | `application/json` |
+
+`v1` is a strict JSON object with required `combat_turns`; the array may be
+empty. Turn and source-reference objects reject unknown fields. An incompatible
+shape change requires a new schema version.
+
+## Wire shape
+
+Each combat turn has these required fields:
+
+| Field | Contract |
+| --- | --- |
+| `actor` | Non-empty acting character or creature name. |
+| `turn_kind` | `turn`, `reaction`, `legendary_action`, `lair_action`, or `other`. |
+| `source_refs` | One or more transcript evidence ranges. |
+
+Each source reference has exactly `source_id`, `start_unit_id`, and
+`end_unit_id`. It identifies an inclusive current-transcript range; unit IDs
+are positive and the start may not follow the end.
+
+```json
+{
+  "combat_turns": [
+    {
+      "actor": "Mira Thorn",
+      "turn_kind": "turn",
+      "source_refs": [
+        {"source_id": "session-7", "start_unit_id": 31, "end_unit_id": 32}
+      ]
+    }
+  ]
+}
+```
+
+## Eligibility, evidence, and normalized form
+
+The extractor requires an approved [scene-description artifact](dnd-scene-description-artifacts.md).
+It emits combat turns only for a chunk with an exact matching scene classified
+`combat`; an exact non-combat scene produces an accepted empty list. The scene
+record controls eligibility only: its title, summary, and reference do not
+become turn evidence. No exact matching scene also produces an empty list and
+the `scene_classification_unavailable` warning.
+
+An optional normalized [NPC artifact](dnd-npc-artifacts.md) can ground an
+actor name. Its registry references are provenance, never combat evidence.
+Normalization trims and, where possible, canonicalizes actor names; orders and
+deduplicates exact source references; orders valid-evidence turns by source
+chronology; and collapses only duplicates with the same actor identity, turn
+kind, and complete valid evidence. It does not infer turns, initiative, or
+actions from registry or scene data.
+
+The [NPC-interaction artifact](dnd-npc-interaction-artifacts.md) records
+broader NPC occurrences. The [JSON output contract](json-output.md) defines
+publication, and [D&D module internals](../internal/dnd.md) describes routing
+and validation mechanics.

docs/integrations/dnd-item-event-artifacts.md

@@ -0,0 +1,78 @@
+# D&D Item-Event Artifact
+
+This contract defines the durable item and currency occurrence list produced by
+`dnd/item-events`. It records source-grounded discoveries and possession
+changes; it does not maintain an inventory, balance, or ledger.
+
+## Identity and compatibility
+
+| Property | Value |
+| --- | --- |
+| Artifact kind | `dnd/item-event-list` |
+| Schema ID | `notarius.dnd.item_events` |
+| Schema name | `notarius_dnd_item_events_v1` |
+| Schema version | `v1` |
+| Media type | `application/json` |
+
+`v1` is a strict JSON object with required `events`; the array may be empty.
+Event and source-reference objects reject unknown fields. An incompatible
+shape change requires a new schema version.
+
+## Wire shape
+
+Every event has required `name`, `kind`, and `source_refs`. `quantity`, `from`,
+and `to` are optional where the event kind permits them.
+
+| Field | Contract |
+| --- | --- |
+| `name` | Non-empty item or currency display name. |
+| `kind` | `discovered`, `acquired`, `lost`, `consumed`, or `transferred`. |
+| `quantity` | Optional positive integer; omit it when no count is established. |
+| `from` | Optional non-empty losing holder, when allowed by `kind`. |
+| `to` | Optional non-empty gaining holder, when allowed by `kind`. |
+| `source_refs` | One or more transcript evidence ranges. |
+
+Each source reference has exactly `source_id`, `start_unit_id`, and
+`end_unit_id`. It identifies an inclusive current-transcript range; unit IDs
+are positive and the start may not follow the end.
+
+```json
+{
+  "events": [
+    {
+      "name": "Silver Pieces",
+      "kind": "acquired",
+      "quantity": 20,
+      "to": "party",
+      "source_refs": [
+        {"source_id": "session-7", "start_unit_id": 2, "end_unit_id": 2}
+      ]
+    }
+  ]
+}
+```
+
+## Holder rules and minimal extraction
+
+`discovered` has neither holder; `acquired` requires `to` and forbids `from`;
+`lost` and `consumed` require `from` and forbid `to`; `transferred` requires
+both holders. `party` denotes collective possession. A transfer cannot use
+`party` for either holder and its two normalized holders must differ.
+
+Only an evidenced discovery or possession change belongs in this artifact.
+It does not infer quantities or holders, convert currency denominations,
+calculate balances, or merge nearby events. Campaign references may
+disambiguate names but are never event evidence. Currency uses the ordinary
+`name` field and an explicit `quantity` only when the transcript establishes
+one; each denomination remains a separate event.
+
+Normalization trims display whitespace, orders and removes exact duplicate
+source references, then orders events by valid source chronology, name identity
+and display value, kind, holders, quantity, and reference sequence. It
+collapses only entries with the same normalized durable fields and complete
+valid evidence.
+
+The [JSON output contract](json-output.md) defines publication. See
+[D&D module internals](../internal/dnd.md) for implementation details and the
+[NPC-interaction artifact](dnd-npc-interaction-artifacts.md) for a distinct
+kind of occurrence.

docs/integrations/dnd-npc-artifacts.md

@@ -0,0 +1,69 @@
+# D&D NPC Artifact
+
+This contract defines the durable NPC registry produced by `dnd/npcs`. It is a
+minimal, source-grounded identity registry for other D&D artifacts, not a
+character sheet or a relationship summary.
+
+## Identity and compatibility
+
+| Property | Value |
+| --- | --- |
+| Artifact kind | `dnd/npc-list` |
+| Schema ID | `notarius.dnd.npcs` |
+| Schema name | `notarius_dnd_npcs_v1` |
+| Schema version | `v1` |
+| Media type | `application/json` |
+| Identity policy | `dnd.npcs.identity.v1` |
+
+`v1` accepts one strict JSON object with required `npcs`; the array may be
+empty. NPC and source-reference objects reject unknown fields. An incompatible
+artifact shape or identity-policy change uses a new version or policy.
+
+## Wire shape and identity
+
+Each NPC has these required fields:
+
+| Field | Contract |
+| --- | --- |
+| `id` | `npc:sha256:` followed by 64 lowercase hexadecimal characters. |
+| `name` | Non-empty canonical display name. |
+| `source_refs` | One or more transcript evidence ranges for the identity. |
+
+A source reference has exactly `source_id`, `start_unit_id`, and `end_unit_id`.
+The source ID identifies the transcript, unit IDs are positive inclusive unit
+identifiers, and the start may not follow the end.
+
+```json
+{
+  "npcs": [
+    {
+      "id": "npc:sha256:99a16589618a04f535a7d21fdcc71a0b1c05d22f752cd492065b1086d97bc3d7",
+      "name": "Mira Thorn",
+      "source_refs": [
+        {"source_id": "session-7", "start_unit_id": 4, "end_unit_id": 5}
+      ]
+    }
+  ]
+}
+```
+
+The ID is deterministic: normalize the name to Unicode NFKC, normalize the
+supported apostrophe forms, collapse whitespace, case-fold it, SHA-256 the
+result, then prefix the lowercase hexadecimal digest with `npc:sha256:`. Each
+canonical identity and ID appears at most once. Normalization collapses records
+with the same canonical identity, retains their earliest position, and merges
+their canonicalized evidence; it does not add aliases, roles, descriptions, or
+relationship fields.
+
+## Scope and consumers
+
+Only individually identifiable NPC names with transcript evidence belong in
+this artifact. Groups, generic roles, invented labels, and descriptive
+enrichment are excluded. Its source references prove registry provenance; they
+do not become evidence for a spell, interaction, or combat occurrence.
+
+This registry can ground actor or caster names in the [spell](dnd-spell-artifacts.md)
+and [combat-turn](dnd-combat-turn-artifacts.md) artifacts. It is required to
+resolve the canonical `name` in an [NPC interaction](dnd-npc-interaction-artifacts.md).
+The [JSON output contract](json-output.md) defines publication, and
+[D&D module internals](../internal/dnd.md) owns pipeline mechanics.

docs/integrations/dnd-npc-interaction-artifacts.md

@@ -0,0 +1,78 @@
+# D&D NPC Interaction Artifact
+
+This contract defines the durable occurrence list produced by
+`dnd/npc-interactions`. It records discrete, source-grounded interactions with
+NPCs already present in a normalized registry; it does not extend that registry
+or summarize the session.
+
+## Identity and compatibility
+
+| Property | Value |
+| --- | --- |
+| Artifact kind | `dnd/npc-interaction-list` |
+| Schema ID | `notarius.dnd.npc_interactions` |
+| Schema name | `notarius_dnd_npc_interactions_v1` |
+| Schema version | `v1` |
+| Media type | `application/json` |
+
+`v1` is a strict JSON object with required `interactions`; the array may be
+empty. Interaction and source-reference objects reject unknown fields. An
+incompatible shape change requires a new schema version.
+
+## Wire shape
+
+Each interaction has these required fields:
+
+| Field | Contract |
+| --- | --- |
+| `name` | Non-empty canonical display name from the required NPC registry. |
+| `kind` | One of the interaction categories below. |
+| `source_refs` | One or more transcript evidence ranges. |
+
+Each source reference has exactly `source_id`, `start_unit_id`, and
+`end_unit_id`. It identifies an inclusive range in the current transcript;
+unit IDs are positive and the start may not follow the end. Extraction evidence
+for an interaction is confined to its accepted chunk.
+
+```json
+{
+  "interactions": [
+    {
+      "name": "Mira Thorn",
+      "kind": "dialogue",
+      "source_refs": [
+        {"source_id": "session-7", "start_unit_id": 12, "end_unit_id": 13}
+      ]
+    }
+  ]
+}
+```
+
+## Interaction categories
+
+| Kind | Meaning |
+| --- | --- |
+| `mentioned` | The NPC is referred to but is not established as present or communicating. |
+| `noncombat_presence` | The NPC is present and relevant without meaningful dialogue or combat participation. |
+| `dialogue` | The NPC speaks, responds, or meaningfully participates in a non-combat exchange. |
+| `combat_ally` | The NPC actively participates in combat on the party's side. |
+| `combat_opponent` | The NPC actively participates in combat against the party. |
+| `other` | A clearly evidenced direct occurrence not covered by another category. |
+
+The categories do not represent motives, relationships, state, or events that
+the cited transcript does not establish. An `other` entry is not a substitute
+for uncertain classification.
+
+## Identity, evidence, and order
+
+The required normalized [NPC artifact](dnd-npc-artifacts.md) resolves `name`.
+Registry references are provenance only and never replace an interaction's own
+evidence. Normalization canonicalizes recognized registry names, orders and
+deduplicates exact source references, then orders interactions by valid source
+chronology, NPC comparison identity, display name, kind, and reference sequence.
+Only entries with the same canonical name, kind, and complete valid evidence
+sequence are collapsed; distinct categories or evidence remain separate.
+
+See the [combat-turn artifact](dnd-combat-turn-artifacts.md) for combat-action
+occurrences and the [JSON output contract](json-output.md) for publication.
+Pipeline mechanics are described in [D&D module internals](../internal/dnd.md).

docs/integrations/dnd-scene-description-artifacts.md

@@ -0,0 +1,69 @@
+# D&D Scene-Description Artifact
+
+This contract defines the durable output of `dnd/scene-descriptions`. Each
+record classifies one accepted transcript chunk and gives it a minimal
+source-grounded title and summary.
+
+## Identity and compatibility
+
+| Property | Value |
+| --- | --- |
+| Artifact kind | `dnd/scene-description-list` |
+| Schema ID | `notarius.dnd.scene_descriptions` |
+| Schema name | `notarius_dnd_scene_descriptions_v1` |
+| Schema version | `v1` |
+| Media type | `application/json` |
+
+`v1` is a strict JSON object with required non-empty `scenes`. Scene and
+source-reference objects reject unknown fields. An incompatible shape change
+requires a new schema version.
+
+## Wire shape
+
+Each scene has exactly these required fields:
+
+| Field | Contract |
+| --- | --- |
+| `id` | Non-empty accepted chunk ID, assigned by Notarius. |
+| `source_ref` | The assigned inclusive source range for that chunk. |
+| `kind` | `combat`, `narrative`, `recap`, or `meta`. |
+| `title` | Non-empty, trimmed, source-grounded title. |
+| `summary` | Non-empty, trimmed, source-grounded summary. |
+
+`source_ref` has exactly `source_id`, `start_unit_id`, and `end_unit_id`.
+Its source ID identifies the input transcript; its positive unit IDs identify
+the chunk's inclusive range, with the start no later than the end.
+
+```json
+{
+  "scenes": [
+    {
+      "id": "chunk-000001",
+      "source_ref": {"source_id": "session-7", "start_unit_id": 1, "end_unit_id": 3},
+      "kind": "narrative",
+      "title": "Arrival at the watchtower",
+      "summary": "The party reaches the ruined watchtower and begins to investigate it."
+    }
+  ]
+}
+```
+
+## Meaning and normalized form
+
+`combat` identifies a chunk where active combat is the central activity.
+`narrative` is current in-world play that is not principally combat, recap, or
+meta discussion. `recap` is primarily a recounting of an earlier session, and
+`meta` is primarily out-of-character discussion. The artifact does not add
+participants, confidence, events, or information absent from the chunk.
+
+Normalization trims title and summary, orders scenes by source position and
+then ID, and removes exact duplicate records. A reused ID with different
+durable fields, or the same source range with different kind, title, or
+summary, is invalid. It does not merge adjacent ranges, alter prose, or infer
+missing scenes.
+
+The [combat-turn artifact](dnd-combat-turn-artifacts.md) uses an exact matching
+`combat` scene only as eligibility control; scene title, summary, and source
+reference never become combat evidence. Publication is defined by the
+[JSON output contract](json-output.md); implementation details live in
+[D&D module internals](../internal/dnd.md).

docs/integrations/dnd-spell-artifacts.md

@@ -1,152 +1,73 @@
-# D&D Spell-Cast Artifacts
+# D&D Spell Artifact
 
-This document is the durable artifact contract for approved
+This contract defines the durable output of the `dnd/spells` extractor and
-`dnd.spell_cast` artifacts produced by the implemented `dnd/spells` extractor.
+normalizer. It records source-grounded spell-casting occurrences; it is not a
+spellbook, a rules lookup result, or a record of hypothetical casts.
 
-## Artifact Identity
+## Identity and compatibility
 
-- Extractor key: `dnd/spells`
+| Property | Value |
-- Artifact type: `dnd.spell_cast`
+| --- | --- |
-- Schema version: `v1`
+| Artifact kind | `dnd/spell-list` |
-- Prompt ID: `dnd.spells`
+| Schema ID | `notarius.dnd.spells` |
-- Response schema key: `dnd_spells`
+| Schema name | `notarius_dnd_spells_v1` |
-- Response schema ID: `notarius.dnd.spells`
+| Schema version | `v1` |
-- Response schema name: `notarius_dnd_spells_v1`
+| Media type | `application/json` |
 
-The extractor requires source chunks and transcript source capability. It
+`v1` is a single strict JSON object. It requires `spell_casts`; the array may
-returns generic artifact candidates that are serialized by the JSON output
+be empty. Each spell-cast object and source-reference object rejects unknown
-module.
+fields. An incompatible shape change requires a new schema version.
 
-## Artifact Envelope
+## Wire shape
 
-Approved artifacts use the generic artifact envelope documented in
+Each `spell_casts` entry has these required fields:
-[JSON Output](json-output.md#artifact-files):
 
-```json
+| Field | Contract |
-{
+| --- | --- |
-  "extractor_key": "dnd/spells",
+| `caster` | Non-empty in-world character or creature name. |
-  "artifact_type": "dnd.spell_cast",
+| `spell` | Non-empty spell name. |
-  "schema_version": "v1",
+| `source_refs` | One or more transcript evidence ranges. |
-  "payload": {
-    "caster": "Aria",
-    "spell": "Cure Wounds",
-    "effect": "heals an injured ally",
-    "narrative_description": "Aria raises her holy symbol and casts Cure Wounds."
-  },
-  "source_refs": [
-    {
-      "source_id": "session-alpha",
-      "start_unit_id": "seg-001",
-      "end_unit_id": "seg-001"
-    }
-  ]
-}
-```
 
-## Payload Fields
+Every source reference has exactly `source_id`, `start_unit_id`, and
-
+`end_unit_id`. The source ID identifies the input transcript; the unit IDs are
-The `payload` object contains:
+positive inclusive unit identifiers, and the start may not follow the end in
-
+that source. References are evidence for the cast, not campaign-reference or
-- `caster`: in-world character or creature casting the spell;
+NPC-registry provenance.
-- `spell`: spell name;
-- `effect`: concise spell effect in the scene;
-- `narrative_description`: short description of the spell cast in context.
-
-All payload fields are strings and must be non-empty after trimming.
-
-`caster` is the in-world caster, not the transcript speaker.
-
-## Source References
-
-Source references live on the artifact envelope as `source_refs`; they are not
-duplicated inside the `payload`.
-
-Each source reference uses the generic source-reference shape:
-
-- `source_id`
-- `start_unit_id`
-- `end_unit_id`
-
-Validation requires:
-
-- at least one source reference;
-- non-empty source ID and unit IDs;
-- source ID matching the source document ID;
-- start and end unit IDs existing in the source document;
-- start unit appearing before or at the same position as end unit.
-
-## Structured LLM Response Shape
-
-The extractor asks the LLM for this top-level response shape:
 
 ```json
 {
   "spell_casts": [
     {
-      "caster": "Aria",
+      "caster": "Mira Thorn",
-      "spell": "Cure Wounds",
+      "spell": "Fireball",
-      "effect": "heals an injured ally",
-      "narrative_description": "Aria raises her holy symbol and casts Cure Wounds.",
       "source_refs": [
-        {
+        {"source_id": "session-7", "start_unit_id": 12, "end_unit_id": 13}
-          "source_id": "session-alpha",
-          "start_unit_id": "seg-001",
-          "end_unit_id": "seg-001"
-        }
       ]
     }
   ]
 }
 ```
 
-`spell_casts` must be present. It may be empty when no spell casts are found.
+## Evidence and normalized form
 
-The response schema asset is embedded at
+An entry represents an actual cast or an unambiguous declared attempt. A spell
-`internal/modules/extract/dnd/spells/assets/schemas/dnd_spells.v1.json`.
+mention, rules discussion, plan, or catalog match alone is not an occurrence.
+The configured catalog checks the name; it does not establish evidence.
 
-## Validators
+When normalization is selected, recognized spell names use the effective
+catalog's canonical display name. Source references are put in canonical source
+order and exact duplicate references are removed. A later entry is collapsed
+only when it has the same canonical spell, the same case- and
+whitespace-insensitive caster identity, and the same complete valid reference
+sequence. Remaining entries retain their merged order.
 
-The extractor supplies two deterministic validators by default:
+The optional normalized [NPC artifact](dnd-npc-artifacts.md) can ground a
+caster name. Its own references remain registry provenance and are never copied
+into `source_refs`.
 
-- `dnd/spells/shape`
+## Related contracts
-- `dnd/spells/source_refs`
 
-Rejection reason codes:
+The [spell-catalog overlay contract](dnd-spell-catalog-overlays.md) defines
-
+the configured catalog additions. The [JSON output contract](json-output.md)
-- `invalid_payload`: payload JSON cannot be decoded as a spell-cast payload.
+defines where this logical artifact is published; [D&D module internals](../internal/dnd.md)
-- `missing_required_field`: `caster`, `spell`, `effect`, or
+describes extraction and validation mechanics.
-  `narrative_description` is blank.
-- `missing_source_ref`: candidate has no source references.
-- `invalid_source_ref`: at least one source reference fails generic source
-  reference validation.
-
-Rejected candidates are written to `rejected.json` by the JSON output module.
-
-## Manifest Metadata
-
-The extractor adds prompt and response-schema provenance under the artifact lane
-manifest metadata:
-
-```json
-{
-  "metadata": {
-    "extractor": {
-      "prompt_id": "dnd.spells",
-      "prompt_version": "v1",
-      "prompt_sha256": "sha256:...",
-      "response_schema_key": "dnd_spells",
-      "response_schema_id": "notarius.dnd.spells",
-      "response_schema_name": "notarius_dnd_spells_v1",
-      "response_schema_version": "v1",
-      "response_schema_sha256": "sha256:..."
-    }
-  }
-}
-```
-
-Raw prompt and schema content are not included in manifest metadata.
-
-## Compatibility Limit
-
-This contract covers only `dnd.spell_cast` artifacts produced by the
-implemented spell-cast extractor.

docs/integrations/dnd-spell-catalog-overlays.md

@@ -0,0 +1,72 @@
+# D&D Spell-Catalog Overlays
+
+This document defines the optional JSON overlay consumed by the D&D spell
+extractor. An overlay contributes campaign spell names and aliases for
+recognition. It does not define spell rules, effects, levels, classes, or
+transcript evidence. Bind the optional `spell_catalog` reference as described
+in [Configuration](../config.md#references-and-ordered-handoffs).
+
+## Contract Identity
+
+| Property | Value |
+| --- | --- |
+| Consumer | D&D spell extraction and normalization |
+| Reference slot | `spell_catalog` |
+| Media type | `application/json` |
+| Required schema version | `notarius.dnd.spell-catalog-overlay.v1` |
+| Base catalog | Embedded D&D 5e 2014 SRD catalog |
+
+At most one overlay document may be bound. The maintained example is
+[dnd-spell-catalog.json](../../examples/dnd-spell-catalog.json).
+
+## Wire Shape
+
+This is a minimal valid overlay:
+
+```json
+{
+  "schema_version": "notarius.dnd.spell-catalog-overlay.v1",
+  "catalogs": [
+    {
+      "id": "campaign.example",
+      "ruleset": "dnd-5e-2014",
+      "source": {"title": "Example campaign spells"},
+      "spells": [{"name": "Aegis of Emberfall"}]
+    }
+  ]
+}
+```
+
+| Field | Required | Meaning and constraints |
+| --- | --- | --- |
+| `schema_version` | Yes | Exactly `notarius.dnd.spell-catalog-overlay.v1`. |
+| `catalogs` | Yes | Non-empty array of catalog objects with unique IDs. |
+| `catalogs[].id` | Yes | Non-empty trimmed string. |
+| `catalogs[].ruleset` | Yes | Exactly `dnd-5e-2014`. |
+| `catalogs[].source.title` | Yes | Non-empty trimmed string. |
+| `catalogs[].source.version` | No | String when present. |
+| `catalogs[].source.url` | No | String when present. |
+| `catalogs[].source.license` | No | String when present. |
+| `catalogs[].spells` | Yes | Non-empty array of spell objects. |
+| `catalogs[].spells[].name` | Yes | Non-empty trimmed string. |
+| `catalogs[].spells[].aliases` | No | Array of non-empty trimmed strings when present. |
+
+Unknown fields are rejected at every object level. The document must contain
+one JSON value; `null` is not accepted for optional strings or aliases.
+
+## Composition And Compatibility
+
+Notarius starts with the embedded base catalog, then applies overlay catalogs
+in ascending catalog-ID order. A new canonical spell name adds a recognition
+entry. If an overlay names an existing canonical spell, it augments that spell
+with aliases while retaining the established display spelling.
+
+Repeated aliases for the same spell are accepted. A canonical-name, canonical-
+to-alias, or alias-to-alias collision between different spells is rejected,
+including a collision with the embedded catalog. Matching uses the catalog’s
+case, whitespace, and apostrophe normalization, so authors should avoid names
+or aliases that normalize to another spell.
+
+The overlay is a recognition aid only. The durable spell-artifact schema and
+source-evidence rules are defined by the
+[D&D spell artifact contract](dnd-spell-artifacts.md).

docs/integrations/evidence-context.md

@@ -0,0 +1,116 @@
+# Published Evidence Context
+
+This contract defines the optional `source/evidence-context` artifact emitted
+by the production JSON output. Its configuration is owned by
+[Configuration](../config.md#module-bindings-and-validators); its logical-file
+discovery is owned by [Published JSON Output](json-output.md).
+
+## Identity And Discovery
+
+When enabled, the JSON bundle contains `evidence-context.json` and an
+`index.json` `evidence_context` descriptor with the same six fields as other
+pipeline-wide artifact descriptors.
+
+| Property | Value |
+| --- | --- |
+| Artifact kind | `source/evidence-context` |
+| Media type | `application/json` |
+| Schema ID | `notarius.source.evidence_context` |
+| Schema name | `notarius_source_evidence_context_v1` |
+| Schema version | `v1` |
+| Logical file | `evidence-context.json` |
+
+Consumers must discover the file from the descriptor, verify all six descriptor
+fields, and decode only a supported schema version. The descriptor is optional:
+its absence means evidence publication was not enabled for that bundle.
+
+## Payload
+
+The v1 payload is a JSON object with required `source_id`, `source_digest`,
+`window_units`, `selected_lanes`, and `contexts` fields. `selected_lanes` and
+`contexts` are always arrays; an enabled configuration with no accepted direct
+evidence publishes `contexts: []`.
+
+```json
+{
+  "source_id": "session-alpha",
+  "source_digest": "sha256:0123456789abcdef0123456789abcdef0123456789abcdef0123456789abcdef",
+  "window_units": 1,
+  "selected_lanes": ["npcs", "spells"],
+  "contexts": [
+    {
+      "context_ref": {
+        "source_id": "session-alpha",
+        "start_unit_id": 10,
+        "end_unit_id": 20
+      },
+      "evidence_refs": [
+        {
+          "lane_id": "spells",
+          "source_ref": {
+            "source_id": "session-alpha",
+            "start_unit_id": 10,
+            "end_unit_id": 10
+          }
+        }
+      ],
+      "units": [
+        {
+          "id": 10,
+          "kind": "transcript_segment",
+          "text": "Aria casts Cure Wounds.",
+          "ref": {
+            "source_id": "session-alpha",
+            "start_unit_id": 10,
+            "end_unit_id": 10
+          }
+        },
+        {
+          "id": 20,
+          "kind": "transcript_segment",
+          "text": "The party regroups.",
+          "ref": {
+            "source_id": "session-alpha",
+            "start_unit_id": 20,
+            "end_unit_id": 20
+          }
+        }
+      ]
+    }
+  ]
+}
+```
+
+Each context requires a `context_ref` object and `evidence_refs` and `units`
+arrays. `context_ref` identifies the first and last included unit. Each
+evidence entry contains a selected `lane_id` and an original `source_ref`. A
+unit uses the existing source-unit shape: required `id`, `kind`, `text`, and
+self `ref`, plus optional JSON-object `metadata`. Fixed payload objects reject
+unknown fields; unit metadata may contain application-defined JSON values.
+
+## Citations And Context
+
+`evidence_refs` are the authoritative citations. They identify the direct
+references emitted by accepted normalized artifacts. `context_ref` and the
+units collection include those cited units plus nearby source units selected by
+the configured window. They are explanatory context, not widened citations.
+
+Only accepted outputs from the configured lane allowlist contribute. Rejected,
+failed, absent, and lane-filtered outputs do not contribute. The artifact never
+contains raw input bytes, prompts, model responses, auxiliary reference
+content, credentials, or filesystem paths.
+
+## Ordering And Compatibility
+
+The selected lane allowlist is lexical. Contexts and units are in source
+document position order, not numeric unit-ID order. Direct evidence entries
+are deterministically ordered by lane and source reference. Overlapping or
+contiguous windows merge, and each source unit appears at most once in the
+resulting contexts.
+
+The artifact is additive to the JSON bundle and is not a lane payload,
+normalized-output count, checkpoint, or generated reference. Consumers that
+do not need it must tolerate the absent optional descriptor. Consumers that do
+use it should preserve the artifact and its schema identity with the run
+provenance, and should treat its source text and metadata as sensitive durable
+content.

docs/integrations/json-output.md

@@ -1,192 +1,133 @@
-# JSON Output
+# Published JSON Output
 
-This document is the durable JSON output file-format contract produced by the
+This document defines the logical JSON bundle emitted by the production JSON
-implemented `json` output module and written by the CLI.
+output encoder. The bundle’s physical destination, atomic publication, and
+retention are operational concerns; see [Operations](../operations.md#output-bundles).
+Output configuration, including chunk-map and evidence-context publication, belongs in
+[Configuration](../config.md#module-bindings-and-validators).
 
-## Output Directory
+## Bundle Layout
 
-The CLI writes logical output files under:
+All paths below are logical, relative, slash-separated bundle paths. The
+encoder always emits the first four JSON files below and adds lane or
+pipeline-wide artifact files when their corresponding artifacts are available:
 
-```text
+A subprocess caller first obtains the physical bundle root from the
-<output-root>/<run-id>/
+[run-result receipt](run-result.md), then resolves `index.json` beneath that
-```
+root for the logical discovery described here.
 
-The default output root is `./notarius-output`. Operational behavior is covered
+| Path | Purpose |
-in [Operations](../operations.md).
+| --- | --- |
+| `index.json` | Entry point that names the other published files and lane payloads. |
+| `manifest.json` | Run provenance and result summaries. |
+| `rejected.json` | Rejected pipeline outputs. |
+| `warnings.json` | Accepted-output and run warnings. |
+| `lanes/<safe-lane-id>.json` | One normalized artifact payload for each lane. |
+| `chunk-map.json` | Optional accepted chunk map, when its export is enabled and available. |
+| `evidence-context.json` | Optional source-context artifact, when evidence publication is enabled. |
 
-## Files
+JSON files are pretty-printed with a trailing newline. Lane payloads are
-
+accepted only when their media type is `application/json`.
-The `json` output module writes:
-
-- `index.json`
-- `manifest.json`
-- `artifacts/<artifact-type>.json`, one file per approved artifact type
-- `rejected.json`
-- `warnings.json`
-
-Files are pretty-printed JSON with a trailing newline.
 
 ## `index.json`
 
-Shape:
+`index.json` is the bundle’s discovery document. An approved run with no
+normalized lanes has this valid minimal index:
 
 ```json
 {
   "manifest_file": "manifest.json",
-  "artifact_files": [
+  "output_files": [],
-    {
-      "artifact_type": "dnd.spell_cast",
-      "file": "artifacts/dnd.spell_cast.json"
-    }
-  ],
   "rejected_file": "rejected.json",
   "warnings_file": "warnings.json"
 }
 ```
 
-`artifact_files` is sorted by artifact type. It is empty when no artifacts are
+| Field | Required | Meaning |
-approved.
+| --- | --- | --- |
+| `manifest_file` | Yes | Always `manifest.json`. |
+| `output_files` | Yes | Lane descriptors sorted by `lane_id`. |
+| `rejected_file` | Yes | Always `rejected.json`. |
+| `warnings_file` | Yes | Always `warnings.json`. |
+| `chunk_map` | No | Descriptor for the pipeline-wide `chunk-map.json`; never a lane descriptor. |
+| `evidence_context` | No | Descriptor for the pipeline-wide `evidence-context.json`; never a lane descriptor. |
+
+Each lane descriptor has required `lane_id` and `file`. It may also include
+`media_type`, `module_key`, `schema_id`, `schema_name`, and `schema_version`
+when supplied by the normalized artifact. Each pipeline-wide artifact
+descriptor (`chunk_map` or `evidence_context`) contains `artifact_kind`,
+`file`, `media_type`, `schema_id`, `schema_name`, and `schema_version`. Their
+payloads are defined by the [Accepted Chunk Map contract](chunk-map.md) and
+[Published Evidence Context](evidence-context.md), respectively.
+
+The lane path is derived from its lane ID. Characters outside letters, digits,
+periods, underscores, and hyphens become underscores; `..` sequences are
+neutralized; leading and trailing periods and underscores are removed. A lane
+that produces an empty name, or two lanes that produce the same path, makes
+output encoding fail.
+
+## Lane Payloads
+
+Each `lanes/<safe-lane-id>.json` file is the codec-owned normalized JSON for
+that lane. Consumers should use the index descriptor’s schema identity rather
+than infer a lane schema from its name. The current D&D payload contracts are
+[spells](dnd-spell-artifacts.md), [NPCs](dnd-npc-artifacts.md),
+[NPC interactions](dnd-npc-interaction-artifacts.md),
+[combat turns](dnd-combat-turn-artifacts.md),
+[item events](dnd-item-event-artifacts.md), and
+[scene descriptions](dnd-scene-description-artifacts.md).
 
 ## `manifest.json`
 
-`manifest.json` contains a run manifest:
+`manifest.json` is published provenance, not a copy of lane payloads or a
+checkpoint store. Fields without a value may be omitted. Its top-level fields
+group into the following externally observable summaries:
 
-```json
+| Group | Fields |
-{
+| --- | --- |
-  "run_id": "run-123",
+| Run identity and result | `run_id`, `pipeline_id`, `pipeline_digest`, `schema_version`, `validation_status`, `started_at`, `completed_at` |
-  "pipeline_id": "dnd-session",
+| Resolved components | `input_module`, `chunker`, `extractors`, `merger`, `normalizer`, `output_encoder`, `artifact_lanes`, `validator_chains`, `module_metadata` |
-  "pipeline_digest": "sha256:...",
+| Source and references | `source_digests`, `references` |
-  "input_module": "seriatim",
+| Published result summaries | `normalized_outputs`, `rejected_outputs` |
-  "chunker": "generic",
+| Execution summaries | `chunk_plan`, `checkpoint_decisions`, `llm_profiles`, `metadata` |
-  "source_digests": ["sha256:..."],
-  "extractors": ["dnd/spells"],
-  "merger": "appendorder",
-  "normalizer": "noop",
-  "output_encoder": "json",
-  "artifact_lanes": [
-    {
-      "id": "spells",
-      "extractor": "dnd/spells",
-      "merger": "appendorder",
-      "normalizer": "noop"
-    }
-  ],
-  "llm_profiles": [
-    {
-      "id": "default",
-      "provider": "openai-compatible",
-      "model": "configured-model"
-    }
-  ],
-  "validation_status": "approved",
-  "started_at": "2026-01-01T00:00:00Z",
-  "completed_at": "2026-01-01T00:00:01Z"
-}
-```
 
-Fields with empty values may be omitted by JSON encoding.
+`references` records provenance such as the target, slot, origin, digest,
+media type, size, and generated-artifact identity. It does not contain
+reference content. `normalized_outputs` and `rejected_outputs` likewise
+summarize results without embedding lane payload bytes. A chunk-plan summary is
+provenance for the plan used by this run; cache records, debug artifacts, and
+other operational state are not published as bundle files.
 
-`validation_status` is `approved` when no candidates were rejected and
+Each `llm_profiles` entry identifies effective, non-secret LLM execution
-`rejected` when one or more candidates were rejected.
+provenance:
 
-## Artifact Files
+| Field | Required | Meaning |
+| --- | --- | --- |
+| `id` | Yes | Selected PromptKit profile identifier. |
+| `provider` | No | Notarius adapter provider identifier. |
+| `model` | No | Effective provider model identifier. |
+| `backend_id` | No | Effective PromptKit backend registration identifier. Endpoint-only profiles omit it. |
+| `reasoning_effort` | No | Effective opaque provider reasoning setting. An empty or explicitly cleared setting is omitted. |
 
-Each artifact file has this shape:
+These values describe observed execution; they are not a backend-registration
+interface. Entries that differ by backend or effective reasoning remain
+distinct even when their profile, provider, and model are otherwise equal.
 
-```json
+## Rejections And Warnings
-{
-  "artifact_type": "dnd.spell_cast",
-  "artifacts": [
-    {
-      "extractor_key": "dnd/spells",
-      "artifact_type": "dnd.spell_cast",
-      "schema_version": "v1",
-      "payload": {},
-      "source_refs": [
-        {
-          "source_id": "session-alpha",
-          "start_unit_id": "seg-001",
-          "end_unit_id": "seg-001"
-        }
-      ]
-    }
-  ]
-}
-```
 
-Artifact envelope fields:
+`rejected.json` is always an object with a `rejected` array. Each entry has
+required `stage` and `message`; `step_id`, `lane_id`, `module_key`, `chunk_id`,
+`chunk_index`, `validator_name`, `reason_code`, `attempt_count`, and
+`diagnostic_artifact_path` are present only when applicable.
 
-- `extractor_key`: extractor module key.
+`warnings.json` is always an object with a `warnings` array. Each warning has
-- `artifact_type`: artifact type.
+`reason_code` and `message`; `scope` is optional. Both arrays are empty when
-- `schema_version`: artifact schema version.
+there is nothing to report.
-- `payload`: artifact-type-specific JSON payload.
-- `source_refs`: optional generic source references.
-- `metadata`: optional artifact metadata.
 
-Artifact file names are produced by sanitizing the artifact type:
+## Compatibility
 
-- characters outside `A-Z`, `a-z`, `0-9`, `.`, `_`, and `-` become `_`;
+The index is the authoritative map from a logical lane to its published
-- repeated `..` sequences are replaced;
+payload. Consumers must tolerate omitted optional manifest and descriptor
-- leading and trailing `.`, `_`, and `-` are trimmed;
+fields, and should rely on the linked artifact contract for each lane’s JSON
-- empty sanitized names are rejected.
+shape. This contract describes the published logical bundle only; it does not
-
+promise a filesystem layout or expose internal state formats.
-For current D&D spell-cast artifacts, the file is
-`artifacts/dnd.spell_cast.json`.
-
-## `rejected.json`
-
-Shape:
-
-```json
-{
-  "rejected": [
-    {
-      "candidate": {
-        "index": 0,
-        "extractor_key": "dnd/spells",
-        "artifact_type": "dnd.spell_cast",
-        "schema_version": "v1",
-        "payload": {},
-        "source_refs": []
-      },
-      "validator_name": "dnd/spells/source_refs",
-      "reason_code": "missing_source_ref",
-      "message": "spell cast candidate must include at least one source ref"
-    }
-  ]
-}
-```
-
-`rejected` is an empty array when no candidates are rejected.
-
-## `warnings.json`
-
-Shape:
-
-```json
-{
-  "warnings": [
-    {
-      "scope": "output",
-      "reason_code": "example_warning",
-      "message": "warning message"
-    }
-  ]
-}
-```
-
-`warnings` is an empty array when no warnings are reported.
-
-## Path Safety
-
-The output module returns slash-separated logical paths. The CLI also validates
-logical output names before writing:
-
-- names must be non-empty;
-- names must be relative;
-- names must be clean;
-- names must use `/`, not `\`;
-- names must not contain `..`;
-- resolved paths must stay under the run output directory.
-
-Durable writes are atomic per file.

docs/integrations/openai-compatible.md

@@ -1,128 +0,0 @@
-# OpenAI-Compatible Structured Output
-
-This document describes the external LLM provider contract implemented by the
-production Notarius LLM client.
-
-## Provider
-
-- Provider key: `openai-compatible`
-- HTTP method: `POST`
-- Endpoint: `<base_url>/chat/completions`
-- Request body: JSON
-- Response mode: chat completions with structured JSON schema output
-
-`base_url` is trimmed of trailing slashes before `/chat/completions` is
-appended. Configure provider settings in [Configuration](../config.md).
-
-## Request
-
-The client sends a JSON object with:
-
-```json
-{
-  "model": "configured-model",
-  "messages": [
-    {
-      "role": "system",
-      "content": "..."
-    },
-    {
-      "role": "user",
-      "content": "..."
-    }
-  ],
-  "response_format": {
-    "type": "json_schema",
-    "json_schema": {
-      "name": "schema_name",
-      "strict": true,
-      "schema": {}
-    }
-  }
-}
-```
-
-Implemented request behavior:
-
-- `model` comes from the structured completion request when set, otherwise from
-  the configured LLM profile.
-- `messages` must be non-empty; each role and content must be non-empty after
-  trimming.
-- `response_format.type` is always `json_schema`.
-- `response_format.json_schema.strict` is always `true`.
-- `response_format.json_schema.name` and `schema` come from the extractor or
-  validator making the call.
-
-If an API key is configured, the client sends:
-
-```text
-Authorization: Bearer <api-key>
-```
-
-The client always sends `Content-Type: application/json`.
-
-## Response
-
-The client expects a JSON response with at least one choice:
-
-```json
-{
-  "model": "provider-model",
-  "choices": [
-    {
-      "message": {
-        "content": "{\"field\":\"value\"}"
-      }
-    }
-  ],
-  "usage": {
-    "prompt_tokens": 10,
-    "completion_tokens": 5,
-    "total_tokens": 15
-  }
-}
-```
-
-`choices[0].message.content` may be either:
-
-- a JSON string whose contents are valid JSON; or
-- raw JSON.
-
-The decoded content is unmarshaled into the caller-provided structured output
-target. If `usage` is present, prompt, completion, and total token counts are
-copied into the completion response.
-
-## Errors And Retries
-
-The client validates base URL, model, response schema name, response schema
-JSON, messages, and output target before or during the call.
-
-Retryable failures:
-
-- HTTP request failure;
-- response body read failure;
-- HTTP `429`;
-- HTTP `5xx`;
-- malformed provider response envelope;
-- missing choices;
-- missing, empty, or invalid assistant JSON content;
-- structured-output decode failure.
-
-Non-retryable provider status codes include non-`429` `4xx` responses.
-
-Provider error bodies are parsed for `error.message` or `message` when present.
-Configured API key values and bearer-token values are redacted from returned
-provider errors.
-
-## Timeouts And Concurrency
-
-The configured profile timeout is applied per provider request when greater
-than zero. Context cancellation is respected.
-
-The production CLI wraps the provider client with the LLM scheduler. Effective
-concurrency is described in [LLM runtime internals](../internal/llm.md).
-
-## Limits
-
-This contract documents only the fields the implemented client sends and reads.
-Provider-specific extensions are ignored unless they affect those fields.

docs/integrations/pkg-promptkit.md

@@ -0,0 +1,106 @@
+# PromptKit Integration
+
+Notarius pins
+[`gitea.maximumdirect.net/eric/promptkit` v0.5.0](https://gitea.maximumdirect.net/eric/promptkit/src/tag/v0.5.0)
+as its in-process prompt engine. The upstream
+[Go package consumer guide](https://gitea.maximumdirect.net/eric/promptkit/src/tag/v0.5.0/docs/consumers/pkg-promptkit.md)
+owns the public engine API, and the upstream
+[format reference](https://gitea.maximumdirect.net/eric/promptkit/src/tag/v0.5.0/docs/formats.md)
+owns prompt, profile, and schema file contracts.
+
+## Supported Boundary
+
+Notarius relies on the root `promptkit` package to:
+
+- construct an `Engine` with filesystem-backed prompt, schema, and optional
+  operator and application-fallback profile sources;
+- prepare one frozen execution from a `RunRequest` with named inline artifacts,
+  variables, a direct session ID, prompt identity, and profile selection, then
+  record credential-redacted details and run that exact execution;
+- return rendered debug material, validated structured output, selected
+  profile, backend, effective model metadata, and token usage;
+- register the optional conventional `local` backend through `BackendLocal`,
+  `LocalBackend`, and `WithBackend`;
+- distinguish structured-output validation failure from execution failure; and
+- identify a missing explicit profile through `ErrProfileNotFound` and backend
+  admission exhaustion through `ErrCapacityExceeded`.
+
+The pinned
+[`BackendLocal`, `LocalBackend`, and `WithBackend` API](https://gitea.maximumdirect.net/eric/promptkit/src/tag/v0.5.0/backends.go)
+owns the registration and backend-capacity contract.
+
+For one completion, the adapter calls `PrepareExecution`, takes a
+caller-owned `Details` snapshot, and calls `RunPrepared` for that same opaque
+prepared execution. It defers `Discard` for every unexecuted handle. Explicit
+profile preflight uses `Engine.InspectProfile`; it does not prepare a synthetic
+prompt. PromptKit's prepared handle, inspection result, and capacity-error
+types stay inside the Notarius LLM adapter.
+
+When a PromptKit profile and runtime override leave `temperature`, `max_tokens`,
+or `top_p` unset, Notarius leaves that control unset as well. Compatible
+providers therefore apply their own defaults; an operator that requires a
+specific sampling value must select it explicitly in the profile or runtime
+override.
+
+Notarius does not use PromptKit's optional `ArtifactReader`. It materializes
+source and reference content itself and supplies owned inline artifacts at the
+adapter boundary. It also retains responsibility for pipeline retries,
+scheduling, debug persistence, redaction, profile provenance, and conversion
+from private model responses into durable domain artifacts.
+
+Notarius sends its trimmed run session through PromptKit's direct session
+field, which is authoritative for provider session behavior. It also retains
+the same value as the `session_id` prompt variable for maintained prompt
+compatibility. Session IDs are stable, non-secret correlation identifiers and
+may be exposed to providers and provider observability.
+
+Notarius records PromptKit's selected backend ID and effective reasoning
+setting as optional run-manifest provenance. Endpoint-only profiles have no
+backend ID. Debug prompt material also retains the selected backend ID and
+PromptKit's stable lower-case `effective_model_params` JSON, which may include
+`backend_id`. Notarius production configuration exposes one optional
+conventional `local` registration. It does not expose a general user-defined
+PromptKit backend registry. Endpoint-only profiles remain supported unchanged.
+
+Notarius retains its application-wide scheduled client around the PromptKit
+adapter. PromptKit may apply a narrower limit for the selected backend;
+endpoint-only profiles have no such backend limit. The adapter translates
+PromptKit capacity rejection into the provider-neutral Notarius
+`ErrLLMCapacityExceeded` contract. It may include the normalized selected
+backend ID in safe diagnostic context, without exposing PromptKit's capacity
+error type, and leaves retries to the calling pipeline stage.
+
+## Profile Sources And Compatibility
+
+Notarius gives PromptKit the configured operator profile source, registered
+application fallback profile assets, and optional backend registration through
+the same construction path for inspection and execution. PromptKit owns the
+resulting source precedence and strict profile parsing: a matching operator
+profile is a complete replacement for a fallback or built-in profile, while an
+invalid matching document fails instead of falling through. The operator
+configuration and deployment workflow are defined in
+[Configuration](../config.md#promptkit-profiles) and
+[Operations](../operations.md#promptkit-profile-deployment).
+
+Notarius supports this boundary against PromptKit v0.5.0. Its fallback source,
+prepared-execution, inspection, and typed capacity APIs are used as public
+upstream contracts; other PromptKit APIs or file-format behavior are not
+implicitly supported. A dependency upgrade requires reviewing the adapter,
+profile-source construction, and this compatibility statement against the
+pinned upstream documentation.
+
+## Notarius Ownership
+
+[LLM Runtime Internals](../internal/llm.md) describes how Notarius mounts
+module assets, maps its transport-neutral completion contract, prepares and
+executes requests, validates output, records provenance, captures debug
+material, redacts errors, and preserves timeout ownership.
+[D&D Module Internals](../internal/dnd.md) owns the embedded
+`dnd-extraction` fallback profile and the maintained D&D prompt defaults.
+[Configuration](../config.md#promptkit-profiles) defines how a Notarius
+configuration selects one PromptKit profile source and optionally registers
+the conventional local backend.
+
+PromptKit API or format changes outside this boundary are not implicitly
+supported. Updating the pinned version requires reviewing the adapter and
+profile/configuration contracts against the upstream documentation.

docs/integrations/run-result.md

@@ -0,0 +1,68 @@
+# Run Result Receipt
+
+`notarius run --json` writes this receipt to standard output when a run
+completes successfully. It lets a subprocess caller discover the physical root
+of the published output bundle without parsing interactive command output.
+Command syntax, streams, and exit statuses are defined in the
+[CLI reference](../cli.md); logical files within the bundle are defined in the
+[Published JSON Output contract](json-output.md).
+
+## Schema
+
+The current schema version is `notarius.run-result.v1`.
+
+| Field | Required | Meaning |
+| --- | --- | --- |
+| `schema_version` | Yes | Exactly `notarius.run-result.v1`. |
+| `run_id` | Yes | The finalized Notarius run identifier. |
+| `pipeline_id` | Yes | The effective pipeline identifier. |
+| `output_directory` | Yes | Absolute path to the published, run-specific output bundle. |
+| `index_file` | For the production JSON output | Logical path `index.json`; omitted for other output modules. |
+| `normalized_output_count` | Yes | Number of final normalized outputs. |
+| `rejected_output_count` | Yes | Number of recorded rejected outputs. |
+| `warning_count` | Yes | Number of final run warnings. |
+| `validation_status` | Yes | The final run manifest validation status. |
+| `debug_directory` | No | Absolute path to the run-specific debug bundle when requested debug capture completed. |
+
+For the production `json` output module, `index_file` is present only when the
+completed run returned exactly one logical output file named `index.json`.
+For another output module, its absence does not indicate a failed run.
+
+```json
+{
+  "schema_version": "notarius.run-result.v1",
+  "run_id": "run-1770000000000000000-0123456789abcdef0123456789abcdef",
+  "pipeline_id": "dnd-session",
+  "output_directory": "/work/results/run-1770000000000000000-0123456789abcdef0123456789abcdef",
+  "index_file": "index.json",
+  "normalized_output_count": 6,
+  "rejected_output_count": 2,
+  "warning_count": 1,
+  "validation_status": "rejected"
+}
+```
+
+## Paths And Bundle Discovery
+
+`output_directory` and `debug_directory`, when present, are lexical absolute
+paths. They identify the paths used by Notarius and do not resolve symlinks.
+`output_directory` is the run-specific bundle, not the configured output root.
+
+The receipt is a summary and discovery document. It does not contain lane
+descriptors, payloads, manifest data, rejections, warnings, or file contents.
+For the production JSON output, resolve `index_file` beneath
+`output_directory`, reject path escapes, and use the
+[Published JSON Output contract](json-output.md) to discover logical files and
+lane payloads.
+
+## Delivery And Compatibility
+
+Notarius writes the receipt only after the output bundle has been published and
+any requested debug terminal reporting has completed. Standard output is not
+transactional: a result-write failure returns a nonzero status and can leave
+partial bytes. Consumers must ignore standard output unless the process exits
+with status 0.
+
+Future versions may add optional fields to this schema. Consumers must tolerate
+unknown fields. An incompatible field or semantic change requires a new
+`schema_version` value.

docs/integrations/seriatim.md

@@ -1,120 +1,73 @@
-# Seriatim Transcript JSON
+# Seriatim Transcript Input
 
-This document is the external input contract for the implemented `seriatim`
+This document defines the JSON transcript accepted by the production Seriatim
-input adapter.
+input adapter. It is a source input, not a durable lane artifact. Configure the
+input adapter through [Configuration](../config.md#production-module-keys).
 
-## Adapter
+## Contract Identity
 
-- Module key: `seriatim`
+| Property | Value |
-- Document kind: `transcript`
+| --- | --- |
-- Unit kind: `transcript_segment`
+| Consumer | Seriatim input adapter |
-- Source format: `application/vnd.seriatim+json`
+| Media type | `application/vnd.seriatim+json` |
-
+| Source document kind | `transcript` |
-The adapter parses raw Seriatim JSON into a generic source document. It owns
+| Source-unit kind | `transcript_segment` |
-transcript-specific JSON parsing and metadata mapping; core source and pipeline
-code stay source-format agnostic.
 
 ## Accepted Shape
 
-The input must be one JSON object with top-level `metadata` and `segments`
+The input is one JSON object containing `metadata` and a non-empty `segments`
-fields. This covers the maintained minimal fixture and Seriatim intermediate
+array. This minimal document is valid:
-output that provides the same required segment fields.
 
 ```json
 {
-  "metadata": {
+  "metadata": {"id": "session-alpha"},
-    "id": "session-alpha",
-    "title": "Synthetic D&D spell session"
-  },
   "segments": [
     {
-      "id": "seg-001",
+      "id": 1,
       "start": 0,
       "end": 4,
       "speaker": "Aria",
-      "text": "Aria raises her holy symbol and casts Cure Wounds."
+      "text": "Aria casts Cure Wounds."
     }
   ]
 }
 ```
 
-The maintained example is
+The maintained two-segment input is
-[examples/seriatim-minimal-transcript.json](../../examples/seriatim-minimal-transcript.json).
+[seriatim-minimal-transcript.json](../../examples/seriatim-minimal-transcript.json).
 
-Top-level metadata entries are preserved. Other segment fields, such as
+| Field | Required | Meaning and constraints |
-`categories`, are ignored.
+| --- | --- | --- |
+| `metadata` | Yes | JSON object. Its entries become source metadata; no particular metadata key is otherwise required. |
+| `segments` | Yes | Non-empty array of segment objects, kept in input order. |
+| `segments[].id` | Yes | Positive canonical decimal integer, supplied as a JSON number or string. IDs must be unique. |
+| `segments[].start` | Yes | Finite, non-negative numeric value, supplied as a JSON number or string. |
+| `segments[].end` | Yes | Finite, non-negative numeric value that is not earlier than `start`. |
+| `segments[].speaker` | Yes | String that is non-empty after trimming. |
+| `segments[].text` | Yes | String that is non-empty after trimming. Its original text is retained. |
 
-Multiple top-level JSON values are rejected.
+Additional top-level and segment fields are ignored. A missing required field,
+`null` in place of an object or array, malformed JSON, or more than one
+top-level JSON value is rejected.
 
-## Validation
+## Source Identity And References
 
-The adapter rejects:
+The adapter chooses the source ID in this order:
 
-- empty raw input;
+1. a non-empty source ID supplied by the calling request;
-- malformed JSON;
+2. non-empty string `metadata.id`;
-- top-level JSON that is not an object;
+3. non-empty string `metadata.source_id`;
-- missing, null, or non-object `metadata`;
+4. `seriatim:` followed by the first 16 hexadecimal characters of the raw
-- missing, null, non-array, or empty `segments`;
+   input’s SHA-256 digest.
-- segment values that are not objects;
-- segment `id` values that are neither strings nor numbers;
-- non-string `speaker` or `text`;
-- empty segment IDs;
-- segment IDs with leading or trailing whitespace;
-- duplicate segment IDs;
-- missing or empty `speaker`;
-- missing, empty, invalid, non-finite, or negative `start`;
-- missing, empty, invalid, non-finite, or negative `end`;
-- `end` values before `start`;
-- missing or empty `text`.
 
-Segment text is preserved as provided, but it must not be empty after trimming.
+Each accepted segment becomes one source unit whose unit ID is `segments[].id`.
+Its self-reference uses the derived source ID and the same segment ID for both
+range endpoints. Artifact contracts use those segment IDs when they cite
+transcript evidence.
 
-## Source Mapping
+## Compatibility
 
-The adapter maps input to `SourceDocument`:
+This adapter accepts only the shape described here. A broader Seriatim export
-
+is usable only when it supplies this object, metadata, and segment shape with
-- `metadata` becomes `SourceDocument.Metadata`;
+the stated types and constraints. Unknown additional fields do not add
-- `SourceDocument.Kind` is `transcript`;
+Notarius behavior.
-- `SourceDocument.Format` is `application/vnd.seriatim+json`;
-- `SourceDocument.Digest` is `sha256:<hex>` of the exact raw input bytes.
-
-`SourceDocument.ID` is selected in this order:
-
-1. the parse request source ID, after trimming;
-2. `metadata.id`, when it is a non-empty string after trimming;
-3. `metadata.source_id`, when it is a non-empty string after trimming;
-4. `seriatim:<first-16-hex-chars-of-raw-sha256>`.
-
-Each segment becomes one `SourceUnit`:
-
-- `segment.id` becomes `SourceUnit.ID`; numeric IDs are converted to their JSON
-  number text, so `1` becomes `"1"`;
-- `segment.text` becomes `SourceUnit.Text`;
-- `SourceUnit.Kind` is `transcript_segment`;
-- `speaker`, `start`, and `end` are stored in source-unit metadata.
-
-## Metadata Keys
-
-Seriatim unit metadata uses these keys:
-
-- `speaker`: string speaker label;
-- `start`: `json.Number` start value;
-- `end`: `json.Number` end value.
-
-The `internal/modules/input/seriatim` package exposes typed accessors for these
-values.
-
-## Capabilities
-
-The module declares these provided capabilities:
-
-- `source.transcript`
-- `transcript.speaker`
-- `transcript.timestamps`
-
-## Compatibility Limit
-
-This contract covers only Seriatim transcript JSON with the top-level
-`metadata` object and `segments` array described here. Broader Seriatim output
-schemas are compatible only when they provide these required fields with the
-accepted types.

docs/internal/cli.md

@@ -0,0 +1,163 @@
+# CLI Internals
+
+This document describes **internal/cli**, Notarius's production composition
+root. The [CLI reference](../cli.md) owns command syntax and exit statuses;
+[Configuration](../config.md) owns configuration values; and
+[Operations](../operations.md) owns filesystem layout, recovery, and operator
+procedures.
+
+## Inputs, Outputs, And Boundaries
+
+The CLI accepts process arguments, standard streams, and injectable options
+used by tests and embedding code. It writes command results to the supplied
+streams and returns a process exit status. For a run, it also creates the
+production catalog and runtime collaborators, hands a prepared pipeline and
+source bytes to the framework, and places the logical files returned by the
+runner.
+
+It is the only boundary allowed to compose concrete registries, LLM clients,
+cache/checkpoint collaborators, debug recorders, and physical output paths.
+Pipeline modules receive interfaces and request data rather than CLI streams or
+filesystem roots. The [Architecture](../policy/architecture.md) defines this
+composition-root boundary; [Pipeline Internals](pipeline.md) owns resolution,
+preparation, and runner mechanics after their inputs are supplied.
+
+## Dispatch And Configuration Handoff
+
+The root dispatcher handles help, configuration validation, pipeline listing,
+and a pipeline run. It normalizes injectable options before dispatch so that a
+missing production dependency fails as a command error rather than reaching
+execution.
+
+Commands that need configuration use one shared loader. The CLI discovers the
+file, parses it through **internal/core/config**, starts from defaults, applies
+the file and supported environment overrides, and then validates it for the
+command. The configured discovery and precedence contract is in
+[Configuration](../config.md), while the loading and resolution mechanics are
+in [Configuration Internals](configuration.md).
+
+Configuration validation without a selected pipeline checks structural
+configuration only. Validation with a selected pipeline also builds the
+effective catalog, resolves the pipeline, and verifies every explicit effective
+PromptKit profile. Selected LLM-backed input, chunk, lane, output, and validator
+profiles are inspected
+against the configured PromptKit source and backend registrations without
+loading a prompt or performing generation, so an unknown or invalid profile
+fails before pipeline preparation. Credential availability remains an
+execution-time concern. Pipeline listing validates configuration before
+returning normalized, sorted identifiers.
+
+## Production Composition
+
+The production composition helper allocates every framework registry and the
+prompt-asset registry, then registers the generic, Seriatim, and D&D module
+families in that order. The resulting registries provide both the module
+catalog used for resolution and the concrete constructors used for preparation.
+Tests may provide a catalog or registries instead; production code must not
+silently merge an injected partial catalog with production registrations.
+
+The production LLM factory builds one PromptKit-backed client from the resolved
+**promptkit.profile_dir** or **promptkit.profile_file** source, attaches the
+profile-provenance recorder, creates one scheduler from the effective global
+LLM limit, and wraps the client before it reaches modules. Registration and LLM
+construction errors are returned before a pipeline is prepared. Configuration
+field definitions remain in [Configuration](../config.md#promptkit-profiles);
+the D&D registrar's fallback profile assets and the adapter mechanics remain in
+[LLM Runtime](llm.md).
+
+The factory also accepts `LLMRuntimeOverrides`, whose reasoning pointer
+preserves inherit, replace, and clear states across the composition boundary.
+Run orchestration constructs this value from the mutually exclusive
+`--reasoning-effort` and `--clear-reasoning-effort` controls. Absence preserves
+a nil pointer, replacement is trimmed, and clear uses a non-nil empty string.
+The same override reaches the one shared production client, checkpoint
+identity, and debug invocation metadata. Persistent reasoning configuration
+remains owned by PromptKit profiles; Notarius configuration has no reasoning
+field.
+
+## Run Orchestration
+
+After parsing and validating a run invocation, the CLI performs this ordered
+handoff:
+
+1. load and validate configuration, then apply command-level operational
+   overrides;
+2. create and validate a safe run identity, then allocate a debug bundle only
+   when requested;
+3. build the effective catalog, resolve requested reference changes, resolve
+   the effective pipeline, and inspect its explicit effective PromptKit
+   profiles;
+4. materialize external or generated references and record redacted invocation
+   and resolution provenance when debug capture is enabled;
+5. construct registries, the scheduled LLM client, prepared modules, and the
+   requested cache/checkpoint collaborators;
+6. read the source input and invoke the framework runner; and
+7. write the runner's logical output files only after a successful run, then
+   complete the command report and user-facing result.
+
+Preparation happens before source parsing, so module construction and
+dependency failures cannot begin stage execution. The CLI also preserves the
+framework's result and warning information when it writes summaries and the
+final command result. Detailed state lifecycle, resume handling, and physical
+path confinement are maintained in [Run State Internals](state.md) and
+[Operations](../operations.md).
+
+For `run --json`, the CLI constructs and encodes its private run-result receipt
+after a successful runner result is available, before it publishes logical
+output files. It writes the prepared receipt to standard output only after
+output publication and requested debug terminalization succeed. A receipt-write
+failure exits with runtime status 1 and may leave partial standard-output bytes,
+but the already-published output bundle remains complete and requested debug
+reporting remains successfully terminalized. The CLI reports a bounded
+command-owned error and does not repeat terminal reporting. The receipt remains
+a CLI reporting concern rather than a framework or output-module responsibility;
+its public contract is the
+[run-result receipt](../integrations/run-result.md).
+
+## Failure Mapping And Terminal Reporting
+
+Argument, flag, and invocation-combination failures are reported to standard
+error before runtime composition and use the syntax error class. Once an
+invocation is syntactically valid, configuration loading and validation,
+resolution, registration, profile checks, reference materialization, module
+construction, input reads, runner failures, output publication, and requested
+debug handling use the runtime failure class. The public status numbers and
+stream contract are defined in the [CLI reference](../cli.md#output-streams-and-exit-statuses).
+
+When debug capture has been allocated, one command-state value records the
+known run result. Guarded terminalization writes a success report once, or
+attempts a failure report and error record once. A persistence failure is
+reported in addition to the original failure and never replaces it. If a debug
+path exists, failure output includes that path so the retained diagnostic data
+is discoverable.
+
+## Invariants To Preserve
+
+- Only the CLI composes production implementations and physical runtime roots.
+- Configuration and resolved composition failures occur before module
+  preparation or source parsing.
+- A runner's logical files are published only after a successful run.
+- Production registries and a caller-supplied catalog or registries are
+  alternative composition sources, not an implicit mixture.
+- A requested debug bundle has one terminal report attempt; its persistence
+  errors supplement rather than obscure the primary command error.
+- User-facing flags, paths, exit codes, and configuration fields are defined
+  by their public documentation, not duplicated here.
+
+## Focused Tests
+
+- **internal/cli/command_contract_test.go** covers dispatch, help, syntax and
+  runtime error classes, discovery, validation, and listing.
+- **internal/cli/run_contract_test.go** covers the run handoff, publication,
+  debug reporting, and command-owned state collaborators.
+- **internal/cli/production_contract_test.go** covers registrar composition,
+  production catalog contents, assets, and representative configuration
+  validation.
+- **internal/cli/reference_contract_test.go** covers CLI reference overrides,
+  origin separation, and materialization boundaries.
+- **internal/cli/state_hardening_test.go** covers safe run identity, state
+  roots, and failure ordering.
+
+Run **go test ./internal/cli** after changing command composition or command
+behavior. Pair it with **go test ./internal/core/config** when the configuration
+handoff changes.

docs/internal/configuration.md

@@ -0,0 +1,144 @@
+# Configuration Internals
+
+This document describes the maintainer-facing configuration boundary in
+**internal/core/config**. The [Configuration](../config.md) reference owns the
+file format, fields, defaults, precedence contract, and selectable keys. The
+[CLI reference](../cli.md) owns command syntax; this document does not redefine
+either interface.
+
+## Boundary
+
+The configuration package turns a selected YAML file and supported environment
+values into a validated, independently owned configuration. It then resolves a
+requested pipeline against a module catalog before the framework prepares or
+runs anything.
+
+| Boundary | Inputs | Outputs | Does not own |
+| --- | --- | --- | --- |
+| Loading | Selected file path and environment lookup | Parsed file model and a populated **Config** | Choosing the file path or reporting a command result. |
+| Validation | **Config** | Structural configuration errors with pipeline, lane, or binding context | Module availability, capabilities, or construction. |
+| Resolution | Valid **Config**, selected pipeline and lanes, runtime reference changes, LLM override, and module catalog | **EffectiveConfig** with a **ResolvedPipeline** | Materializing reference bytes, preparing modules, execution, or filesystem state. |
+| Summary | **Config** or **EffectiveConfig** | Detached redacted payload suitable for debug summaries | Redacting arbitrary process state or provider traffic. |
+
+The CLI discovers a configuration file, invokes this package, and supplies the
+result to the framework. Configuration never reads an input file, constructs a
+module, or creates output, cache, or debug paths. Those responsibilities remain
+at their respective [CLI](cli.md), [pipeline](pipeline.md), and
+[run-state](state.md) boundaries.
+
+## Loading And Validation
+
+The CLI loads configuration in this order:
+
+1. parse the selected YAML file strictly into the file model;
+2. start from **Default**;
+3. apply the file model; and
+4. apply the supported environment overrides.
+
+This establishes the public precedence order without giving environment input a
+second file schema. Loading and application reject malformed YAML, unsupported
+file versions, unknown fields, invalid values, and identifiers that are empty
+or collide after whitespace normalization. The file application also makes the
+effective extraction-worker default follow the effective LLM limit. A present
+PromptKit local-backend object requires and trims its endpoint, defaults its
+omitted concurrency limit to zero, and is copied so the parsed file model
+cannot alias the populated **Config**. A pipeline `llm_profile` is
+presence-aware: omission remains empty, while a present blank value is
+rejected and a non-empty file value is trimmed before it reaches **Config**.
+
+**Config.Validate** checks configuration-only invariants before resolution. It
+rejects incompatible profile sources, invalid state-surface values, unsupported
+concurrency settings, malformed bindings and references, invalid retries, and
+invalid pipeline, step, or lane structure. PromptKit local-backend validation
+accepts only an absolute HTTP or HTTPS endpoint with a host and no user
+information, query, or fragment, and rejects a negative local concurrency
+limit. Its errors retain the closest known pipeline, lane, and binding context.
+It deliberately does not require modules to be registered: that requires a
+catalog and belongs to resolution.
+
+The exact user-selectable values and validation rules are defined in
+[Configuration](../config.md). Keep additions to the file model, an
+environment override, its validation, and that reference in the same change.
+
+## Effective Resolution
+
+**Config.Resolve** first recomputes derived concurrency defaults and validates
+the configuration. It normalizes the requested pipeline ID, copies the selected
+profile, and passes the non-empty command-level LLM profile override, requested
+lane selection, and reference changes to the framework resolver.
+
+After module and validator selection, the resolver applies the effective
+profile policy to LLM-backed bindings only: command override, binding profile,
+pipeline profile, then the prompt default. Deterministic bindings remain
+profile-free, and no second inheritance decision occurs during execution. The
+public field definitions and precedence are owned by
+[Configuration](../config.md#pipelines).
+
+The framework resolver supplies defaults, selects lanes, resolves validator
+chains, checks registered module and artifact compatibility, validates module
+options, and returns the fixed ordered pipeline shape. The resulting
+**EffectiveConfig** retains the selected ID, requested selection and reference
+changes, a clone of the input configuration, and the resolved pipeline.
+Callers may therefore retain or modify their input slices and maps without
+changing the resolved result, and later consumers cannot mutate the original
+configuration through the effective value. This ownership includes the nested
+PromptKit local-backend value.
+
+Resolution failures stop before module construction and source parsing. They
+include an error path for an unconfigured pipeline, missing module, missing
+capability, incompatible artifact variant, invalid option, invalid reference,
+or invalid lane selection. CLI code maps these valid-invocation failures to the
+runtime error class described in the [CLI reference](../cli.md#output-streams-and-exit-statuses).
+
+## Resolved Identity And Redaction
+
+The framework assigns the resolved pipeline a deterministic SHA-256 digest
+after defaults, lane selection, module bindings, reference bindings, validator
+chains, effective LLM profiles, and artifact schema identity have been
+resolved. The digest excludes
+its own stored value. It identifies resolved composition rather than raw YAML
+bytes, a debug payload, or all runtime state. The CLI records it as invocation
+provenance before execution; cache and checkpoint identity have additional
+owners in [Run State Internals](state.md).
+
+Configuration summaries must use **Redacted**, **RedactedSummaryPayload**, or
+**RedactedResolvedPipelinePayload**, never a direct configuration marshal.
+Those methods copy every binding and nested option container, replace values
+whose key is credential-shaped with **[REDACTED]**, and omit materialized
+reference content while retaining safe binding and reference provenance. The
+payload must not alias the source configuration or resolved pipeline.
+PromptKit's local endpoint and concurrency limit are preserved as non-secret
+configuration metadata in the independently owned summary; the object contains
+no credential value. This redaction is deliberately narrow: it protects
+configuration summaries and does not authorize recording arbitrary environment
+values or provider requests.
+
+## Invariants To Preserve
+
+- Defaults, YAML values, and environment values are applied in one direction;
+  later sources may override only their supported operational settings.
+- A configuration is structurally valid before it is resolved, and a resolved
+  pipeline is compatible with the supplied catalog before preparation begins.
+- Whitespace-normalized identifiers are unique wherever they identify a
+  pipeline, step, lane, worker, or reference slot.
+- Resolution and summary generation return detached data. Redaction must cover
+  every configured and resolved binding, including nested validator bindings.
+- The resolved digest changes when resolved composition changes and never
+  includes itself.
+
+## Focused Tests
+
+- **internal/core/config/file_config_contract_test.go** covers strict file
+  parsing, normalization, file application, and structural rejection.
+- **internal/core/config/env_contract_test.go** covers supported operational
+  overrides and their precedence.
+- **internal/core/config/validation_contract_test.go** covers configuration
+  invariants and contextual failures.
+- **internal/core/config/effective_config_contract_test.go** covers defaults,
+  selections, overrides, resolution context, digest changes, and ownership.
+- **internal/core/config/redaction_test.go** covers recursive credential
+  redaction, reference-content exclusion, and non-aliasing payloads.
+
+Run **go test ./internal/core/config** after changing this boundary. Changes to
+the handoff or resolved-composition semantics also need the focused framework
+pipeline tests.

docs/internal/diagnostics.md

@@ -1,88 +0,0 @@
-# Diagnostics Internals
-
-Diagnostics internals live in `internal/core/diagnostics`. Operator-facing run
-behavior is documented in [Operations](../operations.md).
-
-## Purpose
-
-Diagnostics provide local inspection artifacts for a run without becoming the
-durable output contract. Durable user output is produced by output modules and
-written by the CLI.
-
-Diagnostics must not expose secrets.
-
-## Run Directory
-
-`NewRunDirectory(workDir, retention)` creates:
-
-```text
-<workDir>/run-<unix-nanoseconds>/
-```
-
-If `workDir` is empty, it defaults to `/tmp/notarius`. Empty retention defaults
-to `auto`.
-
-The writer makes the work directory if needed, then attempts to create a unique
-run directory. It retries run ID creation a bounded number of times if a
-collision occurs.
-
-## Artifact Writers
-
-Implemented artifact names:
-
-- `invocation.json`
-- `effective-config.json`
-- `resolved-pipeline.json`
-- `source-document.json`
-- `run-manifest.json`
-- `run-report.json`
-- `warnings.json`
-- `error.log`
-
-JSON artifacts are encoded with indentation and a trailing newline. Writes are
-atomic through a temporary file in the target directory followed by rename.
-
-Artifact names must be single relative file names. Absolute paths, path
-separators, and names resolving outside the run directory are rejected.
-
-## Redacted Effective Config
-
-Diagnostics writers accept payloads that implement
-`RedactedDiagnosticsPayload`. `internal/core/config` uses this to redact API
-keys in effective config diagnostics while preserving resolved pipeline context.
-
-The redaction path clones config data before replacing secret values.
-
-## Retention
-
-Retention is decided by `ShouldRetainRunDirectory`.
-
-- Failed runs are always retained.
-- `always` retains successful runs.
-- `never` removes successful runs.
-- `auto` retains successful runs only when warnings exist.
-- Unknown retention values are treated as retain by the retention decision, but
-  config validation rejects unsupported values before normal runs.
-
-`ApplyRetention` removes only the specific run directory.
-
-## CLI Failure Behavior
-
-The CLI creates the diagnostics run directory after config loading and before
-pipeline resolution. Failures before that point do not have diagnostics.
-
-After diagnostics creation, run failures call `WriteErrorLog` and apply
-retention with `RunSucceeded: false`, so the run directory remains available.
-
-When the pipeline returns a partial manifest on failure, the CLI writes that
-manifest before logging the failure.
-
-## Invariants
-
-- Diagnostics paths must be narrow and run-directory scoped.
-- Writes should be atomic where practical.
-- Secrets must be redacted.
-- Diagnostics write failures are command failures because they can hide the
-  information needed for recovery.
-- Durable output file contracts belong to output modules and integration docs,
-  not to diagnostics.

docs/internal/dnd.md

@@ -0,0 +1,155 @@
+# D&D Module Internals
+
+This guide records the conventions shared by the production D&D module family.
+It complements [Module Internals](modules.md), which owns generic registration
+and extension mechanics, and [Configuration](../config.md), which owns the
+selectable keys, bindings, reference syntax, and default validator chains.
+
+## Durable Artifact Contracts
+
+The six lanes have separate durable wire contracts. This guide deliberately
+does not repeat their JSON shapes or schemas.
+
+| Lane | Durable contract |
+| --- | --- |
+| Spells | [spell artifacts](../integrations/dnd-spell-artifacts.md) |
+| NPCs | [NPC artifacts](../integrations/dnd-npc-artifacts.md) |
+| Combat turns | [combat-turn artifacts](../integrations/dnd-combat-turn-artifacts.md) |
+| Item events | [item-event artifacts](../integrations/dnd-item-event-artifacts.md) |
+| NPC interactions | [NPC-interaction artifacts](../integrations/dnd-npc-interaction-artifacts.md) |
+| Scene descriptions | [scene-description artifacts](../integrations/dnd-scene-description-artifacts.md) |
+
+## Family Composition
+
+The D&D registrar registers the family’s artifact codecs, extractors, typed
+append-order mergers, normalizers, validators, prompt assets, fallback LLM
+profile asset, and default validator chains. Each extractor and normalizer has
+a stable module spec, explicit execution class, strict option decoding, and a
+typed builder. Scene chunking, every extractor, and NPC normalization are
+registered as `llm_backed`; the remaining current D&D mergers and normalizers
+are `deterministic`. The metadata is available to catalog inspection and
+resolved-pipeline debug data and determines which selected bindings inherit the
+pipeline profile. Configuration remains the canonical owner of the exact keys,
+profile precedence, and validator order.
+
+Private structured-LLM response schemas are deliberately minimal. They reject
+invalid JSON structure, missing required fields, incompatible types, and
+unknown fields, while preserving semantic candidates for deterministic
+validation. Do not promote a private response envelope into a durable schema;
+the contracts above define durable data.
+
+## Prompt Construction
+
+D&D extractors assemble prompts from an ordered manifest of shared and
+module-owned assets. Reuse the shared D&D system, evidence, identity,
+reference, and transcript assets instead of copying their text into individual
+modules. A manifest’s declared sequence, including cache-control placement, is
+part of the prompt behavior.
+
+Every maintained D&D LLM prompt selects `dnd-extraction` as its default
+profile. The D&D registrar embeds that fallback profile with the maintained
+OpenRouter model, timeout, and service-tier policy. An operator may provide a
+complete profile with the same ID through the configured PromptKit source; that
+definition replaces the fallback rather than merging with it. The fallback
+leaves reasoning and optional sampling controls unspecified. Deployment profile
+selection and the maintained operator example are documented in
+[Configuration](../config.md#promptkit-profiles).
+
+All extraction prompts share this four-message rendered prefix: the system
+message without cache control, the identity message without cache control, the
+campaign-reference message with ephemeral cache control, and the chunk
+transcript message with ephemeral cache control. This gives equivalent
+extraction requests the same reusable prefix through their source material.
+
+Extraction-evidence policy, generated NPC registries, spell catalogs, module
+tasks, and instructions follow the transcript because they are not universal
+across all extraction lanes. The final instructions message carries ephemeral
+cache control; evidence, registry, catalog, and task messages do not. Preserve
+this division when changing an extractor or its assets so prompt-cache behavior
+remains stable.
+
+The other D&D LLM prompts intentionally follow different patterns. Scene
+chunking has no sibling extraction lane with which to share its full transcript,
+so it renders campaign references before its task and instructions, then places
+the cacheable full transcript last. NPC normalization keeps its task and
+cacheable instructions before the candidate collection, followed by the
+cacheable transcript windows: candidates must be available before their
+supporting evidence is evaluated, and those windows are not a cross-lane
+prefix. Mounted assets and their declared message order determine the prompt
+fingerprint, so intentional prompt edits continue to invalidate stale
+checkpoints.
+
+All extractors use the shared prompt-input preparation rules. The current chunk
+is copied into transcript material; player, party, glossary, and compatible
+campaign references are context for disambiguation, not source evidence.
+Reference prompt material is canonically ordered before it is rendered, which
+keeps equivalent inputs stable across runs.
+
+## Evidence, Candidates, And Normalization
+
+The current transcript is the only durable evidence source. Extractors assign
+the current source identity, preserve candidate evidence ranges for validators,
+and canonically order or remove exact duplicate ranges without asking the
+model to repair semantic errors. Campaign context and generated artifacts may
+ground names or control routing, but they never establish evidence for a D&D
+result.
+
+Default chains keep responsibilities separate: structural validators assess the
+candidate, source-reference validators resolve cited ranges against the current
+source, durable-schema validation checks an approved representation, and
+relatedness validators report advisory evidence concerns. The configured order
+is documented in
+[Configuration](../config.md#production-validator-keys-and-default-chains).
+
+Normalizers are deterministic for spells, combat turns, item events, NPC
+interactions, and scene descriptions. They canonicalize display values and
+evidence, use source-document order for stable output, and issue bounded
+warnings for changes or collapsed duplicates. The NPC normalizer is the
+intentional exception: it first produces a deterministic candidate set, then
+uses a bounded structured-LLM proposal to reconcile identity groups. Invalid
+or unusable proposals retain the deterministic result and surface retry or
+fallback diagnostics; the model does not directly replace durable records.
+
+## Generated References And Grounding
+
+Normalized D&D artifacts can be handed to a later step through a generated
+reference binding. The framework verifies artifact compatibility and retains
+producer provenance; consumers resolve the handed-off artifact into an
+immutable, validated projection for each operation. External files are checked
+during preparation, while generated artifacts are resolved at the handoff.
+
+NPC registries are names-only grounding projections: they may canonicalize
+actors for spells and combat turns and are required for NPC interactions, but
+they do not supply evidence. Scene-description registries are eligibility-only
+projections: they retain the current chunk’s classification data, not scene
+prose or evidence, and exist to route combat extraction.
+
+## Lane-Specific Rules
+
+The following differences are intentional and should remain explicit when a
+shared helper changes.
+
+| Lane | Intentional behavior |
+| --- | --- |
+| Spells | May use a spell-catalog overlay and optional NPC grounding; the catalog validator supplies domain-specific semantic checks. |
+| NPCs | Does not consume an NPC registry. Its normalizer is the LLM-assisted reconciliation exception described above. |
+| Combat turns | Requires a scene-description artifact. It calls the LLM only for an exact `combat` classification; exact non-combat classifications return an accepted empty result, while missing or mismatched classifications return an empty result with a bounded warning. Optional NPC grounding never becomes evidence. |
+| Item events | Uses campaign context for disambiguation but has no NPC-registry or scene-description dependency. |
+| NPC interactions | Requires the normalized NPC registry at extraction and normalization, using it for canonical actor grounding only. |
+| Scene descriptions | Produces the classifications consumed by combat routing; it does not consume an NPC registry or provide evidence for combat artifacts. |
+
+The combat and scene-description contracts describe their exact handoff and
+empty-result behavior in more detail:
+[combat turns](../integrations/dnd-combat-turn-artifacts.md) and
+[scene descriptions](../integrations/dnd-scene-description-artifacts.md).
+
+## Focused Verification
+
+When changing D&D behavior, test the affected codec, extractor, normalizer,
+validator, prompt-asset manifest, and registry projection. Also test generated
+handoffs at the integration boundary and run the full D&D module suite:
+
+~~~sh
+go test ./internal/modules/dnd/...
+go test ./internal/modules/integration/...
+~~~

docs/internal/llm.md

@@ -1,116 +1,251 @@
-# LLM Runtime
+# LLM Runtime Internals
 
-The implemented LLM runtime lives in `internal/framework/llm`. It provides
+`internal/framework/llm` is Notarius’s provider-independent structured
-transport-neutral structured completion contracts, an OpenAI-compatible HTTP
+completion boundary. It adapts framework requests to PromptKit, bounds
-adapter, concurrency scheduling, schema registry helpers, retry behavior, and
+provider calls, assembles registered prompt and schema assets, records selected
-secret redaction.
+profiles, and redacts provider errors. The architectural boundary is defined in
+[Architecture](../policy/architecture.md#llm-boundary); profile sources,
+credentials, and concurrency settings belong in
+[Configuration](../config.md#promptkit-profiles) and
+[Configuration](../config.md#concurrency-output-cache-and-debug).
 
-## Contract
+## Structured Completion Boundary
 
-Modules depend on `contracts.StructuredLLMClient`:
+Modules and LLM-backed validators depend only on
+`contracts.StructuredLLMClient`. A completion request supplies a prompt ID and
+version, optional profile and session IDs, named input material, variables, and
+a caller-owned decode target. The successful response returns the validated raw
+structured bytes together with non-secret provider, model, profile, and token
+metadata.
 
-```go
+The caller owns the domain behavior: it chooses the prompt, prepares inputs,
-CompleteStructured(ctx, request, out) (response, error)
+selects the private response schema, and interprets the decoded result. The
-```
+adapter does not own source evidence, artifact conversion, normalization, or
+durable schemas. Those responsibilities remain with the module and its
+[integration contract](../integrations/).
 
-The request contains messages, optional model override, response schema name,
+`PromptKitClient` validates the request target and prompt identity, maps each
-and response schema JSON. The caller supplies a pointer target for decoded
+named material to a PromptKit inline artifact while preserving its origin URI,
-structured output.
+maps the trimmed request session to PromptKit's direct per-run session field,
+retains the same value as the `session_id` prompt variable for maintained
+prompt compatibility, and forwards profile selection. It then creates one
+frozen prepared execution, captures its caller-owned credential-redacted
+details for debug material, and executes that exact snapshot through
+PromptKit's prepared-execution boundary. The direct field
+is authoritative for provider session behavior. A session ID is a stable,
+non-secret correlation identifier and may be exposed to providers and provider
+observability. The adapter returns PromptKit’s validated raw bytes rather than
+re-encoding the decoded target. An empty optional material is represented as
+one space so its named input is retained by PromptKit.
 
-Extractors own prompts and schemas. Provider adapters should not contain
+Client construction may also receive a run-wide reasoning-effort override from
-domain-specific prompt logic.
+the CLI factory boundary. The adapter copies the caller-owned pointer and
+creates a fresh PromptKit execution override for each request: a nil pointer
+inherits the selected profile, a non-empty value replaces it, and an empty
+value clears inherited reasoning. The CLI's mutually exclusive
+`--reasoning-effort` and `--clear-reasoning-effort` controls select those
+states. With neither flag, profile behavior remains unchanged. Because
+production constructs one shared client, the selected state applies uniformly
+to module calls, retries, and LLM-backed validators for the whole run.
 
-## Production Client Construction
+An empty request profile lets the prompt select its configured default. Before a
+run begins, the CLI asks the adapter to inspect every explicit profile on the
+resolved selected LLM-backed bindings and validators, including inherited
+pipeline profiles. Inspection resolves the profile and its selected backend and
+target without loading a prompt, reading credentials, admitting capacity, or
+contacting a provider, so a missing or invalid explicit profile fails before
+stage execution while a valid `api_key_env` may remain unset. Calls record the
+profile actually selected by PromptKit. The recorder trims and deduplicates
+non-secret profile identity, provider, model, selected backend ID, and
+effective reasoning values for manifest use. Entries that differ in backend or
+reasoning remain distinct and deterministically ordered. Endpoint-only profiles
+retain an empty backend ID, which the published JSON omits. Successful
+completion responses and recorded profile manifests identify the adapter
+provider as `promptkit`.
 
-`internal/cli` builds the production LLM client from the effective config:
+The CLI's profile-inspection engine and the production adapter use the same
+profile-source construction to apply the configured profile directory or file,
+the optional registered fallback profile assets, and the optional conventional
+`local` backend. Preflight therefore resolves the same profile sources and
+backend membership as runtime without performing generation. Fallback assets
+are mounted only when at least one source is registered. The production D&D
+registrar contributes its `dnd-extraction` fallback, and the maintained D&D
+prompts select that logical ID by default. PromptKit owns source precedence and
+profile parsing: an operator-provided matching profile takes precedence over a
+fallback profile without Notarius merging either document.
+When the registration is absent, a profile selecting `backend: local` fails
+inspection instead of falling back to a built-in or endpoint-only target.
 
-1. find the effective LLM profile;
+Before execution, the adapter also contributes a non-secret checkpoint
-2. build `OpenAICompatibleClientConfig`;
+fingerprint for the effective PromptKit profile source. It combines the
-3. create an OpenAI-compatible client;
+identity of PromptKit's compiled-in profile catalog with a deterministic digest
-4. create a scheduler from profile or global concurrency;
+of every YAML profile in the configured profile directory, or of the configured
-5. wrap the client with `NewScheduledClient`;
+profile file, and a deterministic digest of the flattened fallback profile
-6. return non-secret LLM profile manifest metadata.
+assets. The fingerprint contains neither profile content nor source paths. It
+covers inherited pipeline profiles, explicit binding profiles, and
+prompt-selected defaults, so changing a model or other profile setting cannot
+reuse checkpoints created under the
+prior profile source. This cache identity is independent of durable
+profile provenance: run manifests continue to list only profiles actually
+observed during LLM calls. When the local backend is registered, a second
+fingerprint hashes its trimmed endpoint behind a stable marker. Changing that
+semantic execution target invalidates checkpoint reuse. The raw endpoint is not
+stored in checkpoint identity, and the local concurrency limit is excluded
+because it changes scheduling rather than execution semantics.
 
-The current run command requires exactly one distinct effective LLM profile for
+## Shared Provider-Call Limit
-the resolved pipeline.
 
-## OpenAI-Compatible Adapter
+Production construction creates one PromptKit client and wraps it in one
+scheduled client. The scheduler has a fixed, positive permit limit, serves
+queued calls in FIFO order, and removes a queued call when its context is
+cancelled. A granted permit is released exactly once on every completion path.
 
-`OpenAICompatibleClient` posts JSON to:
+The scheduled wrapper surrounds every `CompleteStructured` call, so concurrent
+lanes, pipeline retries, and LLM-backed validators share the same provider-call
+ceiling. This ceiling is independent of pipeline worker concurrency; changing
+worker counts cannot exceed the configured LLM limit. The configuration field
+and its effective default are owned by
+[Configuration](../config.md#concurrency-output-cache-and-debug).
 
-```text
+PromptKit applies a second, independent admission limit when the selected
-<base_url>/chat/completions
+profile names a limited backend. It sits beneath the Notarius scheduled client,
-```
+so it may narrow but cannot expand the application-wide limit. Built-in
+OpenRouter profiles select PromptKit's reserved backend and its upstream
+capacity policy. A positive configured local-backend limit bounds active local
+generations inside PromptKit; zero leaves that backend unlimited there.
+Endpoint-only profiles do not select a PromptKit backend and remain limited
+only by the Notarius scheduler.
 
-It sends:
+## Prompt And Schema Assets
 
-- `model`
+An `AssetRegistry` collects prompt, schema, and optional fallback-profile
-- `messages`
+filesystems from production module families. It flattens registered roots into
-- `response_format.type = "json_schema"`
+the corresponding PromptKit filesystems and rejects invalid roots, unreadable
-- `response_format.json_schema.name`
+assets, duplicate paths, and missing prompt or schema files during preparation.
-- `response_format.json_schema.strict = true`
+Fallback assets receive a safe content digest for checkpoint identity; raw
-- `response_format.json_schema.schema`
+paths and bytes are never included. The framework’s `promptfs` helper combines
+module-owned prompt files with reusable domain fragments without making the
+framework depend on D&D content.
 
-If an API key is configured, the adapter sends an `Authorization: Bearer ...`
+Each LLM-backed module owns its prompt declaration, package-specific assets,
-header.
+and private response schema. Shared D&D wording is owned by the D&D shared
+asset package; the detailed D&D conventions are in
+[D&D Module Internals](dnd.md). The mounted prompt assets used by a module also
+determine its prompt fingerprint. Schema loaders validate JSON, attach identity
+and digest metadata, make defensive copies, and expose diagnostics without raw
+schema bytes.
 
-The adapter accepts assistant content either as a JSON string containing JSON or
+Private response schemas validate a model transport envelope. They are not the
-as raw JSON content. It then unmarshals that content into the caller-provided
+durable artifact schema and should not be documented as an external wire
-target.
+contract. Durable formats and compatibility rules remain in the
+[integration contracts](../integrations/).
 
-External wire-contract details belong in the
+## Prompt Maintenance And Backend Caching
-[OpenAI-compatible integration doc](../integrations/openai-compatible.md).
 
-## Retries And Timeouts
+Prompt message order and shared asset bytes are runtime behavior. Backend cache
+reuse depends on identical preceding roles, rendered bytes, and cache-control
+metadata—not merely equivalent meaning. Keep reusable shared assets
+byte-identical and preserve each prompt’s declared ordering and cache controls
+when editing it.
 
-The adapter retries:
+For sibling prompts that can reuse the same source material, order universal
+shared context first, request source material next, and module-specific
+suffixes last. Put a cache boundary at a reusable prefix that is useful to the
+backend. Redundant intermediate cache boundaries do not extend that reusable
+prefix and add no value.
 
-- provider request failures;
+Prompt-family owners may choose a different sequence when their inputs and
-- response read failures;
+reuse pattern differ. The D&D family’s extraction, scene-chunking, and NPC
-- HTTP `429`;
+normalization policies are maintained in [D&D Module Internals](dnd.md#prompt-construction).
-- HTTP `5xx`;
+Do not add tests that enforce prompt prose; prompt tests should verify the
-- malformed provider envelopes;
+meaningful input placement and cache controls of the prompt being changed.
-- malformed assistant JSON;
-- structured-output decode failures.
 
-Non-retryable `4xx` responses are returned without retry. Request timeout comes
+## Validation, Repair, And Retries
-from the effective LLM profile. Context cancellation is respected.
 
-## Scheduler
+PromptKit performs prompt rendering, provider execution, and the prompt’s
+structured-output validation. The adapter reports an empty result, validation
+failure, empty structured body, or decode failure as
+`ErrInvalidStructuredOutput`, while retaining the returned raw bytes and debug
+material when they exist. Provider failures remain operational errors rather
+than output-validation failures.
 
-`Scheduler` bounds concurrent provider calls. It tracks in-flight calls and a
+When PromptKit rejects backend admission before generation, the adapter maps
-FIFO queue of waiters. Cancellation removes queued waiters or releases granted
+`promptkit.ErrCapacityExceeded` to
-permits.
+`contracts.ErrLLMCapacityExceeded`, retaining prompt context and a redacted
+upstream diagnostic without exposing the PromptKit sentinel or capacity-error
+type as a framework contract. When supplied, the normalized selected backend
+ID appears only in that safe application-owned diagnostic context. A canceled
+caller context takes precedence. The adapter does not retry capacity failures;
+the pipeline's existing binding attempt policy sees the operational error and
+decides whether to rerun the complete operation.
 
-`NewScheduledClient` wraps any structured LLM client and runs each completion
+Prompt-declared repair is executed within PromptKit’s structured-output flow.
-inside the scheduler.
+The current production D&D prompt manifests set repair attempts to zero. That
+setting does not replace pipeline retry behavior: a binding’s configured retry
+count reruns its stage attempt after an error or rejection, and an exhausted
+rejection is a recorded output rather than a provider error. The pipeline owns
+attempt lifecycle, validation chains, and retry diagnostics; see
+[Pipeline Internals](pipeline.md#validation-retries-and-output) and the
+[binding reference](../config.md#module-bindings-and-validators).
 
-Effective concurrency is:
+## Timeout Ownership
 
-1. `llm_profiles.<id>.max_concurrency`, when greater than zero;
+The caller context remains the outer cancellation authority. PromptKit applies
-2. `concurrency.total_llm`, when greater than zero;
+a positive effective generation timeout as an inner request deadline; an
-3. `1`.
+explicit zero disables only that generation deadline. The HTTP client timeout
+is a separate transport-wide cap. Notarius forwards the caller context and
+does not install another timeout wrapper around PromptKit.
 
-## Schema Registry
+The selected PromptKit profile owns generation settings. Notarius binding
+retries remain outside the adapter and repeat the complete module operation
+and validation chain. PromptKit does not add a provider retry loop.
+Operator-facing behavior is summarized in
+[Operations](../operations.md#operational-limits), and the pinned upstream
+contract is identified in
+[PromptKit Integration](../integrations/pkg-promptkit.md).
 
-The framework schema registry embeds generic test schemas. It also exposes
+## Observability And Redaction
-helpers for caller-owned schemas:
 
-- `LoadResponseSchema`
+When debug recording is enabled, the pipeline decorates the shared client. The
-- `LookupResponseSchema`
+wrapper records prepared prompt and response material, timing, selected profile
-- `MustLookupResponseSchema`
+and backend, effective model parameters, and call identifiers in the run’s
-- `ResponseSchema.DiagnosticsMap`
+debug bundle, including material available from a failed structured completion.
+Effective parameters use PromptKit's stable lower-case JSON field names and may
+include `backend_id`. For a successful completion, a debug-write failure is
+surfaced; when the completion already failed, its call error remains the
+result. Debug-bundle location, retention, and handling are operational concerns
+documented in [Operations](../operations.md#debug-bundles).
 
-`DiagnosticsMap` omits raw schema content and includes metadata such as key,
+Run manifests receive selected profile summaries, including optional effective
-ID, version, name, and SHA-256.
+backend and reasoning provenance, and component identities—not prompt, schema,
+source, reference, or response content. The published field semantics belong
+to the [JSON output contract](../integrations/json-output.md#manifestjson).
+Provider error text is wrapped with prompt context and bearer credentials are
+redacted before it crosses the runtime boundary. Known-secret redaction is
+available to other runtime collaborators; it does not make prompt or response
+contents safe for general logging.
 
-The D&D spell extractor owns and loads its own embedded response schema.
+## Failure Boundaries
 
-## Secret Redaction
+- Construction fails for missing asset registries, mutually exclusive profile
+  sources, invalid asset registration, or a non-positive scheduler limit.
+- Preparation failures, unavailable explicit profiles, provider failures, and
+  context cancellation propagate to the calling stage with context.
+- Backend admission exhaustion is a provider-neutral operational error and is
+  not classified as invalid structured output or validator rejection.
+- Malformed or schema-invalid provider output is classified separately as
+  invalid structured output so the module or pipeline can apply its own retry
+  and rejection policy.
+- Domain semantic checks, evidence decisions, and deterministic normalization
+  run outside the provider adapter.
 
-Provider errors are passed through `ErrorWithSecretsRedacted` with the API key
+## Focused Verification
-and bearer-token value. Config diagnostics use redacted effective config
-payloads.
 
-Do not add raw provider request bodies, response bodies, API keys, or prompt
+Read the LLM adapter, scheduler, asset registry, schema loader, and redaction
-payloads to diagnostics by default.
+tests when changing this boundary. Prompt changes also require the owning
+module’s asset tests, and retry or debug changes require focused pipeline or
+CLI coverage. The focused runtime and D&D checks are:
+
+~~~sh
+go test ./internal/framework/llm/... ./internal/modules/dnd/...
+~~~

docs/internal/modules.md

@@ -1,165 +1,113 @@
-# Modules
+# Module Internals
 
-Production modules live under `internal/modules`. Each module implements one
+This guide owns the mechanics for implementing and registering production
-contract from `internal/framework/contracts`, exposes a `ModuleSpec`, and
+modules. [Configuration](../config.md) owns selectable keys, binding syntax,
-registers itself with the matching pipeline registry.
+reference configuration, and default validator chains. Durable input and output
+shapes belong in [integration contracts](../integrations/).
 
-The CLI production catalog currently registers only the modules listed here.
+The D&D family has additional shared conventions and domain-specific
+exceptions. See [D&D Module Internals](dnd.md) rather than adding them here.
 
-## Contract Pattern
+## Module Boundary
 
-A production module package should provide:
+A module is a typed implementation registered for one pipeline stage. Its
+`ModuleSpec` is the public-to-the-framework declaration of its stable key,
+stage, execution class, required and provided capabilities, artifact kind, and
+accepted reference slots. The execution class states whether a module is
+`deterministic` or `llm_backed`; registries retain it for catalog inspection and
+resolved-pipeline debug data without constructing the module. The framework
+uses the declaration to resolve a configured binding before it builds the
+implementation. After selection, the resolver applies profile inheritance only
+to bindings whose declared execution class is `llm_backed` and rejects a
+binding-specific profile on a deterministic module. The user-facing precedence
+contract belongs in [Configuration](../config.md#pipelines).
 
-- a stable module key;
+Implementations that accept options must provide both an option validator and
-- a constructor such as `New`;
+a builder. The validator is used while resolving configuration; the builder
-- the relevant contract implementation;
+decodes the same options and constructs the implementation from the prepared
-- `ModuleSpec`;
+`BuildRequest`. Reject unknown options in both paths. A builder receives only
-- `Register`;
+the dependencies and materialized references that the framework prepared for
-- focused tests for registration, options, contract behavior, and errors.
+that operation, so it must not re-read configuration or files.
 
-Module specs should describe capabilities accurately. Resolution uses specs to
+Registry helpers register the typed builder for a stage-specific registry.
-reject incompatible pipelines before execution.
+They are preferable to hand-written untyped registration because they retain
+the artifact type at the framework boundary. Registrars validate the registries
+they need, register each leaf implementation, and add any family-owned assets
+or default validator chains. They return contextual errors so production
+composition fails at startup rather than at the first run.
 
-## `seriatim` Input
+An artifact family can register an optional typed evidence projector alongside
+its codec. The projector returns defensive copies of the artifact's direct
+generic source references and must use the codec's exact Go type. It does not
+interpret surrounding context or publish files; the pipeline validates the
+capability during preparation and the output boundary owns publication. See
+the [Published Evidence Context contract](../integrations/evidence-context.md)
+for the durable result.
 
-Package: `internal/modules/input/seriatim`
+## Production Composition
 
-The `seriatim` adapter parses Seriatim transcript JSON into a generic source
+Production composition is intentionally split by family:
-document. It owns transcript JSON details, source ID selection, source digest
-creation, transcript segment validation, and segment metadata mapping.
 
-Provides:
+- The generic registrar provides the unit chunker, generic JSON validators,
+  and JSON output encoder.
+- The Seriatim registrar provides the transcript input adapter. Its external
+  input behavior is defined by the [Seriatim contract](../integrations/seriatim.md).
+- The D&D registrar provides its codecs, extractors, mergers, normalizers,
+  validators, prompt assets, fallback profile asset, and default chains. Its behavioral conventions
+  are documented in [D&D Module Internals](dnd.md).
 
-- `source.transcript`
+The CLI owns the composition that invokes these registrars. A module package
-- `transcript.speaker`
+may register its own family but must not assemble the CLI or make framework
-- `transcript.timestamps`
+packages depend on production extensions.
 
-External JSON shape belongs in the Seriatim integration doc.
+## Adding Or Changing A Module
 
-## `generic` Chunker
+1. Choose the pipeline stage and the typed artifact boundary. Put external
+   input or durable artifact formats in the relevant integration contract,
+   not in this guide or in a private LLM response type.
+2. Define a stable `ModuleSpec` with an explicit execution class, the exact
+   capabilities, and reference slots needed for the operation. Model a
+   producer/consumer handoff as an artifact-compatible slot; configuration
+   then chooses an external file or a generated binding.
+3. Implement strict option decoding, construction, and the typed stage
+   interface. Preserve caller ownership: do not retain mutable request data
+   and return defensive copies where an implementation exposes stored data.
+4. Register the module through its typed registry helper and add it to the
+   owning family registrar. Add a default validator chain only when that
+   family owns the behavior; otherwise require an explicit compatible chain.
+5. Update the selectable-key and chain reference in
+   [Configuration](../config.md#production-module-keys), the applicable
+   integration contract, and focused tests. Keep the configuration document
+   as the sole list of production keys and validator order.
 
-Package: `internal/modules/chunk/generic`
+## Validation And References
 
-The `generic` chunker splits source units into ordered chunks. It validates the
+Validators operate on the value produced at their configured stage. A default
-source document, clones source units, assigns chunk IDs such as `chunk-000001`,
+chain is ordered behavior, not a set: JSON parsing, structural checks,
-and records chunk metadata for start unit, end unit, and unit count.
+domain-specific checks, durable-schema checks, and advisory checks may have
+different responsibilities and failure handling. The active default chains and
+override rules are maintained in
+[Configuration](../config.md#production-validator-keys-and-default-chains).
 
-Options:
+Reference slots are part of the module specification. They describe the
+accepted artifact kind, media type, size, and whether a binding is required;
+the framework validates those constraints before construction. An external
+reference is materialized during preparation. A generated reference is a
+compatible normalized artifact handed from an earlier pipeline step at
+operation time. The configuration reference rules, including precedence and
+ordered-handoff requirements, are maintained in
+[Configuration](../config.md#references-and-ordered-handoffs).
 
-- `max_units`: positive integer, default `50`;
+## Focused Verification
-- `overlap_units`: non-negative integer, default `0`, and less than
-  `max_units`.
 
-Provides:
+Exercise the leaf implementation and its registration path when changing a
+module. Registry and registrar tests cover duplicate keys, required registries,
+and typed construction; pipeline resolution tests cover capabilities, options,
+and reference compatibility. Domain packages should additionally test their
+codecs, validators, normalizers, and any integration handoffs they own.
 
-- `chunks`
+Run the affected package tests while iterating. The complete module suite is:
 
-## `dnd/spells` Extractor
+~~~sh
-
+go test ./internal/modules/...
-Package: `internal/modules/extract/dnd/spells`
+~~~
-
-The `dnd/spells` extractor owns D&D spell-cast artifact semantics. It renders
-embedded prompts, loads the embedded structured response schema, calls the
-structured LLM client, converts spell-cast responses into artifact candidates,
-and supplies deterministic validators.
-
-Requires:
-
-- `chunks`
-- `source.transcript`
-
-Provides:
-
-- `dnd.spell_casts`
-
-Artifact type and schema version:
-
-- artifact type: `dnd.spell_cast`
-- schema version: `v1`
-
-The extractor adds prompt and response-schema provenance to lane manifest
-metadata. Durable artifact payload details belong in the
-[D&D spell artifact contract](../integrations/dnd-spell-artifacts.md).
-
-## D&D Spell Validators
-
-The spell extractor returns two built-in validators:
-
-- `dnd/spells/shape`: rejects malformed payloads and missing required fields.
-- `dnd/spells/source_refs`: rejects candidates without valid source references.
-
-Reason codes include:
-
-- `invalid_payload`
-- `missing_required_field`
-- `missing_source_ref`
-- `invalid_source_ref`
-
-These validators are supplied by the extractor when no validators are configured
-for the lane.
-
-## `appendorder` Merger
-
-Package: `internal/modules/merge/appendorder`
-
-The `appendorder` merger clones and appends candidates in chunk order. It does
-not deduplicate or reconcile candidates.
-
-Provides:
-
-- `merged`
-
-## `noop` Normalizer
-
-Package: `internal/modules/normalize/noop`
-
-The `noop` normalizer clones merged candidates and returns them unchanged.
-
-Requires:
-
-- `merged`
-
-Provides:
-
-- `normalized`
-
-## `json` Output
-
-Package: `internal/modules/output/json`
-
-The `json` output encoder converts approved artifacts, rejected artifacts,
-warnings, and the run manifest into logical JSON output files. It groups
-approved artifacts by artifact type and sanitizes artifact-type file names.
-
-Requires:
-
-- `normalized`
-
-Provides:
-
-- `encoded`
-
-Durable output file shapes belong in the
-[JSON output contract](../integrations/json-output.md). Operator behavior
-belongs in [Operations](../operations.md).
-
-## Production Registration
-
-Production registration is centralized in `internal/cli/catalog.go`.
-
-Do not make framework code import production modules. The CLI wires production
-modules at the application boundary; tests may provide fake registries or fake
-catalogs directly.
-
-## Adding A Module
-
-When adding a module, keep source-format and extraction-domain boundaries clear:
-
-- input modules may know external source formats;
-- extract modules may know artifact semantics and prompt/schema assets;
-- merge and normalize modules own candidate combination and reconciliation;
-- output modules own serialization, not diagnostics or CLI reporting.
-
-Update [Development](../policy/development.md), [Configuration](../config.md),
-internal docs, integration docs, and examples when the new module becomes
-implemented production behavior.

docs/internal/overview.md

@@ -1,86 +1,58 @@
 # Internal Overview
 
-This directory documents implemented Notarius internals for developers and LLM
+This document is the implemented component map for Notarius. Normative
-coding agents. It complements [Architecture](../policy/architecture.md), which
+boundaries and dependency direction belong in
-is the durable policy for boundaries and invariants.
+[Architecture](../policy/architecture.md). User and operator contracts belong
+in the [CLI](../cli.md), [Configuration](../config.md),
+[Operations](../operations.md), and [integration contracts](../integrations/).
 
-## Executable And CLI
+## Execution Path
 
-`cmd/notarius` calls the CLI package. `internal/cli` owns:
+~~~
+cmd/notarius -> internal/cli -> configuration and production composition
+             -> internal/framework/pipeline -> logical output files
+             -> internal/cli -> durable output and optional state/debug data
+~~~
 
-- command parsing and usage;
+The CLI is the application boundary: it discovers configuration, composes
-- config discovery and loading;
+production registries and runtime collaborators, invokes the framework, and
-- production module catalog and registry wiring;
+places returned files. The framework resolves and prepares a fixed extraction
-- production LLM client construction;
+pipeline, then returns logical results without owning process behavior or
-- run directory creation;
+physical state roots.
-- durable output writes;
-- user-facing stdout, stderr, and exit codes.
 
-The CLI should stay thin around framework contracts. Domain extraction behavior
+## Components
-belongs in modules, not in command handlers.
 
-## Core Packages
+| Area | Implemented owners | Responsibility |
+| --- | --- | --- |
+| Executable and command boundary | **cmd/notarius**, **internal/cli** | Process entry, command dispatch, configuration discovery, production composition, runtime collaborator setup, durable file placement, and user-facing reporting. |
+| Configuration | **internal/core/config** | Defaults, strict YAML parsing, environment overrides, structural validation, effective resolution, redaction, and resolved-composition summaries. |
+| Generic models | **internal/core/source**, **internal/core/artifacts**, **internal/framework/contracts** | Source documents and chunks, manifests and provenance, plus typed artifact, reference, validation, output, and structured-completion contracts. |
+| Pipeline framework | **internal/framework/pipeline** | Registries, profile and reference resolution, typed preparation, validation, retry coordination, ordered execution, handoff, and result assembly. |
+| LLM and prompt runtime | **internal/framework/llm**, **internal/framework/promptfs** | Provider-neutral structured completions, scheduling, profile recording, prompt assets, schema registration, and credential-shaped-value redaction. |
+| Runtime state | **internal/core/fileio**, **internal/core/debugbundle**, **internal/framework/checkpoint**, **internal/framework/chunkplan**, **internal/framework/chunkmap**, **internal/framework/debug** | Confined atomic files, debug bundles, checkpoint and chunk-plan state, accepted chunk maps, and pipeline-facing debug recording. |
+| Production extensions | **internal/modules/generic**, **internal/modules/seriatim**, **internal/modules/dnd** | Domain-neutral extensions, Seriatim input support, and D&D extraction families registered into the production catalog. |
 
-- `internal/core/artifacts`: artifact candidates, approved artifacts, rejected
+Generic core and framework packages do not depend on production extensions.
-  artifacts, validation decisions, and run manifests.
+Concrete extensions depend inward on their contracts and are registered only at
-- `internal/core/config`: defaults, YAML config parsing, environment overrides,
+the CLI composition boundary.
-  validation, redaction, and resolved pipeline config.
-- `internal/core/diagnostics`: per-run diagnostics directory creation,
-  diagnostics artifact writers, atomic writes, and retention decisions.
-- `internal/core/source`: source documents, source units, source references, and
-  validation.
 
-Core packages should remain deterministic and concrete. They should not import
+## Focused Documentation
-production modules.
 
-## Framework Packages
+- [Configuration Internals](configuration.md): loading, validation, effective
+  resolution, redaction, and resolved-composition identity.
+- [CLI Internals](cli.md): command dispatch, production composition, run
+  orchestration, and terminal reporting.
+- [Pipeline Internals](pipeline.md): resolution, preparation, execution,
+  validation, typed handoff, and framework state hooks.
+- [Run State Internals](state.md): output, cache, debug collaborator
+  composition, and path safety.
+- [LLM Runtime](llm.md): structured completion, scheduling, prompt assets,
+  profiles, and secret handling.
+- [Module Internals](modules.md): generic extension registration, module
+  construction, validation, and reference mechanics.
+- [D&D Module Internals](dnd.md): shared D&D extractor conventions, generated
+  reference projections, and lane-specific exceptions. Durable D&D and
+  Seriatim data shapes remain in the [integration contracts](../integrations/).
 
-- `internal/framework/contracts`: interfaces and request/result structs for
+Use this map to find an owner, then read the focused document and its tests
-  input adapters, chunkers, extractors, mergers, normalizers, validators, output
+before changing behavior.
-  encoders, and structured LLM clients.
-- `internal/framework/pipeline`: module registries, module specs, profile
-  resolution, capability checks, run orchestration, warnings, validation, and
-  manifest population.
-- `internal/framework/llm`: OpenAI-compatible structured-output client,
-  scheduler, schema registry, retries, and secret redaction.
-- `internal/framework/prompt`: embedded prompt registry and template rendering.
-- `internal/framework/validate`: validator decision helpers and cardinality
-  enforcement.
-
-Framework code should stay source-agnostic and domain-agnostic.
-
-## Module Packages
-
-Production module packages live under `internal/modules`:
-
-- `input/seriatim`
-- `chunk/generic`
-- `extract/dnd/spells`
-- `merge/appendorder`
-- `normalize/noop`
-- `output/json`
-
-Each module package owns its contract implementation, module spec,
-registration, options, focused tests, and module-specific errors.
-
-## Fixtures And Tests
-
-The repository uses focused package tests plus a fixture-driven CLI workflow.
-
-- CLI acceptance tests cover maintained examples under `examples/`.
-- Pipeline tests cover registry composition and end-to-end framework behavior
-  with fakes.
-- Module tests cover implemented module contracts without requiring real
-  provider calls.
-- LLM tests use local test servers and fakes.
-
-Do not use real external services in tests. Use fakes, fixtures, or local test
-servers.
-
-## Boundary Reminders
-
-- Source-format details stay in input modules and integration docs.
-- Extraction-domain details stay in extract modules and artifact docs.
-- Provider wire details stay in the LLM runtime and provider integration docs.
-- Durable output contracts belong in integration docs.
-- Operator procedures belong in `docs/operations.md`, not internal docs.

docs/internal/pipeline.md

@@ -1,127 +1,185 @@
 # Pipeline Internals
 
-The implemented pipeline runner lives in `internal/framework/pipeline`. It
+This document describes the framework-owned pipeline mechanics in
-executes the fixed workflow defined by the architecture policy:
+**internal/framework/pipeline**. [Configuration](../config.md) owns selectable
+profiles, bindings, and retry settings; [Operations](../operations.md) owns
+state lifecycle and recovery; and the [integration contracts](../integrations/)
+own durable output shapes. Concrete production extensions are covered by
+[Module Internals](modules.md).
 
-```text
+## Boundary
+
+The pipeline framework accepts a resolved composition, registries, shared
+dependencies, input bytes, and state/debug collaborators. It returns logical
+output files, normalized artifacts, recorded rejections and warnings, manifest
+provenance, and checkpoint decisions. The CLI owns process arguments,
+configuration discovery, physical roots, and placement of returned output
+files.
+
+The framework has one fixed shape:
+
+~~~
 input -> chunk -> extract -> merge -> normalize -> output
-```
+~~~
 
-Pipeline execution is serial. The runner executes the resolved lanes one after
+Input and chunking are pipeline-wide. A selected artifact lane owns extract,
-another in the fixed workflow order.
+merge, and normalize; output aggregates the terminal lane outcomes. A pipeline
+is an ordered list of steps, not an arbitrary workflow graph.
 
-## Profile Resolution
+## Resolve, Materialize, Prepare
 
-Config loading produces `pipeline.PipelineProfile` values. Resolution happens
+Resolution turns a configured pipeline profile into a **ResolvedPipeline**.
-before execution:
+It normalizes the pipeline and lane identities, applies stage defaults, selects
+requested lanes where that is supported, resolves validator chains, checks
+module capabilities and typed artifact compatibility, validates options, and
+assigns a deterministic resolved-composition digest. The resolved pipeline
+contains bindings and declared reference targets, not external reference bytes.
+After selection, the resolver applies command, binding, and pipeline profile
+precedence to LLM-backed bindings and validators only; prompt defaults remain
+an empty resolved binding profile. Deterministic bindings remain profile-free.
+These effective values are part of the digest, so execution and checkpoint
+consumers do not repeat profile inheritance.
+Configuration resolution supplies the selected profile and catalog; see
+[Configuration Internals](configuration.md).
 
-1. `internal/core/config.Config.Resolve` validates config and finds the named
+External reference materialization happens before preparation. The materializer
-   pipeline.
+checks that each slot is declared by the selected module, resolves a file path
-2. The optional lane selection is passed to `pipeline.ResolvePipeline`.
+relative to the correct configuration or working-directory origin, reads
-3. Module bindings are defaulted:
+UTF-8 text, verifies media type and size limits, and retains bounded
-   - chunk: `generic`
+provenance. A generated-artifact selector remains declared but has no bytes
-   - merge: `appendorder`
+until its producing step completes.
-   - normalize: `noop`
-   - output: `json`
-   - LLM profile: `default`
-4. The module catalog is checked for each bound module key.
-5. Module capabilities are checked in workflow order.
-6. A digest is calculated from the resolved pipeline without the digest field.
 
-The CLI writes the resolved pipeline and digest to diagnostics.
+Preparation is the construction boundary. It validates the resolved shape and
+registry set, clones the resolved data, then constructs the input adapter,
+chunker, stage-local validators, every typed lane, and output encoder with
+cloned options, references, and shared dependencies. It also collects stable
+checkpoint fingerprints. Missing registrations, incompatible typed entries,
+nil implementations, and constructor failures are reported before source
+parsing or any stage operation begins.
 
-## Registries And Module Specs
+An output encoder can opt into source-evidence publication through its output
+policy. Preparation keeps the configured lane allowlist and active lanes
+separate, then verifies an exact typed evidence projector and registered codec
+for each active lane. The resulting private plan is immutable; lanes excluded
+by invocation filtering remain configured but do not acquire a projector for
+that run.
 
-`pipeline.Registries` holds concrete constructors for execution. A
+## Typed Lanes And References
-`pipeline.ModuleCatalog` exposes module specs for config validation and
-resolution.
 
-Every production module registers a `ModuleSpec` with:
+Each resolved lane has one artifact kind, codec, and exact Go type. The
+framework uses private type erasure only around those typed operations; every
+handoff checks exact type and codec identity and reports incompatibility as an
+error rather than panicking. Encoding through the registered codec is the
+boundary for output, checkpoints, debug records, and generated references.
 
-- `Key`: module key used in config;
+Reference targets are stage- and lane-specific. External reference bytes are
-- `Stage`: module kind such as input, chunk, extract, merge, normalize,
+cloned into the operation request. Generated references are built at the next
-  validate, or output;
+step boundary from exactly one accepted normalized producer output. The
-- `Provides`: capabilities added after that module runs;
+framework decodes and re-encodes that output with the registered producer
-- `Requires`: capabilities that must already be available.
+codec, checks its complete schema and media identity, and records a content
+digest plus bounded producer provenance. A missing, ambiguous, invalid, or
+incompatible producer prevents the consumer step from starting.
 
-Capability checks prevent incompatible pipeline composition before a run starts.
+## Execution And Ordering
 
-## Runner Input And Output
+The runner validates its input, installs no-op state collaborators when none
+were supplied, and serially performs source parsing and chunk-plan selection.
+An accepted plan is materialized into source-addressed chunks and passes the
+configured chunk validators before any lane runs. A chunk rejection is a
+recorded pipeline outcome: lanes do not start, but the output stage can encode
+the terminal result.
 
-`pipeline.RunInput` carries:
+For each ordered step, the runner first builds generated reference sets from
+the accepted normalized outputs of earlier steps. It then executes the step's
+lanes. Later steps do not begin until the current step is terminal and its
+generated handoffs have succeeded.
 
-- a `ResolvedPipeline`;
+Within a step, the lane engine dispatches extraction jobs in deterministic
-- optional source ID, input path, and raw input bytes;
+chunk-first, lane-second order to a bounded worker group. When all extraction
-- a structured LLM client;
+jobs for one lane are terminal, a bounded continuation group can run that
-- run ID, start time, LLM profile manifest metadata, and CLI metadata.
+lane's merge and normalize work while extraction for other lanes continues.
+The framework does not create an unbounded goroutine per chunk or lane.
 
-`pipeline.RunOutput` carries:
+Completion timing does not determine public results. The coordinator restores
+lane and chunk order before merging results, and selects a framework error by
+stable stage, lane, and chunk position. A validator rejection records a lane
+outcome without cancelling unrelated work. A framework error or parent
+cancellation cancels derived work, prevents queued work from starting, waits
+for started workers, and prevents output encoding.
 
-- run manifest;
+## Validation, Retries, And Output
-- approved artifacts;
-- rejected artifacts;
-- warnings;
-- logical output files returned by the output encoder.
 
-The CLI owns durable file writes and diagnostics writes after the runner returns.
+Every chunk, extract, merge, and normalize candidate passes its resolved
+validator chain. Validators receive immutable canonical input appropriate to
+their target: chunks, typed values, or serialized codec bytes. They may
+approve, approve with warnings, reject, or fail. A rejection is an ordinary
+pipeline result; a validator error is a framework error.
 
-## Execution
+The runner applies the binding's retry policy around a stage operation and its
+complete validation chain. It preserves warnings only from the final accepted
+or rejected attempt. Cancellation stops retries. Normalizer-specific retry
+directives consume this same budget and validate any final safe fallback through
+the normalizer chain.
 
-The runner:
+After terminal lane work, the runner assembles manifest provenance, normalized
+artifacts, rejections, warnings, and an optional accepted chunk map. When an
+output policy selected evidence lanes, it decodes accepted serialized normalize
+outputs through their registered codecs and invokes the prepared typed
+projectors. Rejected or absent lanes contribute nothing. This reconstruction is
+also used after normalized-checkpoint reuse, so no second typed output channel
+is retained. The runner passes the resulting owned artifact to the output
+encoder, which returns logical files and does not choose a physical directory.
+The CLI publishes those files only after the runner returns without a framework
+error. Logical file names and schemas are defined by the [output integration
+contracts](../integrations/).
 
-1. validates run input and registries;
+## Checkpoint And Debug Hooks
-2. builds the input adapter and parses the raw input into a source document;
-3. validates the source document;
-4. builds the chunker and produces source chunks;
-5. runs each selected artifact lane in sorted resolved order;
-6. builds the output encoder and validates logical output file names.
 
-Within an artifact lane, the runner:
+The runner receives checkpoint and debug interfaces rather than roots. It
+records workflow transitions and reuse decisions through the supplied
+collaborators, and clones reusable artifacts before they re-enter normal typed
+handoff. Generated-reference dependencies participate in checkpoint decisions.
+Selective recomputation can require a canonical accepted normalized predecessor
+before a dependent lane starts.
 
-1. builds the extractor, merger, and normalizer;
+Debug recording is attempt-scoped and application-owned. A failure to persist
-2. records module manifest metadata when modules provide it;
+required debug data is a framework error. State roots, persistence, reason-code
-3. extracts candidates from each chunk;
+meanings, resume, and cleanup are intentionally owned by
-4. normalizes candidate envelope fields such as index, extractor key, artifact
+[Run State Internals](state.md) and [Operations](../operations.md).
-   type, and schema version;
-5. merges candidates;
-6. normalizes merged candidates;
-7. validates candidate envelope consistency;
-8. runs validators;
-9. converts approved candidates to artifacts.
 
-## Validators
+## Invariants To Preserve
 
-If a lane declares validators in config, the runner builds those validators from
+- The six fixed stages remain explicit; a pipeline is not a general DAG.
-the validator registry. Otherwise it uses validators returned by the extractor.
+- Resolution and preparation reject statically discoverable incompatibility
+  before parsing or execution.
+- Every typed lane uses one compatible artifact kind, codec, and exact Go type.
+- Generated references come only from one earlier accepted normalized producer
+  and carry canonical identity rather than an unverified value.
+- Rejections are recorded outcomes; framework errors cancel derived work and
+  prevent output encoding.
+- Public ordering and selected errors are independent of goroutine completion
+  order.
+- Pipeline modules receive collaborators and data, never CLI streams or
+  physical output, cache, or debug roots.
 
-Each validator must return exactly one decision for each eligible candidate. The
+## Focused Tests
-runner enforces decision cardinality with `internal/framework/validate`.
-Rejected candidates are removed before the next validator runs. Approved
-candidates continue through the chain.
 
-The production CLI currently registers no standalone validator modules. The
+- **internal/framework/pipeline/profile_test.go** and
-current D&D spell extractor supplies deterministic shape and source-reference
+  **typed_resolution_test.go** cover resolution, defaults, ordered steps,
-validators.
+  compatibility, validators, references, and resolved identity.
+- **internal/framework/pipeline/preparation_test.go** covers complete
+  construction before execution and contextual construction failures.
+- **internal/framework/pipeline/references_test.go** and **handoff_test.go**
+  cover external materialization, generated references, provenance, and typed
+  producer checks.
+- **internal/framework/pipeline/runner_concurrency_test.go** covers bounded
+  execution, ordered steps, stable error selection, rejections, and
+  cancellation.
+- **internal/framework/pipeline/runner_chunk_plan_test.go**,
+  **runner_typed_checkpoint_test.go**, and
+  **runner_accepted_checkpoint_test.go** cover state hooks and reuse behavior.
+- **internal/framework/pipeline/runner_attempt_debug_test.go** and
+  **runner_terminal_debug_test.go** cover attempt and terminal debug behavior.
 
-## Warnings And Failures
+Run **go test ./internal/framework/pipeline ./internal/cli** after changing a
-
+pipeline boundary. Use the more focused tests above while iterating.
-Warnings from chunking, extraction, merging, normalization, validation, and
-output encoding are accumulated in `RunOutput.Warnings`.
-
-Errors wrap the operation and module key or lane context. If execution fails
-after a manifest exists, the returned manifest is marked `failed` and receives a
-completion timestamp.
-
-On successful execution, the manifest validation status is:
-
-- `approved` when no candidates were rejected;
-- `rejected` when at least one candidate was rejected.
-
-## Manifest Population
-
-The manifest records run ID, pipeline ID, pipeline digest, module keys, artifact
-lanes, LLM profile metadata, source digest, validation status, and timing.
-
-Modules can add non-secret manifest metadata by implementing
-`contracts.ManifestMetadataProvider`. The D&D spell extractor uses this for
-prompt and response-schema provenance.

docs/internal/state.md

@@ -0,0 +1,147 @@
+# Run State Internals
+
+This document describes the implementation collaborators behind output, cache,
+and debug state. User-visible fields belong in [Configuration](../config.md),
+and physical layout, retention, recovery, reason codes, and cleanup belong in
+[Operations](../operations.md).
+
+## Composition
+
+`internal/cli` is the only physical-path composition root. It resolves the
+effective configuration, selects exact roots, allocates requested debug bundles,
+constructs cache collaborators, writes logical output files, and reports paths.
+Pipeline modules receive interfaces and request data, never output, cache, or
+debug roots.
+
+The CLI creates no chunk-plan store in bypass mode. It creates a checkpoint
+recorder only when recording is enabled and a checkpoint loader only for a
+resume invocation. It allocates debug state only after a safe run identity has
+been generated and only when debug capture was requested. These choices keep
+the three state families independently composable.
+
+## Output And Cache
+
+The pipeline runner returns logical output files. After validating every
+logical name, the CLI exclusively creates the run directory beneath the
+selected output root and performs confined, atomic file writes within it.
+The runner supplies an accepted chunk map as an optional, defensively owned
+output-request artifact. The JSON encoder alone decides whether its explicit
+option writes the map and optional index descriptor; neither the map payload
+nor its annotations are copied into the run manifest. The durable fields are
+owned by the [Accepted Chunk Map contract](../integrations/chunk-map.md).
+
+`internal/framework/chunkplan` owns source-addressed plan storage, validation,
+and atomic publication. Its store is constructed only when the selected mode is
+not `bypass`.
+
+`internal/framework/checkpoint` owns checkpoint identity, manifests, payload
+codecs, loader, and recorder. The CLI constructs a recorder whenever checkpoint
+recording is enabled and constructs a loader only for a `--resume` invocation.
+Identity incorporates explicit stable semantic fingerprints collected from
+prepared modules and validators in addition to configuration, input,
+references, runtime overrides, observed LLM profiles, and the LLM runtime's
+non-secret effective profile-source identity. A profile source change therefore
+causes a cold miss even when the configured profile ID remains unchanged.
+The serialized
+`workspace_schema_version` identifiers are frozen wire-compatibility fields;
+they do not describe a current public state surface.
+
+Ordered-step lane checkpoints include the step identity in their storage scope.
+When a later lane consumes a generated artifact, its dependency fingerprints
+include the producer's artifact kind, complete schema identity, media type,
+canonical content digest, and size. Ordinary resume compares those fingerprints
+when progressively loading consumer stage checkpoints, so changed producer
+content produces `dependency_invalidated` rather than stale downstream reuse.
+Selective recomputation instead requires each unselected producer's accepted
+normalized artifact; invalid accepted state records its specific bounded reason
+and stops before the dependent. The selected step and its transitive dependents
+record `forced_recompute`.
+
+Ordinary resume loads extract, merge, and normalize checkpoints progressively
+and may execute later lane stages after an earlier cache miss. Selective
+recomputation instead asks the loader for the required producer's accepted
+normalize artifact. That lookup reuses the existing normalize files, requires
+workspace schema v3 plus an exact non-empty invocation identity, and deliberately
+does not require extract or merge checkpoint files or dependency fingerprints.
+The runner performs canonical codec and producer-provenance validation before
+cloning the artifact into normal step output. Success restores only stored
+normalize warnings and emits one normalize decision; failure retains the files,
+records the decision, and stops without executing the producer or consumer.
+
+The loader assigns a typed category and reason code at each validation site;
+diagnostic prose is not classified after the fact. The runner then applies
+forced-execution policy, validates reusable artifact bytes through the prepared
+codec once, returns the canonical hydrated value to the stage, and records the
+final decision before enforcing a required-predecessor failure. That failure
+names only the step, lane, and stable reason code. Decision detail is selected
+from code-owned descriptions by reason code and then UTF-8 normalized and
+bounded; callers cannot supply arbitrary diagnostic prose. Typed categories and
+codes remain intact through pipeline events and become strings only in manifest
+and debug-summary JSON.
+[Operations](../operations.md#checkpoint-recording-resume-and-recompute) is the
+canonical operator-facing reason-code reference.
+
+`internal/core/fileio` provides confined atomic file writes used by state
+collaborators. The chunk-plan store retains its stronger entry validation.
+
+The CLI constructs selective-recomputation policy from resolved generated
+artifact dependencies. It forces the selected step and transitive consumers,
+while marking unforced producers as required reusable inputs. The runner owns
+the actual hydration and rejection decisions; the [Operations guide](../operations.md#checkpoint-recording-resume-and-recompute)
+owns the operator workflow and stable reason-code meanings.
+
+## Debug Bundles
+
+`internal/core/debugbundle` allocates an explicitly requested per-run bundle
+with `summary/` and `trace/` roots. `SummaryWriter` persists redacted command,
+resolution, run, warning, and failure artifacts. `internal/framework/debug`
+implements the pipeline-facing trace recorder under the trace root.
+
+The CLI allocates a bundle before pipeline resolution and treats requested
+summary or trace persistence failures as command failures. The pipeline's debug
+boundaries redact sensitive metadata and credential-shaped bytes while allowing
+application-owned trace material. Debug data is never a checkpoint source or
+cache input.
+
+Generated reference bytes exist only in cloned operation requests and are not
+written as paths into checkpoints, manifests, or debug summaries. Those state
+surfaces retain canonical identities and bounded producer provenance so that a
+resume decision can be explained without copying generated campaign content.
+
+After allocation, one CLI-owned state value accumulates the known report paths,
+pipeline outcome counts, and validation status. A single guarded terminalization
+operation writes the success report, or makes one attempt each to write the
+failure report and error log. Terminal persistence failures are reported
+separately and never replace the command's primary error.
+
+## Invariants To Preserve
+
+- Modules receive state collaborators and request data, never physical roots.
+- Output logical paths are validated before a run directory is allocated, and
+  files are atomically written within that directory.
+- Chunk-plan publication occurs only for accepted plans; bypass does not
+  construct or touch a plan store.
+- Checkpoint recording and checkpoint loading remain separate collaborators.
+- Debug state is opt-in, is not cache input, and terminal reporting does not
+  obscure the command's primary failure.
+
+## Tests To Inspect
+
+- `internal/cli/run_contract_test.go`: command-owned state allocation,
+  terminalization, and output/report boundaries.
+- `internal/cli/cache_contract_test.go`: cache-mode precedence, root selection,
+  and resume collaborator construction.
+- `internal/cli/state_hardening_test.go`: independent roots, reuse, failures,
+  permissions, cleanup, and redaction.
+- `internal/cli/recompute_policy_test.go`: forced dependents and required
+  reusable predecessors for selective recomputation.
+- `internal/cli/recompute_execution_contract_test.go`: selective recomputation,
+  filesystem recovery, deterministic decisions, and failed predecessor state.
+- `internal/cli/production_contract_test.go`: production composition and
+  configuration validation at the CLI boundary.
+- `internal/cli/example_contract_test.go`: maintained example ownership.
+- `internal/core/debugbundle/*_test.go`: bundle allocation and summary writes.
+- `internal/framework/checkpoint/*_test.go`: checkpoint serialization and
+  reuse.
+- `internal/framework/chunkplan/store_test.go`: plan envelope, confinement,
+  publication, and permissions.

docs/operations.md

@@ -1,132 +1,291 @@
 # Operations
 
-This is the canonical reference for operating implemented Notarius runs.
+This is the canonical guide for operating Notarius runtime state. The
+[CLI reference](cli.md) owns command syntax and exit statuses, while
+[Configuration](config.md) owns fields, defaults, and precedence. Maintainers
+who need implementation mechanics should read [Run State Internals](internal/state.md).
 
-## Normal Run
+## State Surfaces
 
-A run reads one source file, resolves one configured pipeline, calls the
+Each run can use independent roots with different retention and access-control
-configured OpenAI-compatible LLM profile, writes durable JSON output, and writes
+needs.
-diagnostics for inspection.
 
-```sh
+| Surface | Purpose | Created when | Retention |
-go run ./cmd/notarius run dnd-session \
+| --- | --- | --- | --- |
-  --config examples/dnd-spells.config.yml \
+| Output | Durable user-facing result bundle | A pipeline completes and returns logical output files | Keep until consumers no longer need it. |
-  --input examples/seriatim-minimal-transcript.json \
+| Chunk-plan cache | Reconstructible source-addressed plan | The configured cache mode permits cache I/O | Keep while reuse is useful. |
-  --output-dir ./notarius-output \
+| Checkpoint cache | Reconstructible execution and recovery state | Checkpoint recording is enabled | Keep only while recovery or reuse is useful. |
-  --diagnostics-dir /tmp/notarius
+| Debug bundle | Explicit diagnostic record | A run requests debug collection | Keep only under an intentional sensitive-data retention policy. |
-```
 
-The command prints a success line with the pipeline ID, approved and rejected
+Output, cache, and debug roots are never merged or cleaned automatically. Use
-artifact counts, and the output path.
+separate locations and permissions for operators or services that must not
+share application data.
 
-## Output Directory
+## Roots And Permissions
 
-Durable output is written to:
+The configured output and debug directories are exact roots. An empty cache
+directory selects a per-user root:
 
-```text
+~~~
+<os.UserCacheDir>/notarius/chunk-plans
+<os.UserCacheDir>/notarius/checkpoints
+~~~
+
+The field definitions and configuration examples are in [Configuration](config.md).
+On supported Unix systems, output directories and files are created with
+requested modes **0755** and **0644**. Chunk-plan, checkpoint, and debug
+directories and files use **0700** and **0600**. The operating system's umask
+may impose stricter output modes. Cache and debug roots may contain sensitive
+source-derived data, so provision them for one trusted account or service. An
+output bundle can also contain source content when its JSON output enables
+evidence publication. Apply an appropriate umask and output-root access policy
+before enabling that option; the requested output modes alone may not be
+suitable for transcript-bearing bundles.
+
+## PromptKit Profile Deployment
+
+Profile deployment has four distinct layers:
+
+| Layer | Owner | Operational role |
+| --- | --- | --- |
+| Prompts and schemas | Notarius module families | Embedded request and structured-output definitions. They are not deployment profile files. |
+| Fallback profiles | Notarius module families | Embedded application defaults, including D&D's `dnd-extraction` profile. |
+| Built-in profiles | PromptKit | Upstream catalog entries available when no higher-precedence source defines an ID. |
+| Operator profiles | Deployment filesystem | Complete environment-specific definitions selected by `promptkit.profile_file` or `promptkit.profile_dir`. |
+
+The maintained D&D pipeline uses the workload ID `dnd-extraction`. The
+embedded fallback makes that ID usable without an operator file. Production,
+development, and local deployments can each install a different complete
+definition for the same ID, retaining the pipeline while choosing their own
+model, backend, timeout, or reasoning policy. An operator definition wins over
+the fallback; it is not merged with it. The configuration field and full
+precedence rules are owned by [Configuration](config.md#promptkit-profiles).
+
+Use a profile source owned by the service account, keep it readable only by
+the intended operator, and supply provider credentials through the service
+environment—not in the Notarius configuration or profile YAML. The maintained
+[operator profile](../examples/profiles/dnd-extraction.yml) is secret-free and
+can be copied as a format starting point. Validate a deployment without a
+provider call or credentials:
+
+~~~sh
+notarius config validate --config /etc/notarius/config.yml --pipeline dnd-session
+~~~
+
+Profile paths are currently resolved from the process working directory, not
+from the configuration file. The complete example's
+`./examples/profiles/dnd-extraction.yml` path is valid for a repository-root
+invocation only. Use absolute paths such as
+`/etc/notarius/profiles/dnd-extraction.yml` for services and containers.
+
+## Run Lifecycle
+
+Use the [run command](cli.md#run) to start a pipeline. A valid invocation loads
+and resolves configuration before module preparation and source parsing. It
+then performs any permitted cache lookup, executes the pipeline, and publishes
+logical output files only after a successful runner result.
+
+On success, the command reports the output bundle path. A warning-bearing run
+still succeeds and reports its warning count on standard error. Errors and
+their exit classes are defined in the [CLI reference](cli.md#output-streams-and-exit-statuses).
+
+## Output Bundles
+
+Each successful run receives a generated safe run identifier and writes beneath:
+
+~~~
 <output-root>/<run-id>/
-```
+~~~
 
-The default output root is `./notarius-output`. Use `--output-dir` to choose a
+The [JSON output contract](integrations/json-output.md) owns the logical files
-different root.
+and their schemas. Before creating the run directory, Notarius validates every
+logical output path. It refuses an existing run directory without changing it.
+Files are written atomically; if a later write fails, the newly created partial
+run directory remains for inspection and is never removed automatically.
 
-The `json` output module writes these files:
+Treat an output bundle as durable user data. Do not use cache-cleanup policy to
+remove it. An optional accepted chunk map is also durable output and can carry
+source- or model-derived annotations; its content and compatibility contract
+are defined in [Accepted Chunk Map](integrations/chunk-map.md). An optional
+[evidence context](integrations/evidence-context.md) contains source-unit text
+and metadata. It is not a cache or debug artifact: retain it with the output
+bundle only for as long as consumers need it, and apply source-content access
+controls to the entire bundle. Selected lanes may collectively cite most of a
+transcript, so a broad allowlist can make the evidence artifact nearly as
+sensitive and large as the source itself.
 
-- `index.json`: file index with paths to the manifest, artifact files,
+## Chunk-Plan Cache
-  rejected artifacts, and warnings.
-- `manifest.json`: run manifest with resolved pipeline provenance, module keys,
-  validation status, and timing.
-- `artifacts/<artifact-type>.json`: approved artifacts grouped by artifact
-  type. For the current D&D spell extractor, this includes
-  `artifacts/dnd.spell_cast.json` when spell-cast artifacts are approved.
-- `rejected.json`: rejected candidates and validator decisions.
-- `warnings.json`: warnings reported by pipeline modules or the output encoder.
 
-Output writes are atomic per file. Logical output file names must be clean,
+Chunk plans live beneath the selected chunk-plan root:
-relative, slash-separated paths and must not contain `..`.
 
-## Diagnostics Directory
+~~~
+<chunk-plan-root>/<source-sha256-hex>/plan.json
+~~~
 
-Diagnostics are written under:
+One validated canonical plan is active for each source digest. The plan stores
+boundaries and provenance, not a second copy of the entire source. This
+source-addressed policy is recorded in [ADR-0005](adr/0005-cache-canonical-chunk-plans-by-source.md).
 
-```text
+The configured cache mode controls one invocation:
-<diagnostics-work-dir>/<run-id>/
-```
 
-The default diagnostics work directory is `/tmp/notarius`. It can be set with
+- **auto** looks for a valid active plan. Missing or invalid state causes a new
-`diagnostics.work_dir`, `NOTARIUS_WORK_DIR`, or `--diagnostics-dir`.
+  plan to be generated; an accepted new plan is atomically published.
+- **refresh** skips lookup, generates a plan with the configured chunker, and
+  atomically replaces the active plan after it is accepted.
+- **bypass** performs no chunk-plan cache I/O. It does not resolve or create a
+  chunk-plan root.
 
-Implemented diagnostics artifacts:
+A reused plan is still materialized and validated against the current source.
+If a prior plan no longer gives acceptable results, use a refresh run rather
+than editing cache files. Deleting a plan is recoverable but can repeat costly
+chunking work.
 
-- `invocation.json`: command metadata such as operation, config path, input
+## Checkpoint Recording, Resume, And Recompute
-  path, selected lanes, run ID, and pipeline digest when available.
-- `effective-config.json`: resolved config with API keys redacted.
-- `resolved-pipeline.json`: resolved module bindings and pipeline digest.
-- `run-manifest.json`: the same run manifest written to durable output when it
-  is available.
-- `warnings.json`: warning list.
-- `run-report.json`: counts, status, output path, diagnostics path, and run ID.
-- `error.log`: failure message, written after diagnostics directory creation
-  when a run fails.
 
-`source-document.json` is supported by the diagnostics writer but is not written
+Checkpoint recording is an explicit configuration choice and is disabled by
-by the current CLI run workflow.
+default. When enabled, each run records stage transitions and the state needed
+for compatible recovery. A run records checkpoints even when it does not ask
+to reuse them. Checkpoint payloads can contain source-derived and intermediate
+application data, so treat the entire root as sensitive.
 
-## Retention
+Checkpoint loading is separate: [**--resume**](cli.md#run) asks a run to reuse
+compatible recorded work. A resume request fails when checkpoint recording is
+disabled. Without **--resume**, a recording-enabled run executes normally and
+does not load checkpoint state. Compatibility includes the resolved pipeline,
+input, selected lanes, runtime overrides, reference provenance, LLM-profile
+provenance, the effective PromptKit profile-source fingerprint, and
+prepared-component fingerprints. When a local PromptKit backend is configured,
+compatibility also includes a non-secret fingerprint of its endpoint. Changing
+profile content or the local endpoint causes a cold miss; changing only the
+local concurrency limit does not. A changed identity produces a cold miss;
+Notarius does not migrate, rewrite, or delete older checkpoint directories.
+Reasoning-effort inheritance, replacement, and explicit clearing are distinct
+runtime identities, so checkpoints created under one state are not reused by
+either of the others.
 
-Diagnostics retention is configured with `diagnostics.retention`,
+Checkpoint state is confined below an identity-specific path:
-`NOTARIUS_DIAGNOSTICS_RETENTION`, or the default `auto`.
 
-- `auto`: keep failed runs and successful runs with warnings; remove successful
+~~~
-  warning-free runs.
+<checkpoint-root>/<pipeline-id>/<input-key>-<source-or-input-digest-prefix>/<pipeline-digest-prefix>/<identity-digest-prefix>/
-- `always`: keep every diagnostics run directory.
+~~~
-- `never`: remove successful run directories; failed runs are still retained.
 
-Unknown retention values are rejected during config validation.
+### Selective Recompute
 
-## Failures
+[**--recompute-step**](cli.md#run) requires both **--resume** and enabled
+checkpoint recording. It forces the selected ordered step and every lane that
+depends on it through generated artifact references. Unrelated lanes remain
+eligible for reuse.
 
-Failures before diagnostics directory creation, such as a missing config file or
+For an earlier producer required by a forced consumer, Notarius requires a
-an unusable diagnostics work directory, are printed to stderr and may not have a
+compatible accepted normalized artifact. It validates that artifact before
-diagnostics run directory.
+hydrating it and does not silently rerun the producer. If that state is
+missing, rejected, corrupt, non-canonical, or incompatible, the run stops
+before its dependent starts. Rerun the required producer deliberately instead
+of copying or editing checkpoint files.
 
-Failures after diagnostics directory creation are printed to stderr and written
+## Checkpoint Decisions And Recovery
-to `error.log`. Depending on where the failure occurred, diagnostics may also
-include invocation metadata, redacted effective config, resolved pipeline data,
-the run manifest, warnings, and a run report.
 
-If durable output writing fails after the pipeline completes, diagnostics are
+Checkpoint events classify work as **executed**, **reused**,
-retained for inspection and may include `run-manifest.json`, `warnings.json`,
+**forced_recompute**, or **dependency_invalidated**. Their stable reason codes
-`run-report.json`, and `error.log`.
+are written to run diagnostics and provenance. Use the code, not a copied
+error message, to decide what to repair.
 
-## Warnings
+| Reason code | Recovery meaning |
+| --- | --- |
+| **loading_disabled** | This invocation did not permit checkpoint loading. |
+| **checkpoint_missing**, **checkpoint_path_invalid**, **checkpoint_read_failed**, **checkpoint_decode_failed** | The stored checkpoint could not be located or read safely; normal resume work can execute again. |
+| **workspace_schema_incompatible**, **identity_mismatch**, **stage_mismatch**, **step_mismatch**, **lane_mismatch**, **module_mismatch** | Stored state belongs to a different compatible scope or identity; allow a fresh run to create new state. |
+| **status_not_reusable** | The recorded operation did not end in reusable state. |
+| **dependency_mismatch** | A dependency changed; dependent work is invalidated rather than reused. |
+| **artifact_payload_invalid**, **artifact_digest_mismatch**, **artifact_codec_incompatible**, **artifact_not_canonical** | A stored artifact cannot safely be hydrated; rerun the producer instead of modifying the cache. |
+| **checkpoint_reused** | A normal checkpoint passed compatibility checks. |
+| **accepted_artifact_reused** | A required predecessor's accepted normalized artifact was safely hydrated. |
+| **recompute_step** | Selective recomputation deliberately forced this work. |
 
-A successful run with warnings exits with code `0`, prints a warning count to
+Reason detail is bounded code-owned text. It is diagnostic information, not a
-stderr, and writes warnings to durable output and diagnostics when retained.
+path-discovery or data-recovery mechanism, and does not contain checkpoint,
+source, reference, credential, or environment content.
 
-The run manifest `validation_status` indicates whether final artifacts were
+## Debug Bundles
-approved or rejected after validation.
+
+Only a [debug-enabled run](cli.md#run) creates a bundle:
+
+~~~
+<debug-root>/<run-id>/
+  summary/
+  trace/
+~~~
+
+The summary contains redacted invocation and resolution information plus run,
+warning, checkpoint, chunk-plan, and terminal reporting artifacts. The trace
+contains allowlisted application diagnostic records and can include source or
+derived application data. Neither surface is a cache input. Do not treat a
+debug bundle as safe to share merely because its configuration summary is
+redacted. Invocation metadata omits reasoning effort when it is inherited,
+records the replacement value when one is supplied, and records an empty value
+when inherited reasoning was explicitly cleared.
+
+Notarius never creates debug state without an explicit request and never
+automatically deletes a requested bundle. If allocation succeeds, the command
+reports its path on both success and later failure. A summary, trace, or
+terminal-report persistence failure fails the command while preserving any
+already-written diagnostic data for inspection.
 
 ## Cleanup
 
-It is safe to remove specific old run directories after their output and
+Cleanup is manual and destructive. First inspect the exact leaf directory,
-diagnostics are no longer needed:
+then remove only that leaf; do not use a glob or a parent root as the target.
 
-```sh
+~~~
-rm -rf /tmp/notarius/run-1234567890
+rm -rf -- /srv/notarius/output/run-1721300000000000000-0123456789abcdef0123456789abcdef
-rm -rf ./notarius-output/run-1234567890
+rm -rf -- /srv/notarius/chunk-plans/0123456789abcdef0123456789abcdef0123456789abcdef0123456789abcdef
-```
+rm -rf -- /srv/notarius/checkpoints/example/seriatim-0123456789abcdef/0123456789abcdef/0123456789abcdef
+rm -rf -- /srv/notarius/debug/run-1721300000000000000-0123456789abcdef0123456789abcdef
+~~~
 
-Use exact run-directory paths. Avoid broad cleanup commands against parent
+Deleting output permanently removes user data. Deleting chunk plans or
-directories unless they are part of your own operational policy.
+checkpoints is recoverable but may repeat expensive provider or pipeline work.
+Deleting a debug bundle removes troubleshooting evidence and a retained copy of
+application data. Notarius has no cache garbage collector, rollback operation,
+or automatic cleanup command.
 
 ## Operational Limits
 
-There is no command to resume a failed run. Re-run `notarius run` after fixing
+Provider execution settings and the generation timeout come from the selected
-the cause.
+PromptKit profile. The invocation-only **--reasoning-effort** and
+**--clear-reasoning-effort** controls may replace or clear that profile setting
+for all LLM-backed calls in one run without changing the profile. PromptKit
+v0.5.0 does not add a provider retry loop. Notarius binding retries rerun the
+complete module operation and validation chain as defined by
+[module bindings](config.md#module-bindings-and-validators).
 
-Provider retries are limited to the OpenAI-compatible client retry behavior
+Timeouts are layered. Caller cancellation is the outer authority. A positive
-configured by the effective LLM profile. There is no separate CLI retry command.
+effective generation timeout adds an inner request deadline, while zero
+disables only that generation deadline. The HTTP client timeout remains a
+transport-wide cap. Notarius does not add another timeout around PromptKit.
+The pinned upstream boundary and profile-format links are in
+[PromptKit Integration](integrations/pkg-promptkit.md).
 
-Notarius writes local files only. Remote storage and archive management are not
+Concurrency has two independent layers. Notarius **total_llm** is the
-part of the implemented CLI.
+application-wide provider-call limit shared by all backends, modules, retries,
+and validators. PromptKit may impose a narrower admission limit for the
+selected backend. The effective active-generation bound is the intersection of
+both limits and can therefore be lower than **total_llm**. Built-in OpenRouter
+profiles use PromptKit's upstream backend limit; endpoint-only profiles have no
+PromptKit backend limit and remain bounded by Notarius. For the configured
+local backend, a zero **concurrency_limit** leaves only the Notarius scheduler
+as a call limit. A positive value makes the effective active local-generation
+bound the smaller of **total_llm** and that local limit.
+
+For a positive local limit, PromptKit owns its default waiting capacity and
+admission behavior. When a PromptKit backend has admitted all active and queued
+work, a new call fails as capacity exhaustion before generation. The adapter
+maps that failure to Notarius's existing provider-neutral capacity error and
+does not retry it. The calling stage's configured retry policy applies
+normally, and the run fails if those attempts are exhausted. Caller
+cancellation remains authoritative. Configuration contracts are documented
+under [PromptKit profiles](config.md#promptkit-profiles) and
+[concurrency](config.md#concurrency-output-cache-and-debug). Extract-worker
+limits and actual provider-call limits are independent. Notarius writes local
+filesystem state only; remote storage, archival, and retention automation are
+outside the implemented CLI.

docs/policy/architecture.md

@@ -1,210 +1,237 @@
 # Architecture
 
-This document defines Notarius development policy. It is inward-facing:
+This document defines the intended high-level architecture of Notarius and the
-developers and LLM coding agents should use it to preserve the project's shape,
+invariants that changes must preserve. Implemented component details belong in
-boundaries, and invariants as the code evolves.
+[Internal Overview](../internal/overview.md) and its linked documents. The
+reasoning behind significant architectural choices belongs in
+[ADRs](../adr/).
 
-Keep this document concise. It should describe durable architectural rules, not
+## System Shape
-CLI syntax, configuration reference material, module catalogs, or roadmap items.
 
-## Project Shape
+Notarius is a small, dependency-light Go application for extracting structured
+artifacts from source material. It is a general extraction platform whose
+source formats, extraction domains, validation policies, LLM providers, and
+output formats are isolated behind explicit boundaries.
 
-Notarius is a small, explicit, dependency-light Go application for extracting
+The application has one fixed pipeline shape:
-structured artifacts from source material using modular pipeline stages.
-
-The application is contract-first but not abstraction-heavy. Add interfaces and
-extension points when they protect a real boundary:
-
-- external source formats;
-- pipeline stage modules;
-- validators;
-- LLM providers and runtime plumbing;
-- output schemas and embedded assets.
-
-Avoid abstractions that only anticipate hypothetical complexity. Prefer narrow
-contracts that can be exercised by tests and real modules.
-
-## Core Invariants
-
-The framework must remain source-agnostic and domain-agnostic.
-
-Source-format details belong in input modules. Transcript-specific concepts such
-as segments, speakers, timestamps, and transcript schemas must not spread into
-runner, extractor, validator, or LLM framework code.
-
-Extraction-domain details belong in domain modules. D&D-specific concepts such
-as spells, NPCs, items, combat turns, and encounters must not spread into core
-source, runner, or LLM framework packages.
-
-Extracted facts should be grounded with source references. Source references
-should point to generic source units, not transcript-only structures. Framework
-code should preserve source-reference ranges exactly and should not merge or
-rewrite overlapping ranges unless a module explicitly owns that behavior.
-
-The application workflow is fixed:
 
 ```text
 input -> chunk -> extract -> merge -> normalize -> output
 ```
 
-These stages should remain explicit in the architecture. Chunking, merging, and
+Pipelines are configured compositions of this shape. They are not arbitrary
-normalization must not be hidden inside domain extractors when they represent
+DAGs or a general workflow language. Every stage remains explicit; general
-general pipeline behavior.
+chunking, merging, or normalization behavior must not be hidden inside an
+extractor.
 
-Pipelines are fixed-shape templates for this workflow, not arbitrary DAGs or a
+Input and chunking are pipeline-wide. Each selected artifact lane owns its
-general workflow language. Module selection should be configuration- and
+extract, merge, and normalize stages, and the output stage aggregates the run's
-registry-driven, not scattered through conditionals.
+lane outcomes.
 
-## Package Boundaries
+Notarius is contract-first without being abstraction-heavy. Interfaces and
+extension points should protect demonstrated boundaries. New abstraction is not
+itself an architectural goal.
 
-Prefer fewer, larger framework packages until a boundary proves itself through
+## Layers And Dependency Direction
-import direction, ownership, test seams, or substantial file size.
 
-Core packages should contain deterministic models and policy. Framework
+The application boundary is the composition root and may depend on concrete
-packages should contain reusable orchestration and provider plumbing. Concrete
+implementations. Domain-neutral model and framework layers provide reusable
-business logic should live under stage-oriented module packages:
+policy, contracts, and orchestration. Concrete input, pipeline, output, and
+validation extensions depend inward on those generic layers.
 
-```text
+Generic layers must not depend on production extensions. Concrete extensions
-internal/modules/input/...
+must not compose the application or take ownership of process behavior. The
-internal/modules/chunk/...
+current packages implementing these layers are inventoried in
-internal/modules/extract/...
+[Internal Overview](../internal/overview.md).
-internal/modules/merge/...
-internal/modules/normalize/...
-internal/modules/output/...
-```
 
-Use short, lowercase, idiomatic Go package names. Avoid package names that repeat
+The following dependency boundaries are mandatory:
-parent-stage context.
 
-Input modules translate external source formats into the core source model.
+- extractors and validators do not depend on concrete input adapters;
-They may know about external schema details, source-specific metadata, and
+- provider-specific types do not cross the LLM runtime boundary;
-format-specific validation rules. They should not own extraction-domain
+- external dependency types do not leak across internal package boundaries
-decisions.
+  unless that dependency is the package's explicit contract.
 
-Extract modules own artifact semantics, prompt usage, structured response schema
+Shared helpers may support demonstrated common needs, but must not move
-selection, validator defaults, and domain-specific interpretation. They should
+source-format or extraction-domain knowledge into generic framework packages.
-depend on framework contracts and core source/artifact types, not concrete input
+External dependencies require a clear correctness, security, interoperability,
-module packages.
+or complexity benefit.
 
-Merge modules combine extracted candidates. Normalize modules reconcile merged
+## Source And Domain Boundaries
-candidates for semantic consistency. Generic behavior may exist for simple
-artifact types, but domain-specific behavior belongs in modules for the relevant
-stage.
 
-Output modules serialize final artifacts and may report warnings out of band.
+Input modules translate external source formats into the generic source model.
-CLI, diagnostics, and reporting layers are responsible for surfacing those
+Format-specific schemas, fields, and validation remain with the input module
-warnings.
+and its integration contract.
+
+Framework stages operate on source documents, source units, and source
+references rather than format-specific structures. A source reference identifies
+an ordered range of generic source units. Framework code preserves those ranges
+and does not merge or rewrite them unless a stage module explicitly owns that
+behavior. Every source unit carries a validated self-reference to its containing
+document and its own unit ID.
+
+Extract modules own artifact semantics, prompt use, response schemas, and
+domain interpretation. Domain-specific concepts remain in the relevant module,
+validator, shared domain helper, and artifact contract.
+
+Typed artifact registrations declare one stable artifact kind and exact Go
+type from extraction through merge, normalization, and semantic validation.
+Pipeline resolution requires a compatible codec and matching kind-specific
+variants before a typed lane can be accepted. Framework-owned erasure remains
+private and must report type incompatibility as an error rather than a panic.
+
+An artifact kind may additionally provide a typed evidence projection that
+copies its direct generic source references. Preparation proves that projection
+matches the artifact codec's exact Go type before retaining it for an output
+policy. The runner reconstructs evidence only from accepted serialized
+normalized artifacts, and the output boundary owns any resulting publication.
+Generic framework code never infers evidence by inspecting domain JSON or
+depends on domain artifact types.
+
+Auxiliary references provide context or disambiguation. They are not source
+evidence and must not be converted into source references.
+
+## Pipeline Composition And Ownership
+
+Module selection is configuration- and registry-driven. The framework resolves
+named pipeline definitions, applies explicit defaults and runtime overrides,
+and verifies module availability and capabilities before execution. Structural
+pipeline choices must not be scattered through conditionals or hidden behind
+ad hoc command flags.
+
+Resolution validates every selected module and validator option set. A separate
+preparation boundary then constructs the complete input, chunk, lane,
+validation, and output implementation set in pipeline order. The runner accepts
+only that prepared set, so construction and dependency failures occur before
+source parsing or any other module operation.
+
+Stage ownership is explicit:
+
+- input modules convert external material into the generic source model;
+- chunk modules partition source material for extraction;
+- extract modules produce domain artifacts from chunks;
+- merge modules combine accepted extraction outputs;
+- normalize modules reconcile merged output;
+- output modules encode accepted results and run outcomes into logical files.
+
+Chunk modules produce source-addressed chunk plans rather than materialized
+chunks. The framework validates and materializes those plans into the generic
+chunk representation before chunk validation and lane execution. Plan reuse is
+therefore independent of the configured pipeline, module options, references,
+lanes, validators, and LLM profile: the canonical source digest selects the
+plan, while the current run still applies its configured chunk validators to
+the materialized chunks.
+
+The framework owns orchestration and handoff provenance. Modules return logical
+results and warnings; they do not own CLI reporting, physical output, cache, or
+debug roots, durable file placement, or checkpoint and debug lifecycle.
+
+After pipeline-wide chunking, extraction uses bounded framework concurrency.
+One run-wide worker pool receives chunk-scoped lane jobs in deterministic
+chunk-first, lane-second order. A lane may begin its merge and normalize
+continuation only after all of its extract jobs are terminal; that continuation
+remains serial within the lane, while bounded continuations for different lanes
+may overlap. The framework must not create unbounded goroutines per lane or
+chunk.
+
+Completion timing does not choose public ordering or errors. The coordinator
+orders accepted artifacts, warnings, rejections, checkpoint events, and
+framework errors by stable pipeline scope. Rejections do not cancel unrelated
+work. A framework error cancels derived work, prevents undispatched work from
+starting, waits for started work, and prevents output encoding.
 
 ## Validation
 
-Validators should be independently testable and composable.
+Validation is a framework-managed boundary around outputs from chunk, extract,
+merge, and normalize stages. Validators receive immutable stage output
+and make an explicit whole-output decision: approve, approve with warnings, or
+reject.
 
-Deterministic validators should run before LLM-backed validators when both are
+Typed artifact validators receive the domain value directly. Chunk validators
-present. Validator decision semantics should be explicit: each candidate
+receive source-zone chunks, while serialized validators receive immutable
-artifact evaluated by a validator should receive exactly one decision from that
+representation bytes and declared schema metadata. A validator registered for
-validator.
+one target or artifact kind cannot satisfy an incompatible selection.
 
-LLM-backed review belongs in module-owned validator chains, not in an implicit
+Rejection is a recorded pipeline outcome, not a framework execution error.
-global review phase. Extract and normalize modules may both use deterministic
+Validator execution failures are framework errors. Rejected output does not
-and LLM-backed validators.
+advance to the next stage.
 
-Shared validator runtime mechanics belong in framework code. Concrete validator
+Default validator chains are production composition policy and are registered
-behavior belongs in module or validator implementation packages.
+centrally by stage and module. Configuration may replace a stage-local default,
+including with an explicitly empty chain. Configured validator order is
+authoritative; the framework must not silently reorder it.
 
-## LLM Runtime
+## LLM Boundary
 
-LLM provider details belong behind transport-neutral framework contracts.
+Modules and validators use transport-neutral structured completion contracts.
+Provider request and response types, authentication, transport behavior, and
+provider error adaptation remain inside the LLM runtime.
 
-Provider-specific HTTP request and response types should stay inside the LLM
+The caller of the LLM owns prompt selection, prompt inputs, response schema,
-runtime package. Prompt construction should stay in extractors, validators, or
+and interpretation of structured output. Provider adapters do not own source-
-shared prompt helpers; provider adapters should not own domain prompt logic.
+or domain-specific prompt logic.
 
-Errors, diagnostics, reports, manifests, and redacted configuration must not
+LLM calls and other external operations accept cancellation and respect
-expose secrets.
+timeouts. Concurrency control belongs in shared runtime plumbing rather than in
+individual modules.
 
-## Configuration
+The application-wide LLM scheduler bounds actual provider calls independently
+of framework worker limits. Every LLM-backed module, retry, and validator uses
+the single injected scheduled client, including work performed by overlapping
+lanes. Provider runtime adapters may enforce a narrower backend-specific limit
+beneath this mandatory application-wide scheduler.
 
-Configuration should make pipeline composition explicit and discoverable.
+## Configuration And Provenance
 
-Centralize configuration loading, precedence, defaults, and validation. Structural
+Configuration loading, precedence, defaults, environment overrides, redaction,
-pipeline choices should come from named pipeline definitions, not ad hoc command
+and validation are centralized. Named pipeline definitions make structural
-flags. Operational overrides may be handled separately when they do not obscure
+composition explicit and discoverable. Operational overrides are permitted
-the configured pipeline structure.
+when they do not obscure the configured pipeline structure.
 
-Module registries should expose module metadata and capabilities without
+Run preparation fails before stage execution when statically discoverable
-requiring module construction. Configuration validation should fail fast when a
+modules, capabilities, reference bindings, or explicitly selected profiles are
-pipeline binds incompatible or unknown modules.
+invalid or incompatible.
 
-Run manifests should record enough resolved pipeline provenance to make a run
+Run manifests record enough resolved pipeline, module, source, reference, and
-auditable after named configuration changes over time.
+LLM provenance to make a run auditable after configuration changes. Manifests
+record identities and summaries rather than secret or large payload content.
 
-## Dependencies
+## State, Output, And Safety
 
-Prefer the Go standard library where practical.
+Notarius exposes three filesystem surfaces with independent roots and
+lifecycle:
 
-Use external dependencies only when justified by correctness, security,
+- output is durable user data; output modules define logical files and the CLI
-interoperability, or substantial complexity reduction. Good reasons include
+  owns their placement;
-widely used file formats, complex validation behavior, or secure transport
+- cache is reconstructible state, with separate chunk-plan and checkpoint
-handling.
+  families; and
+- debug is explicitly requested inspection data, combining a redacted summary
+  with a detailed trace.
 
-Avoid dependencies for small conveniences. Do not let external dependency types
+Chunk plans are keyed only by canonical source digest. Configured checkpoint
-leak across internal package boundaries unless the dependency is itself the
+recording is independent of checkpoint reuse; checkpoints are loaded only for
-explicit contract of that package.
+an invocation that explicitly requests resume. Debug is never a cache input and
+is never created without an explicit request. Pipeline modules receive
+collaborator interfaces and never physical roots.
 
-## State, Files, and Safety
+Writes are atomic where practical. Paths for writes, moves, overwrites, and
+deletion must be narrow and explicit. Notarius never automatically deletes
+output or requested debug bundles; cache cleanup is explicit and recoverable.
 
-If the application writes durable state, writes should be atomic where
+Secrets must not appear in errors, logs, output, cache, debug summaries,
-practical. Multi-step workflows should preserve enough diagnostics to support
+traces, manifests, documentation, examples, or redacted configuration. Debug
-inspection after failure.
+collection is allowlisted to application-owned payloads and must not capture
+unrelated process environment values or filesystem content. Trace data may
+contain application data and therefore inherits its sensitivity; operators own
+access controls and retention. Physical layout and operation are defined in
+[Operations](../operations.md).
 
-Code that deletes, moves, or overwrites files must use narrow, explicit paths.
+## Architectural Non-Goals
-Avoid broad parent-directory operations. Cleanup that can cause data loss must
-be opt-in.
 
-## Errors and Logging
+Notarius does not aim to provide:
 
-Errors should be actionable and preserve context. Wrap errors with operation and
+- an arbitrary workflow graph or general workflow language;
-path or resource context. CLI code should convert internal errors into concise
+- source-format or extraction-domain behavior in generic framework packages;
-user-facing messages.
+- provider-specific contracts exposed to modules;
-
+- structural pipeline composition through ad hoc CLI flags;
-Errors and logs must not expose secrets. Logs should describe operations,
+- implicit cross-stage behavior that bypasses the fixed pipeline;
-external calls, retries, and failure causes, but should not include large source
+- abstractions introduced solely for hypothetical future complexity.
-or artifact payloads by default.
-
-Long-running operations should accept `context.Context`. External calls,
-subprocesses, HTTP requests, storage operations, LLM calls, and multi-stage
-workflows should respect cancellation and timeouts.
-
-## Testing
-
-Core logic should be testable without real external services. Use fakes,
-fixtures, or local test doubles for input modules, extract modules, validators,
-and LLM clients where practical.
-
-Contract-first work should include fake implementations that prove interfaces
-compose before real modules depend on them.
-
-Maintain a fixture-driven walking skeleton that exercises the full pipeline with
-fake modules and fake external clients. This protects stage composition as real
-modules evolve.
-
-Important CLI and configuration workflows should have tests. Adapter, extractor,
-validator, and stage contracts should have focused tests that do not require
-running the full application unless end-to-end coverage is intentional.
-
-## Documentation
-
-Documentation should follow the project documentation policy. Keep user docs
-focused on implemented behavior. Put future, planned, or aspirational work only
-under `docs/roadmap/`.
-
-Core documentation should use generic terms such as source document, source
-unit, source reference, input adapter, extractor, chunker, merger, normalizer,
-artifact, validator, and run manifest.
-
-Source-format details belong in input module or integration docs.
-Domain-specific extraction details belong in extract module or artifact docs.
-
-When changing architecture, config, CLI behavior, stage modules, extractor
-contracts, validator contracts, LLM runtime behavior, or artifact schemas, update
-the relevant docs and examples in the same change.

docs/policy/development.md

@@ -1,139 +0,0 @@
-# Development
-
-This document defines contributor workflow for Notarius. For architectural
-invariants and package boundaries, read [Architecture](architecture.md) first.
-
-## Required Reading
-
-Before changing the repository, review:
-
-- [Architecture](architecture.md)
-- [Documentation Policy](documentation.md)
-
-Keep current-behavior documentation limited to implemented behavior. Put planned
-or deferred behavior under `docs/roadmap/`.
-
-## Repository Layout
-
-- `cmd/notarius`: executable entry point.
-- `internal/cli`: CLI parsing, production catalog wiring, config loading, run
-  command orchestration, output writes, and user-facing errors.
-- `internal/core`: deterministic models and policy for artifacts, source
-  documents, config, and diagnostics.
-- `internal/framework`: reusable contracts, pipeline orchestration, prompt
-  helpers, validation helpers, and LLM runtime plumbing.
-- `internal/modules`: concrete input, chunk, extract, merge, normalize, and
-  output modules.
-- `docs`: policy, user/operator docs, internal docs, integration docs, and
-  roadmap files.
-- `examples`: maintained, secret-free examples covered by tests where practical.
-
-## Validation Commands
-
-Run focused tests for the area changed, then run the broader checks when the
-change affects shared contracts, CLI behavior, or documentation examples.
-
-```sh
-go test ./...
-go vet ./...
-go build ./cmd/notarius
-```
-
-Useful focused checks:
-
-```sh
-go test ./internal/cli
-go test ./internal/core/config
-go test ./internal/framework/pipeline
-go test ./internal/framework/llm
-go test ./internal/modules/input/seriatim
-go test ./internal/modules/extract/dnd/spells
-go test ./internal/modules/output/json
-```
-
-## Go Conventions
-
-- Prefer the standard library unless a dependency is justified by correctness,
-  security, interoperability, or substantial complexity reduction.
-- Keep package names short, lowercase, and idiomatic.
-- Preserve import direction: framework and core code must not depend on concrete
-  production modules.
-- Use `context.Context` for long-running operations and external calls.
-- Return contextual errors that name the operation and relevant module, path, or
-  resource.
-- Do not include secrets in errors, logs, diagnostics, manifests, or docs.
-
-## Adding Config Fields
-
-Config behavior is centralized under `internal/core/config`.
-
-When adding a file config field:
-
-1. Update file config structs and YAML parsing in `file_config.go`.
-2. Apply the field over defaults in config application code.
-3. Add validation in `validation.go` when the field has constraints.
-4. Add environment override support in `env.go` only for operational overrides.
-5. Update redaction if the field can contain secrets.
-6. Add focused config tests.
-7. Update [Configuration](../config.md) and maintained examples when behavior
-   changes.
-
-Pipeline composition should remain config-driven. Do not add command flags that
-silently replace structural pipeline definitions.
-
-## Adding CLI Flags Or Commands
-
-CLI behavior lives in `internal/cli`.
-
-When adding CLI surface:
-
-1. Keep syntax explicit and update usage text.
-2. Validate arguments before running expensive work.
-3. Convert internal errors into concise user-facing messages.
-4. Add CLI tests for success, syntax errors, and failure modes.
-5. Update [CLI Reference](../cli.md), and update
-   [Operations](../operations.md) or [Troubleshooting](../troubleshooting.md)
-   if run behavior changes.
-
-## Adding Modules Or Adapters
-
-Concrete modules live under `internal/modules/<kind>/...` and implement the
-interfaces in `internal/framework/contracts`.
-
-For a new production module:
-
-1. Implement the relevant contract.
-2. Expose a `ModuleSpec` with the correct module key, module kind, provided
-   capabilities, and required capabilities.
-3. Expose a `Register` function that registers the module with its registry.
-4. Add focused module tests for contract behavior, registration, options,
-   validation, and errors.
-5. Register the module in `internal/cli/catalog.go` only when it is production
-   ready.
-6. Update internal docs and user-facing docs only for implemented behavior.
-
-Source-format behavior belongs in input modules and integration docs.
-Extraction-domain behavior belongs in extract modules and artifact docs.
-
-## Updating Examples
-
-Examples must be valid, secret-free, and small.
-
-- Prefer environment-based secret configuration.
-- Keep `examples/dnd-spells.config.yml` loadable by CLI tests.
-- Keep `examples/seriatim-minimal-transcript.json` compatible with the Seriatim
-  adapter.
-- Do not add expected-output fixtures unless they are validated or have a clear
-  regeneration procedure.
-
-## Documentation Updates
-
-Update docs in the same change when behavior changes.
-
-- CLI syntax: `docs/cli.md`
-- Config fields and defaults: `docs/config.md`
-- Output, diagnostics, retention, or recovery: `docs/operations.md`
-- Common user-facing failures: `docs/troubleshooting.md`
-- Internal architecture and contracts: `docs/internal/`
-- External file formats and durable integration contracts: `docs/integrations/`
-- Future or planned work only: `docs/roadmap/`

docs/policy/documentation.md

@@ -1,446 +1,144 @@
-# Go Project Documentation Policy
+# Documentation Policy
 
 ## Purpose
 
-Project documentation must help five audiences:
+This policy assigns each documentation topic to one canonical owner. Its goal is
-
+to keep Notarius documentation accurate, concise, discoverable, and resistant
-1. users who need to run the application;
+to drift for users, operators, developers, integrators, and LLM coding agents.
-2. administrators/operators who need to configure and operate it;
-3. developers who need to understand and change it safely;
-4. LLM coding agents that need clear scope, boundaries, and invariants;
-5. developers and LLM coding agents integrating this project from another codebase.
-
-Docs should be accurate, concise, task-oriented, and organized by audience. Prefer links to canonical docs over repetition.
 
 ## Core Rules
 
-### 1. Keep docs concise
+### One Canonical Owner
 
-Each document should cover a defined scope and only the essentials for that scope.
+Each authoritative fact belongs in one document. A non-owning document may give
+a short, stable summary for orientation, but it must link to the canonical owner
+instead of repeating volatile details.
 
-Avoid:
+Volatile details include commands, flags, configuration fields and defaults,
-- long background explanations;
+module keys, schemas, file names, paths, status codes, retry behavior, and
-- repeated reference material;
+runtime guarantees. If readers could reasonably treat a statement as a
-- implementation detail in user-facing docs;
+contract, maintain it only in the owning document.
-- aspirational language outside roadmap docs;
-- verbose examples where one minimal example is clearer.
 
-### 2. Document only implemented behavior outside roadmap files
+### Current And Future Behavior
+
+Outside `docs/roadmap/`, documentation describes implemented behavior only.
+Partial features may be described only to their implemented boundary.
+
+ADRs are the narrow exception: an ADR may record an accepted architectural
+decision before implementation, but acceptance must not be presented as proof
+that the behavior exists. The roadmap owns implementation status and sequencing
+until the decision is implemented. Current architecture, user, operator,
+integration, and internal documentation are updated when the behavior lands.
 
-Unimplemented, planned, aspirational, experimental, or future work may be described only under:
+### Audience And Detail
+
+Write for the document's stated audience and include only the detail needed for
+its owned topic. User and operator docs should not expose implementation detail.
+Developer docs should link to user-facing and external contracts rather than
+restate them.
 
-- `docs/roadmap/`
+### Examples
 
-No other documentation file, including `README.md`, should describe code, features, modules, stages, commands, config fields, or behaviors that do not currently exist.
+Complete copyable files belong in `examples/`. Documentation may use the
-
+smallest illustrative snippet needed to explain its owned topic, but should link
-If a feature is partial, non-roadmap docs may describe only the implemented portion and its current boundary.
+to maintained examples instead of embedding a second complete copy.
-
+
-### 3. Use canonical homes
+Examples must be valid, secret-free, and tested where practical. Commands and
-
+configuration used in documentation should match the application.
-Each type of information should have one canonical location.
+
-
+### Security And Privacy
-Canonical homes:
+
-
+Documentation and examples must not contain real credentials, private keys,
-- project purpose and quickstart: `README.md`
+private environment dumps, sensitive source material, or private infrastructure
-- development principles: `docs/policy/architecture.md`
+details unless intentionally public. Document secret-handling mechanisms, not
-- public HTTP API reference: `docs/api.md`
+secret values.
-- configuration reference: `docs/config.md`
+
-- CLI reference: `docs/cli.md`
+## Canonical Ownership
-- operations and recovery: `docs/operations.md`
+
-- troubleshooting: `docs/troubleshooting.md`
+| Topic | Canonical owner | Owned content | Content owned elsewhere |
-- public API/package consumer guidance: `docs/consumers/`
+| --- | --- | --- | --- |
-- implemented internals: `docs/internal/`
+| Product orientation and minimal end-to-end quickstart | `README.md` | What Notarius is, why it is useful, one shortest successful invocation, and links onward. | Complete command reference, configuration reference, operational procedures, implementation detail. |
-- external protocol, service, and file-format contracts: `docs/integrations/`
+| Contributor entry point | `docs/development.md` | Task-oriented reading guide, minimal contributor orientation, baseline validation commands, and links to canonical docs. | Package inventory, architecture rules, subsystem behavior, detailed change recipes. |
-- future work: `docs/roadmap/`
+| Current application architecture | `docs/policy/architecture.md` | System shape, normative ownership, dependency direction, architectural boundaries, invariants, safety properties, and non-goals. | Concrete package inventory, implementation mechanics, contributor procedures, decision history, future work. |
-- contributor workflow: `docs/policy/development.md`
+| Documentation organization | `docs/policy/documentation.md` | Documentation ownership, audience boundaries, maintenance rules, and ADR/document lifecycle. | Application architecture or product behavior. |
-- copyable examples: `examples/`
+| Testing policy | `docs/policy/testing.md` | Test philosophy, risk-based sufficiency, test boundaries, doubles, coverage guidance, regression-test policy, and criteria for adding, rewriting, or deleting tests. | Subsystem behavior, application contracts, subsystem-specific test inventories, and implementation plans. |
-
+| CLI contract | `docs/cli.md` | Commands, arguments, flags, invocation semantics, and exit codes. | End-to-end operating procedures, configuration field definitions, runtime filesystem layout, module implementation details. |
-Other files should summarize briefly and link to the canonical source.
+| Configuration contract | `docs/config.md` | Discovery and precedence, file schema, fields, defaults, environment overrides, validation rules, and user-selectable module or validator keys. | Complete example files, CLI syntax, runtime state lifecycle, module implementation details. |
-
+| Operations | `docs/operations.md` | Runtime workflows, physical filesystem and state layout, output, cache, and debug handling, resume, cleanup, permissions, recovery, and operational limits. | CLI flag syntax, configuration field definitions, logical output schemas, implementation mechanics. |
-### 4. Keep examples real
+| Public HTTP contract, if introduced | `docs/api.md` | Routes, authentication, media types, request and response schemas, status codes, pagination, caching, idempotency, rate limits, and HTTP retry semantics. | Client walkthroughs, upstream or downstream integration internals, implementation detail. |
-
+| Consumer guidance, if a public package or API is introduced | `docs/consumers/` | Task-oriented use of the public interface, minimal client examples, and consumer responsibilities. | HTTP wire semantics, external protocol contracts, internal implementation detail. |
-Examples should be valid, maintained, and free of secrets.
+| External and durable integration contracts | `docs/integrations/` | External file formats and protocols, upstream and downstream contracts, logical output bundle paths and schemas, media types, and compatibility behavior. | Physical runtime placement and lifecycle, internal transformations, CLI syntax, configuration defaults. |
-
+| Implemented component inventory | `docs/internal/overview.md` | Current packages and components, their implemented responsibilities, and links to focused internal docs. | Normative architecture, contributor reading policy, external contracts. |
-Where practical:
+| Internal component behavior | Other files under `docs/internal/` | Implementation flow, internal collaborators and state transitions, package-local guarantees and failures, and relevant tests. | Global architecture invariants, configuration definitions and defaults, external schemas, operator procedures. |
-- example configs should load successfully;
+| Architectural decision history | `docs/adr/` | Significant decisions, context, alternatives, rationale, consequences, and supersession history. | Current behavior reference, implementation status, task sequencing. |
-- example commands should match real CLI syntax;
+| Future work and implementation status | `docs/roadmap/` | Proposed, accepted, deferred, or rejected work; implementation status; sequencing; and task breakdowns. | Implemented behavior reference and architectural decision rationale. |
-- important examples should be covered by tests.
+| Complete copyable artifacts | `examples/` | Maintained configuration, inputs, and other files intended to be copied or run. | Field-by-field reference, command reference, prose explanation. |
-
+
-## Documentation Profiles
+Documents that do not exist are required only when the corresponding interface
-
+or responsibility exists. Do not create placeholder API, consumer, integration,
-All projects require:
+or operations documents for behavior the application does not have.
-
+
-- `README.md`
+## Boundary Rules
-- `docs/policy/architecture.md`
+
-
+### Orientation
-Additional docs depend on the project.
+
-
+The README owns product orientation. The developer guide routes contributors.
-### Small library
+Architecture owns normative structure. Internal overview owns the current
-
+concrete component map. These documents may link to one another but should not
-Recommended:
+maintain parallel package or behavior descriptions.
-- `docs/policy/development.md`, if contributor conventions are non-obvious
+
-
+### Commands, Configuration, And Operations
-### Simple CLI
+
-
+CLI documentation answers how to invoke the application. Configuration
-Required:
+documentation answers what settings mean. Operations answers what happens to
-- `docs/cli.md`
+runtime state and how to operate or recover the application. When a workflow
-
+crosses these topics, choose the document that owns the task and link to the
-Recommended:
+other contracts.
-- `docs/policy/development.md`
+
-
+### Contracts And Implementation
-### Config-driven CLI
+
-
+Integration and API documents define externally observable shapes and
-Required:
+semantics. Internal documents explain how Notarius implements or consumes those
-- `docs/cli.md`
+contracts. Internal docs may name a field, file, or protocol to identify a
-- `docs/config.md`
+dependency, but must link to its canonical contract for the definition.
-
+
-Recommended:
+### Security Topics
-- `examples/`
+
-- `docs/policy/development.md`
+This policy owns what documentation and examples may contain. Architecture owns
-
+application security invariants. Configuration owns credential-supply
-### Stateful or operator-facing application
+mechanisms. Operations owns permissions and handling of sensitive runtime
-
+artifacts. Internal docs own implementation mechanisms only.
-Required:
+
-- `docs/cli.md`, if CLI-based
+## Architecture Decision Records
-- `docs/config.md`, if config-driven
+
-- `docs/operations.md`
+Use sequentially numbered ADR filenames such as
-
+`0001-record-architecture-decisions.md`. Follow the lightweight Nygard format:
-Recommended:
+
-- `docs/troubleshooting.md`
+1. title;
-- `examples/`
+2. status;
-- `docs/policy/development.md`
+3. date;
-
+4. context;
-### Modular, service-oriented, or orchestration application
+5. decision;
-
+6. alternatives considered;
-Required:
+7. consequences.
-- `docs/cli.md`, if CLI-based
+
-- `docs/config.md`, if config-driven
+Treat the decision content of an accepted ADR as immutable. When a decision
-- `docs/operations.md`
+changes, create a new ADR and update the earlier ADR's status to superseded.
-- `docs/internal/`
+Rejected architectural alternatives belong in the ADR; rejected product ideas
-- `docs/policy/development.md`
+belong in the roadmap.
-
+
-Recommended:
+## Maintenance
-- `docs/troubleshooting.md`
+
-- validated examples under `examples/`
+When behavior changes, update its canonical owner in the same change. If
-
+ownership moves, remove the old definition and replace it with a link where
-### Public HTTP API service
+navigation remains useful.
-
+
-Required:
+Before completing documentation work:
-- `docs/api.md`
+
-- `docs/cli.md`, if CLI-based
+- verify affected behavior and examples;
-- `docs/config.md`, if config-driven
+- check commands, flags, fields, defaults, schemas, and paths against their
-- `docs/operations.md`
+  implementation;
-- `docs/internal/`
+- keep unimplemented behavior in the roadmap, subject to the ADR exception;
-- `docs/policy/development.md`
+- remove stale references and validate links;
-
+- confirm that non-owning documents summarize and link rather than redefine;
-Recommended:
+- confirm that no secrets or sensitive private data were added.
-- `docs/troubleshooting.md`
-- `docs/consumers/`, for task-oriented client integration guides
-- `docs/integrations/`, for upstream/downstream service contracts
-- validated examples under `examples/`
-
-### Project with public packages or consumer APIs
-
-Required:
-- `docs/consumers/api.md`
-- one `docs/consumers/pkg-<name>.md` file per public package, if public packages exist
-
-Recommended:
-- copyable consumer examples under `examples/`, if practical
-
-## Required Documents
-
-### README.md
-
-**Audience:** users, administrators, operators
-
-The README is the outward-facing project orientation page.
-
-It should include, in order:
-
-1. concise description;
-2. elevator pitch;
-3. shortest useful command or usage example;
-4. links to targeted docs.
-
-The README should be short. It is not a manual.
-
-The “shortest useful command” means the simplest command that performs the project’s core use case. (It does not mean `app --help`.)
-
-### docs/policy/architecture.md
-
-**Audience:** developers, LLM coding agents
-
-`docs/policy/architecture.md` is required for every project.
-
-It is an inward-facing development policy document. It should describe how the project is intended to be built and changed.
-
-It should include:
-
-- project shape;
-- core design principles;
-- package and boundary philosophy;
-- state/persistence philosophy, if applicable;
-- external integration philosophy, if applicable;
-- error-handling and logging principles;
-- testing expectations;
-- documentation expectations;
-- architectural invariants;
-- explicit non-goals, if useful.
-
-Notably, this file should prescribe a core development *policy* that should remain unchanged as the application evolves. It is not a place for details (e.g., CLI flags) that could change over time.
-
-The contents of `architecture.md` should be trim and concise. LLMs may be directed to review it routinely via AGENTS.md, CLAUDE.md, or similar.
-
-### docs/api.md
-
-**Audience:** external HTTP API consumers, developers, LLM coding agents integrating by HTTP
-
-Required for projects whose primary public interface is HTTP.
-
-`docs/api.md` is the canonical public HTTP API contract. It should be normative for external consumers and should not be duplicated by README, operations docs, consumer guides, or integration docs.
-
-It should include:
-
-1. base URL conventions;
-2. authentication and authorization behavior, if implemented;
-3. response envelope;
-4. supported media types and content negotiation behavior;
-5. shared query parameters;
-6. endpoint reference grouped by route family;
-7. request parameters and validation rules;
-8. response fields, units, nullability, and optionality;
-9. error response shape and status codes;
-10. pagination, caching, rate-limit, idempotency, and retry behavior, if implemented;
-11. compact request and response examples.
-
-It must document only implemented endpoints and behavior. Planned endpoints, proposed fields, future filters, and experimental response shapes belong only under `docs/roadmap/`.
-
-For HTTP API projects, `docs/consumers/` may provide task-oriented client integration guides, but those guides should link to `docs/api.md` for the authoritative endpoint contract.
-
-### docs/policy/development.md
-
-**Audience:** developers, LLM coding agents
-
-Required for projects maintained by humans and LLM coding agents.
-
-It should include:
-
-- repository layout;
-- build/test commands;
-- coding conventions;
-- dependency policy;
-- how to add config fields;
-- how to add CLI flags;
-- how to add modules or adapters, if applicable;
-- how to update examples;
-- documentation update expectations.
-
-### docs/config.md
-
-**Audience:** administrators, operators, advanced users
-
-Required for applications with configuration files.
-
-It should include, in order:
-
-1. config file locations and discovery precedence;
-2. minimal working config;
-3. production-oriented config;
-4. full configuration reference;
-5. secrets handling, if applicable;
-6. links to maintained examples.
-
-The full configuration reference should be canonical.
-
-### docs/cli.md
-
-**Audience:** users, administrators, operators
-
-Required for CLI applications.
-
-It should include, in order:
-
-1. shortest useful command;
-2. command overview;
-3. complete flag reference;
-4. common workflows;
-5. diagnostic or recovery commands, if applicable.
-
-Explain when commands are useful, not just their syntax.
-
-### docs/operations.md
-
-**Audience:** administrators, operators
-
-Required for applications that maintain state, support resume behavior, run multi-step workflows, write durable artifacts, use remote storage, or require recovery procedures.
-
-It should cover:
-
-- normal workflow;
-- filesystem layout;
-- remote storage layout, if applicable;
-- logs and manifests;
-- resume/retry behavior;
-- cleanup behavior;
-- archive/backup behavior;
-- safe recovery procedures;
-- operational caveats.
-
-### docs/troubleshooting.md
-
-**Audience:** administrators, operators
-
-Recommended once recurring failure modes exist.
-
-Each entry should include:
-
-- symptom;
-- likely cause;
-- diagnostic command or inspection step;
-- safe fix;
-- relevant links.
-
-### docs/consumers/
-
-**Audience:** developers and LLM coding agents integrating this project from another codebase
-
-Required for projects with public packages, SDKs, client APIs, plugin APIs, or other application-facing integration surfaces.
-
-This directory describes how an external codebase should consume the project's public API. It should be task-oriented and copyable where useful. It is not the place for internal implementation details or operator procedures.
-
-For projects whose public API is HTTP, `docs/consumers/` is not required, and it should not duplicate the endpoint reference in `docs/api.md`. If present, it may provide practical integration workflows, client-specific examples, or migration notes that link back to `docs/api.md`.
-
-`docs/consumers/api.md` should provide the consumer-facing overview and primary implementation workflow. It should include:
-
-1. intended consumer audience and use cases;
-2. required inputs supplied by operators or deployment configuration;
-3. recommended public package or API workflow;
-4. minimal copyable example;
-5. consumer responsibilities and boundaries;
-6. retry, idempotency, or status behavior, if applicable;
-7. links to package-specific docs and canonical integration contracts.
-
-Package-specific docs should be named `pkg-<name>.md` and should include:
-
-1. import path;
-2. intended use cases;
-3. primary types and functions needed by consumers;
-4. minimal examples;
-5. validation, error, retry, and boundary behavior;
-6. links to canonical file-format or wire-protocol contracts.
-
-### docs/internal/
-
-**Audience:** developers, LLM coding agents
-
-Required for modular, service-oriented, or orchestration projects.
-
-This directory describes implemented internal components. It is not the roadmap.
-
-Use one file per major component where useful.
-
-Each component doc should include:
-
-1. purpose;
-2. inputs and outputs;
-3. boundaries;
-4. config fields used;
-5. external adapters used;
-6. state or manifest behavior, if applicable;
-7. skip/resume behavior, if applicable;
-8. failure behavior;
-9. tests to inspect before changing;
-10. architectural invariants.
-
-### docs/roadmap/
-
-**Audience:** maintainers, developers, LLM coding agents
-
-This is the only place for planned, future, aspirational, experimental, or unimplemented work.
-
-Roadmap docs should clearly distinguish:
-
-- proposed work;
-- accepted plans;
-- deferred ideas;
-- rejected ideas;
-- implementation prompts or task breakdowns, if useful.
-
-Roadmap docs should not be confused with current behavior.
-
-### docs/integrations/
-
-**Audience:** developers, LLM coding agents
-
-Required for projects that depend on external CLIs, APIs, services, protocols, or file formats where the integration contract is important to maintain.
-
-This directory contains concise, versioned reference notes for external integration contracts. It should document only the parts of the external system that this project actually uses or exposes.
-
-For public HTTP API services, `docs/integrations/` should document upstream, downstream, storage, protocol, or runtime contracts that the service depends on or bridges. It should not become a second copy of the public HTTP endpoint reference; that belongs in `docs/api.md`.
-
-Use one file per integration where useful.
-
-## Examples Directory
-
-Projects with non-trivial configuration or workflows should include `examples/`.
-
-Useful examples include:
-
-- minimal working config;
-- production-oriented config;
-- full annotated config;
-- local development config;
-- remote/object-storage config;
-- minimal session/input file.
-
-Examples should be valid, maintained, tested when practical, and linked from relevant docs.
-
-## Security and Privacy
-
-Docs and examples must not include:
-
-- real API keys;
-- tokens;
-- passwords;
-- private keys;
-- private environment dumps;
-- sensitive user data;
-- raw private transcripts;
-- private infrastructure details unless intentionally public.
-
-Document secret-handling mechanisms, not actual secret values.
-
-## Maintenance Rules
-
-When docs change, verify the affected behavior.
-
-Where practical:
-
-- load example config files in tests;
-- test CLI examples or command parser behavior;
-- validate documented flags against real flags;
-- remove stale references;
-- update links after renames;
-- keep roadmap content out of non-roadmap docs.
-
-If documentation and code disagree, fix the documentation and/or open a roadmap item; do not leave aspirational behavior in current-behavior docs.
-
-Documentation is complete only when it matches the current code.
-
-## Documentation Change Checklist
-
-Before merging documentation changes, verify:
-
-- README is concise and orientation-focused.
-- `docs/policy/architecture.md` describes development principles.
-- `docs/api.md` is the canonical HTTP contract for HTTP API services.
-- Future work appears only under `docs/roadmap/`.
-- User-facing docs avoid unnecessary internals.
-- Consumer-facing docs explain public APIs without duplicating HTTP endpoint or integration contracts.
-- Developer-facing docs preserve boundaries and invariants.
-- Config examples match the schema.
-- CLI examples match real commands and flags.
-- Defaults appear in the canonical config reference.
-- No secrets or private data are included.
-- Links are accurate.

docs/policy/testing.md

@@ -0,0 +1,296 @@
+# Testing Policy
+
+## Purpose
+
+Our tests exist to make **incorrect changes expensive and correct changes cheap**.
+
+We do not optimize for test count, line coverage, exhaustive isolation, or the fewest possible tests. We optimize for sufficient confidence in important behavior while imposing as little unnecessary friction as possible on future development.
+
+## Every test has a cost
+
+Testing is not an unqualified good. Every test imposes both an immediate cost and a continuing lifetime cost.
+
+A test must be:
+
+- written and reviewed;
+- understood by future maintainers and coding agents;
+- executed in local and CI workflows;
+- diagnosed when it fails;
+- updated when legitimate behavior changes;
+- maintained as fixtures, APIs, and dependencies evolve; and
+- removed or rewritten when it becomes redundant, brittle, misleading, or obsolete.
+
+Tests also create cognitive and architectural friction. They can constrain refactoring, duplicate policy, slow feedback loops, add noise to failures, and cause harmless implementation changes to require unrelated edits across the suite.
+
+A test is warranted only when the confidence it provides justifies these costs.
+
+Apply this cost-benefit analysis at two levels:
+
+1. **Per test:** What realistic defect does this test detect, how consequential would that defect be, and is that protection worth the test's lifetime cost?
+2. **Across the suite:** Does this collection provide materially more confidence than a smaller, simpler suite would?
+
+The preferred test suite is a **lean suite that provides sufficient confidence in the risks that matter, without redundant or low-value tests**. We seek sufficient confidence with the least unnecessary testing friction, not the fewest possible tests.
+
+Some friction is intentional. Tests should make dangerous changes—such as breaking compatibility, corrupting data, violating security boundaries, or reintroducing subtle bugs—require deliberate review. They should not make ordinary internal changes needlessly expensive.
+
+The cost of a test is not a reason to omit testing by default. Do not cite maintenance cost abstractly. When omitting a plausible test, be able to state why the protected failure is low-risk, already covered, obvious, reversible, or cheaper to detect elsewhere. For consequential, subtle, or difficult-to-observe behavior, the presumption should favor testing.
+
+## Default testing style
+
+Use a **classical/Detroit-style** approach:
+
+- Test observable behavior, resulting state, contracts, and invariants.
+- Use real internal collaborators when they are fast and deterministic.
+- Use fakes, stubs, or mocks primarily at expensive, nondeterministic, destructive, or external boundaries.
+- Prefer package-level behavioral tests over tests coupled to private helpers or internal call sequences.
+- Treat exact collaborator interactions as testable behavior only when the interaction itself is a requirement.
+
+Examples of appropriate seams include clocks, randomness, subprocesses, remote APIs, object storage, email, and paid LLM calls.
+
+## Test execution requirements
+
+Tests in the default suite must be deterministic, offline, and independent of real credentials. They must not invoke paid APIs or depend on mutable external services. Tests that require live infrastructure must be explicitly opt-in and clearly separated from the default suite.
+
+Control clocks, randomness, environment variables, and other process-global or machine-specific state when they affect behavior. Tests should be safe to run repeatedly and alongside other tests without depending on execution order or state left by an earlier test.
+
+## What deserves tests
+
+Prioritize tests for:
+
+1. Public and package-level contracts.
+2. Domain rules and important invariants.
+3. Boundary conditions and malformed input.
+4. Failure handling, cancellation, retries, recovery, and partial success.
+5. Serialization, schemas, compatibility, and round trips.
+6. Previously observed or plausible regressions.
+7. Representative integration and end-to-end workflows.
+
+A package-level contract is behavior relied upon by another package or major collaborator, not every observable detail of a package implementation.
+
+For behavior involving **data integrity, destructive operations, compatibility, security, concurrency, idempotency, or recovery**, presume that durable tests are required unless the behavior is already credibly protected at another layer.
+
+Do not add tests merely because a function, branch, or line exists. Do not add a test when the same meaningful risk is already adequately protected elsewhere.
+
+## Choose the right test boundary
+
+Test through the narrowest stable boundary that expresses the behavior clearly.
+
+This is often the package API, but it may instead be:
+
+- a smaller pure function when dense domain logic is most clearly isolated there;
+- a package-level operation when several internal collaborators jointly produce the behavior; or
+- a larger integration boundary when correctness emerges from interaction with a real dependency.
+
+Do not force all behavior through oversized end-to-end tests. Do not test every private helper merely because it exists. Choose the boundary that gives durable confidence with the least incidental coupling.
+
+## Test behavior, not implementation
+
+A test should protect a decision, contract, or invariant—not memorialize the current implementation.
+
+Before adding or retaining a test, ask:
+
+> What realistic defect would this test catch?
+
+A test is suspect when its main purpose is to detect that someone:
+
+- changed an internal constant;
+- renamed or split a private helper;
+- reordered equivalent internal operations;
+- changed incidental formatting;
+- replaced one correct algorithm with another; or
+- refactored internal object structure without changing behavior.
+
+Refactoring should normally require no test edits unless the refactored structure is itself part of the contract.
+
+A test can be factually correct and still have negative value. Accurately describing current behavior is not enough; the protected behavior must be important enough to justify the future friction.
+
+## Expected effects of different changes
+
+Use the following expectations when evaluating test failures and test maintenance:
+
+| Change | Expected effect on tests |
+|---|---|
+| Internal refactor that preserves behavior | Existing tests should normally remain unchanged and continue to pass. |
+| Change to an internal default with no contractual significance | Behavioral tests should normally remain unchanged; tests should derive expectations from configuration or relationships rather than duplicate the old value. |
+| Intentional change to public behavior, policy, schema, or compatibility guarantees | The relevant tests should be reviewed and changed deliberately. |
+| Accidental violation of a contract or invariant | Tests should fail; fix the production code rather than rewriting the tests to accept the defect. |
+
+A test failing is not the same as a test needing to be edited. Many tests may correctly fail because of one production defect. The maintenance smell is a correct internal change that requires unrelated expectation updates throughout the suite.
+
+## Separate mechanism from policy
+
+Configurable thresholds and defaults must not be duplicated throughout the test suite.
+
+For example, do not encode an internal concurrency limit indirectly:
+
+```go
+// Production policy:
+const maxConcurrency = 4
+
+// Brittle test:
+err := startProcesses(5)
+require.Error(t, err)
+```
+
+Instead, test the mechanism relationally:
+
+```go
+const limit = 2
+runner := NewRunner(limit)
+
+require.NoError(t, runner.Start(limit))
+require.ErrorIs(t, runner.Start(limit+1), ErrTooMuchConcurrency)
+```
+
+The test should prove:
+
+- the configured limit is accepted; and
+- one beyond the configured limit is rejected.
+
+The production default should be tested exactly only when its literal value is itself a public, operational, safety, protocol, or compatibility requirement.
+
+Apply the same rule to limits, timeouts, capacities, retry counts, and ranges: test relationships and behavior, not duplicated literals.
+
+For concurrency limits, test both kinds of behavior when relevant:
+
+1. **Configuration enforcement:** invalid or excessive requested values are handled correctly.
+2. **Runtime enforcement:** observed peak concurrency never exceeds the configured limit.
+
+Use a test-controlled limit and measure the behavior relative to that limit. Do not merely assert today's default value.
+
+## Avoid semantic duplication across layers
+
+Each behavior should have a clear test owner.
+
+- Parser tests own parsing cases.
+- Validator tests own validation rules.
+- Domain tests own transformations and invariants.
+- Adapter tests own external integration behavior.
+- Orchestrator tests own coordination and failure propagation.
+- CLI tests own argument and configuration mapping.
+- End-to-end tests prove that representative assembled workflows work.
+
+Higher-level tests should not repeat every lower-level case. A single intentional policy change should not require unrelated edits across many test files.
+
+Tests that are individually reasonable may still be collectively redundant. Evaluate the marginal value of each additional test in light of the protection already provided by the rest of the suite.
+
+## Use test doubles deliberately
+
+Choose the least elaborate test double that provides the required control or observation.
+
+As a default:
+
+1. Prefer real collaborators when they are fast and deterministic.
+2. Use small in-memory fakes when realistic stateful behavior is helpful.
+3. Use stubs when a dependency only needs to provide controlled responses.
+4. Use mocks when the interaction itself is contractual.
+
+Mocks are appropriate when the contract includes facts such as:
+
+- a notification is sent exactly once;
+- a transaction is committed only after successful writes;
+- cancellation reaches a subprocess;
+- an expensive API is called no more than once; or
+- a security audit event is emitted.
+
+Do not use mocks merely to isolate every object or reproduce the implementation's call graph.
+
+## Go-specific guidance
+
+Use:
+
+- table-driven tests for meaningful behavioral categories and boundaries;
+- `t.TempDir()` for real filesystem behavior;
+- `httptest.Server` for realistic HTTP interactions;
+- fuzz tests for parsers, normalization, path handling, and broad input spaces;
+- golden files only when the complete output is intentionally stable;
+- integration tests where correctness depends on component interaction; and
+- a small number of representative end-to-end tests.
+
+Avoid exact error-string assertions unless the wording is itself contractual. Prefer `errors.Is`, `errors.As`, typed errors, or structured error fields.
+
+At CLI boundaries, prefer exit classifications, structured output, and the smallest stable semantic fragment needed to identify the error. Do not snapshot complete diagnostic wording unless it is contractual.
+
+Golden-file updates must require an explicit local flag. CI must not update golden files automatically, and reviewers must inspect the semantic diff before accepting an update.
+
+Keep tests readable and direct. Test helpers and fixture frameworks must earn their own maintenance cost; do not build elaborate test infrastructure for small or isolated needs.
+
+## Coverage
+
+Coverage is a diagnostic, not a target.
+
+Use it to find untested critical branches and unexpectedly weak packages. Do not write low-value tests solely to increase a percentage, and do not infer test quality from coverage alone.
+
+Pure domain logic will often warrant higher coverage than CLI wiring or external adapters. Uneven coverage is acceptable when it reflects risk.
+
+Increasing coverage is valuable only when the newly covered behavior protects a meaningful risk at an acceptable cost.
+
+## Regression tests
+
+A bug fix should normally include a regression test that fails before the fix and passes afterward.
+
+Retain the test when the defect could realistically recur and its consequences justify the ongoing cost. Prefer the narrowest durable test of the violated contract or invariant; do not preserve accidental implementation details from the original bug.
+
+Not every historical bug requires a permanent test. If the underlying design has made recurrence impossible, the test has become redundant, or a stronger invariant test now subsumes it, remove or consolidate it.
+
+## Deleting or rewriting tests
+
+Tests are maintained code, not permanent historical artifacts.
+
+Delete or rewrite a test when its maintenance cost exceeds the confidence it provides.
+
+Strong candidates include tests that:
+
+- require updates after harmless internal changes;
+- directly assert private constants without protecting a real contract;
+- duplicate the same policy across several layers;
+- verify mock choreography rather than outcomes;
+- snapshot large amounts of incidental output;
+- test trivial private helpers already exercised through stable package behavior;
+- protect risks already covered more effectively elsewhere;
+- are flaky, misleading, obsolete, or disproportionately expensive to diagnose; or
+- no longer correspond to a plausible failure mode.
+
+Several brittle tests may encode one genuine requirement. Replace them with one durable behavior-level or invariant test rather than preserving all of them.
+
+Deleting a low-value test can improve the quality of the suite by reducing noise, maintenance burden, and friction around legitimate change.
+
+## Reviewing a proposed test
+
+Use the following questions when the value, boundary, or durability of a proposed test is not self-evident. Significant test additions should be reviewable against them, but written answers are not required for every routine test.
+
+1. What realistic defect would it catch?
+2. How likely is that defect?
+3. How consequential would it be?
+4. Is the behavior already protected elsewhere?
+5. At which layer should this behavior be owned?
+6. Does the test assert a durable contract or an incidental implementation detail?
+7. Could the implementation be refactored without changing the behavior and without editing this test?
+8. What should cause this test to fail?
+9. What legitimate changes should not cause this test to fail?
+10. What ongoing maintenance, execution, and diagnostic cost will the test impose?
+11. Is there a smaller or more direct test that protects the same risk?
+
+Do not add the test when its expected lifetime cost exceeds its expected protective value.
+
+When deciding not to test plausible behavior, record or be able to explain why the risk is low, already protected, obvious, reversible, or cheaper to detect elsewhere.
+
+## Definition of sufficient
+
+A test suite is sufficient when:
+
+- important contracts and invariants are protected;
+- meaningful boundaries and failure modes are exercised;
+- realistic and consequential regressions are credibly protected against silent recurrence;
+- behavior involving data integrity, destructive operations, compatibility, security, concurrency, idempotency, and recovery is credibly protected;
+- important external boundaries have realistic integration coverage;
+- representative complete workflows are tested;
+- failures provide useful signal rather than redundant noise;
+- legitimate internal changes usually do not require test edits; and
+- additional tests would mostly repeat existing protection or preserve inconsequential implementation details.
+
+Sufficiency is a risk judgment, not a coverage percentage or test count. Reassess it as the application, its users, and the consequences of failure evolve.
+
+The governing rule is:
+
+> Test heavily where failure is consequential, subtle, or difficult to detect after the fact. Test lightly where failure is obvious, reversible, and inexpensive—and retain no test whose lifetime cost exceeds the confidence it provides.

docs/roadmap/future.md

@@ -0,0 +1,223 @@
+# Future Work
+
+Current Notarius behavior is documented in the canonical README, CLI,
+configuration, operations, internal, and integration docs. This roadmap records
+future work only. Items are ordered roughly by current value and specificity,
+not as committed release dates.
+
+## Near-Term D&D Pipeline
+
+### Combat Enemy Ledger
+
+- Add a D&D artifact that identifies enemies faced during combat and supports
+  an end-of-session encounter ledger.
+- Track each enemy's observed state using a small controlled vocabulary such as
+  `active`, `killed`, `fled`, `captured`, or `incapacitated`, while preserving
+  an explicit unresolved state when the transcript does not establish an
+  outcome.
+- Preserve the evidence for enemy participation and state changes rather than
+  inferring a terminal outcome from combat ending or an enemy disappearing
+  from the conversation.
+- Define how repeated mentions, groups of unnamed enemies, summoned or allied
+  creatures, and the same enemy appearing in multiple combats affect identity
+  and ledger entries.
+- Evaluate whether the ledger should be extracted directly, derived from
+  combat-turn artifacts, or use a sequential pipeline that consumes combat
+  turns and the normalized NPC registry as grounding references.
+
+### Location Extraction
+
+- Add a D&D artifact for locations visited by the party or otherwise mentioned
+  in the transcript.
+- Distinguish observed visits from references, plans, recalled places, and
+  uncertain or inferred locations so a mention alone is not reported as a
+  visit.
+- Preserve transcript evidence for each visit or mention and reconcile aliases,
+  nested places, and repeated appearances without collapsing distinct
+  locations that share a generic name.
+- Define how the location artifact should ground later narrative reports and
+  whether future event artifacts should retain canonical location identities.
+
+### Evaluate Spell Extraction And Normalization
+
+- Evaluate ordinary extraction retries and the completed normalization path
+  against a human-reviewed transcript set before adding repair-aware retries or
+  an LLM-backed semantic validator.
+- Maintain a small set of human-reviewed transcripts and outputs for prompt,
+  validator, and normalizer development. Treat model-quality review as an
+  iterative human evaluation aid, not a deterministic correctness gate.
+
+### Evaluate The Shared D&D Scene Plan
+
+- Reassess whether one shared scene plan provides enough context for NPC,
+  spell, combat, interaction, and scene-description lanes after real-world use.
+  Add more complex chunking only in response to demonstrated failures.
+
+## Cross-Cutting LLM Runtime
+
+### Deterministic Prompt Session Identity
+
+- Replace the source-document-ID default for prompt sessions with one
+  predictable, procedurally generated session ID for the complete
+  source-processing workload.
+- Preserve an explicit non-empty `--session-id` as the highest-precedence
+  override. Otherwise, derive the default only from the effective input module
+  identity and the exact raw input bytes.
+- Use a versioned, bounded representation such as
+  `notarius:v1:<sha256(input-module + NUL + raw-input)>`. The exact encoding
+  must fit PromptKit's session length contract and must not embed source
+  content.
+- Keep the derived session stable across runs, pipelines, selected lanes,
+  ordered steps, retries, resume, recomputation, LLM profiles, reasoning
+  overrides, and output, debug, or cache settings.
+- Do not include file-backed references, generated references, reference
+  contents, or the composition of a reference bundle in session derivation.
+  References may change between prompt calls within one pipeline without
+  changing routing affinity.
+- Resolve the authoritative session before checkpoint construction and use the
+  same value for checkpoint runtime identity, every prompt-facing module,
+  PromptKit's direct session field, the compatibility `session_id` prompt
+  variable, run-manifest metadata, and debug metadata.
+- Keep routing identity separate from cache and checkpoint content identity.
+  Exact prompt prefixes, reference contents, model settings, and other
+  generation-affecting inputs must continue to participate in their existing
+  hashes and checkpoint fingerprints even though they do not change the
+  session.
+- Treat the generated value as a provider-visible, stable pseudonymous
+  correlation identifier. Do not introduce an installation-specific HMAC or
+  secret unless a concrete multi-tenant or privacy requirement justifies
+  sacrificing deterministic identity across installations.
+
+### Raise The Default Application-Wide LLM Limit
+
+- Raise the default `concurrency.total_llm` value from 1 to 16 so ordinary
+  single-backend runs can use PromptKit's expected OpenRouter capacity and
+  lower-capacity local backends without an unnecessarily narrower Notarius
+  limit.
+- Keep the Notarius application-wide scheduler mandatory and require
+  `total_llm` to remain a positive integer. Do not make the default unlimited:
+  endpoint-only profiles, an unrestricted local backend, injected clients, and
+  aggregate work across several backends may have no narrower PromptKit limit.
+- Continue defaulting `concurrency.stage_workers.extract` to the effective
+  `total_llm`, making its default 16 as part of the same change. Preserve an
+  explicit lower extract-worker setting when an operator wants less queued or
+  concurrent extraction work.
+- Define effective provider concurrency as the intersection of the Notarius
+  application-wide limit, the selected PromptKit backend limit when present,
+  and the work made available by stage execution. A Notarius limit of 16 does
+  not narrow a backend already limited to 16, while a local backend limited to
+  4 remains bounded at 4.
+- Treat the default as an application-wide safety ceiling across profiles,
+  backends, modules, retries, and validators. A run that intentionally needs
+  the combined capacity of several backends may configure a higher
+  `total_llm` and an appropriate extract-worker count explicitly.
+- Retain the existing configuration and environment override surfaces. Update
+  canonical configuration, operations, and internal documentation together
+  when the default changes.
+- Reconsider decoupling the extract-worker default from `total_llm` only after
+  mixed-backend workloads demonstrate a need for a high global emergency
+  ceiling with a lower default work-production rate.
+
+## Shared Normalization And Quality Work
+
+### Generic LLM-Assisted Deduplication
+
+- Add a reusable normalizer that asks an LLM to identify duplicate sets in a
+  list and propose one replacement element for each set.
+- Define the minimum domain-neutral input contract, initially an ordered list
+  whose elements have stable unique IDs. Artifact-kind registrations or
+  adapters may expose that structure without moving domain rules into the
+  generic package.
+- Keep mutation deterministic: parse and validate the model's duplicate groups,
+  require every referenced ID to exist, reject overlapping or malformed groups,
+  prevent unrelated insertion or deletion, and apply only approved replacement
+  operations in code.
+- Preserve provenance needed for audit and downstream validation, and emit
+  warnings describing every collapsed group.
+- Evaluate batching and context-window limits before applying the normalizer to
+  large artifact collections.
+
+The model may use its own domain knowledge to judge semantic duplication; the
+generic implementation is responsible only for the common proposal contract,
+safety checks, and deterministic application of accepted changes.
+
+### Validation And Review
+
+- Add domain validators and production default chains alongside each new D&D
+  artifact.
+- Add production LLM-backed validators only when a concrete review policy
+  benefits from model judgment and deterministic checks are insufficient.
+- Add validator diagnostics and timing summaries if operators need more detail
+  than the current [durable output bundle](../integrations/json-output.md)
+  provides.
+- Add validator compatibility metadata if deployments need config-time proof
+  that a validator is suitable for a particular stage, module, or artifact
+  kind.
+- Add media-type validators when non-JSON artifact representations are
+  introduced.
+
+## Further Reference Evolution
+
+- Make prior-run artifacts easier to bind as references without changing the
+  existing module-facing reference-item contract.
+- Add structured or parsed references, such as typed NPC registries, rosters,
+  or spell catalogs, when opaque UTF-8 prompt material is no longer sufficient.
+- Add per-slot or per-chunk inclusion policies so large references are not
+  repeated in every prompt unnecessarily.
+- Add token budgeting and model context-window management for reference
+  content.
+- Add reference caching, preprocessing, summarization, embedding, or retrieval
+  only when reference size and observed model behavior justify them.
+- Extend generated references to prior-run artifacts or derived summaries only
+  after same-run ordered handoffs establish the required provenance and
+  lifecycle semantics.
+
+## Design Considerations To Revisit
+
+These concerns are relevant to ordered artifact dependencies but are not
+committed near-term features.
+
+### Cross-artifact identity links
+
+Evaluate whether downstream D&D artifacts should retain canonical NPC IDs from
+the generated NPC reference in addition to normalized display names. Any such
+contract must define player-character, unknown-actor, missing-NPC, and
+superseded-identity behavior before implementation. Deterministic validation
+may confirm that a linked ID exists in the consumed NPC artifact, but the link
+must never substitute for transcript evidence that the downstream event
+occurred.
+
+### Artifact contract evolution
+
+Define compatibility and migration policy before generated-reference chains
+must span multiple schema versions or long-lived historical artifacts. The
+policy should address stable identifier semantics, which schema changes permit
+checkpoint reuse, when an older artifact may be decoded or adapted, and when a
+producer or all dependents must be recomputed. Do not add a general migration
+framework until an actual contract change requires one.
+
+## Blue-Sky Platform And Operations
+
+These ideas are intentionally less specified. Promote one into an earlier
+section only after a concrete workflow, contract, and priority emerge.
+
+### Platform Extensions
+
+- Additional input adapters, such as Markdown or note-export formats.
+- Additional output encoders.
+- Concurrent cross-lane entity normalization or broader workflow composition.
+- Batching or specialized context-window controls for LLM-backed validators.
+
+### Distribution And Operations
+
+- Packaged release artifacts for alpha distribution.
+- A documented versioning and release process.
+- Optional generated example-output fixtures with a regeneration procedure.
+- Additional diagnostics or reporting views.
+
+### Workspace And Storage
+
+- Default-idempotent run behavior with an explicit force override.
+- Remote workspace storage.
+- Workspace garbage collection and archival policies.
+- Cross-machine checkpoint reuse.

docs/roadmap/implementation.md

@@ -0,0 +1,586 @@
+# PromptKit v0.5 Implementation Plan
+
+## Objective
+
+Implement the target state in
+[PromptKit v0.5 Integration And LLM Profile Policy](promptkit.md). Each numbered
+stage is intended to be one implementation prompt for a GPT-5.6-Terra coding
+agent. Complete stages in order and leave the repository buildable, tested, and
+internally coherent after every stage.
+
+Follow [Architecture](../policy/architecture.md),
+[Testing Policy](../policy/testing.md), and
+[Documentation Policy](../policy/documentation.md) throughout. Preserve
+unrelated user changes. Use `apply_patch` for source and documentation edits,
+run `gofmt` on changed Go files, and add only tests that protect the behaviors
+and risks assigned to that stage.
+
+Do not implement the separate deterministic session-ID or default-concurrency
+roadmap items as part of this plan. Do not perform paid or credentialed LLM
+calls.
+
+## Background Summary
+
+Notarius currently pins PromptKit v0.3.0, calls `Prepare` and then `Run` for one
+completion, validates profiles through a synthetic prompt, has no application
+fallback profile source, and accepts LLM profiles only at individual bindings
+or through the run-wide CLI override. PromptKit v0.5.0 is source-compatible
+with the current tree; a temporary v0.5.0 module override has already passed
+`go test ./...`.
+
+The implementation must nevertheless treat the upstream optional-parameter
+change as intentional: unset `temperature`, `max_tokens`, and `top_p` remain
+unset and are omitted from compatible provider requests. Do not restore the old
+implicit `top_p: 1` default.
+
+## Stage 1: Upgrade The PromptKit Dependency
+
+### Goal
+
+Establish a clean PromptKit v0.5.0 baseline before adopting its new APIs.
+
+### Work
+
+- Update `go.mod` and `go.sum` from PromptKit v0.3.0 to v0.5.0 and run
+  `go mod tidy`.
+- Change the PromptKit built-in profile-catalog marker in
+  `internal/framework/llm/promptkit_profile_fingerprint.go` to identify
+  v0.5.0. This deliberately invalidates LLM checkpoints tied to the prior
+  catalog identity.
+- Review PromptKit-facing compile errors or test failures against the v0.4.0
+  and v0.5.0 release guides. Do not adopt prepared execution, inspection, or
+  fallback profiles in this stage.
+- Replace the existing test assertion for one exact built-in fingerprint hash
+  with durable assertions that the fingerprint is deterministic, non-empty,
+  non-secret, and changes when a semantic profile source changes. Do not add a
+  new version-constant or exact-hash change detector.
+- Update `docs/integrations/pkg-promptkit.md` to pin and link v0.5.0 and state
+  the implemented dependency-level behavior: unset optional sampling controls
+  are provider defaults. Do not document later stages as implemented.
+- Update any other canonical text that explicitly claims the dependency is
+  v0.3.0, but defer descriptions of unimplemented v0.5 APIs.
+
+### Tests And Validation
+
+- `go test ./internal/framework/llm ./internal/cli`
+- `go test ./...`
+- `go vet ./...`
+- `go build ./cmd/notarius`
+- `rg -n 'promptkit v0\.3\.0|promptkit@v0\.3\.0|PromptKit v0\.3\.0' .`
+- `git diff --check`
+
+### Completion Criteria
+
+- The repository directly pins v0.5.0 and all default offline checks pass.
+- The profile-source fingerprint identifies the new upstream catalog without a
+  brittle literal-hash test.
+- Current documentation no longer identifies v0.3.0 as the supported version.
+
+## Stage 2: Execute One Frozen Prepared Snapshot
+
+### Goal
+
+Make Notarius debug details and generation use one exact PromptKit preparation.
+
+### Work
+
+- Refactor `PromptKitClient.CompleteStructured` to call
+  `PrepareExecution`, immediately defer `Discard`, obtain a caller-owned
+  `Details` value, and execute with `RunPrepared`.
+- Preserve the existing Notarius request mapping, cancellation precedence,
+  validation classification, raw structured bytes, response decoding,
+  profile recording, usage reporting, and credential redaction.
+- Ensure every preparation, execution, validation, empty-result, and decode
+  error retains useful Notarius prompt context without exposing prepared handle
+  state or secrets.
+- Use `errors.As` to obtain `*promptkit.CapacityError` on admission rejection.
+  Preserve `contracts.ErrLLMCapacityExceeded` as the stable classification and
+  add a nonblank backend ID only to safe application-owned diagnostic context.
+  Do not expose `promptkit.CapacityError` outside the LLM adapter.
+- Update `docs/internal/llm.md` and the implemented-mechanics portion of
+  `docs/integrations/pkg-promptkit.md` to describe the single frozen execution
+  snapshot and structured capacity adaptation.
+
+### Tests And Validation
+
+- Adapt existing PromptKit client tests to the prepared-execution path.
+- Retain or add one behavioral test proving that the debug prompt details match
+  the request actually passed to generation when a backing prompt source could
+  otherwise change between independent preparations. Test the resulting
+  snapshot consistency, not a private helper call count.
+- Retain capacity tests proving `errors.Is` reaches
+  `contracts.ErrLLMCapacityExceeded`, the selected backend can appear in safe
+  diagnostic context, and provider calls are not made after rejected
+  admission.
+- Run `go test ./internal/framework/llm` and
+  `go test -race ./internal/framework/llm`.
+- Run `go test ./...` and `git diff --check`.
+
+### Completion Criteria
+
+- `CompleteStructured` no longer calls independent `Prepare` and `Run`
+  operations for one request.
+- Debug prompt material and generation result originate from the same frozen
+  PromptKit snapshot.
+- Capacity remains a provider-neutral Notarius error classification.
+
+## Stage 3: Replace Synthetic Profile Validation With Inspection
+
+### Goal
+
+Validate profiles through PromptKit's exact profile-inspection boundary and
+centralize engine profile-source construction.
+
+### Work
+
+- Introduce a small provider-adapter-owned profile inspection or validation
+  function in `internal/framework/llm`. Its public internal signature must use
+  Notarius-owned configuration and result/error types rather than returning
+  PromptKit types to the CLI.
+- Share the code that applies `profile_dir`, `profile_file`, and registered
+  backend options between the production PromptKit engine and the inspection
+  engine. Preserve the mutual-exclusion and local-backend rules.
+- Change CLI explicit-profile preflight to use `Engine.InspectProfile` through
+  that LLM boundary.
+- Remove `profileCheckPromptID`, `profileCheckPromptFS`, the `testing/fstest`
+  production dependency, and the synthetic `Prepare` request.
+- Preserve distinct, useful errors for an absent profile, invalid profile,
+  unknown backend registration, cancellation, and invalid profile source.
+- Do not require `api_key_env` to be populated during configuration validation.
+  Inspection may report credential requirements internally, but actual
+  preparation remains responsible for credential availability before a model
+  call.
+- Update current-behavior sections in `docs/internal/cli.md` and
+  `docs/internal/llm.md`. Keep field definitions in `docs/config.md`.
+
+### Tests And Validation
+
+- Replace synthetic-prompt tests with profile inspection tests covering:
+  configured local backend success; missing local backend failure; absent
+  profile; malformed profile; and an otherwise valid profile whose credential
+  environment variable is intentionally unset.
+- Prove validation performs no provider HTTP call and remains offline.
+- Run `go test ./internal/framework/llm ./internal/cli` and `go test ./...`.
+- Run `git diff --check`.
+
+### Completion Criteria
+
+- No production synthetic profile-check prompt remains.
+- Profile validation uses the same ordinary profile source and backend
+  registrations as execution.
+- Configuration validation succeeds for structurally valid profiles without
+  reading credential values.
+
+## Stage 4: Add Application Fallback Profile Asset Plumbing
+
+### Goal
+
+Allow module families to register application-owned fallback profile YAML
+without placing domain policy in generic LLM code.
+
+### Work
+
+- Extend `internal/framework/llm.AssetRegistry` with a separate fallback
+  profile source collection, registration method, flattened filesystem, and
+  safe content digest.
+- Reuse the existing asset-source path validation and flattening behavior where
+  appropriate. Reject invalid roots, unreadable assets, and duplicate flattened
+  paths. Do not parse PromptKit profile YAML in Notarius.
+- Add `promptkit.WithFallbackProfileFS` to production engine options only when
+  at least one fallback profile source is registered.
+- Supply the identical assembled fallback source to the profile-inspection
+  engine. Adjust CLI composition so pipeline-aware profile validation can use
+  the production LLM asset registry without exposing PromptKit types.
+- Extend profile-source checkpoint identity to include the exact fallback
+  profile asset digest in addition to the PromptKit catalog marker and operator
+  source. Keep the resulting fingerprint hash-only and path/content/credential
+  free.
+- Keep operator source precedence owned by PromptKit. Do not implement profile
+  merging or duplicate PromptKit source resolution in Notarius.
+- Update `docs/internal/llm.md` only for the new implemented generic asset and
+  fingerprint mechanics. No domain fallback exists until Stage 5.
+
+### Tests And Validation
+
+- Add focused AssetRegistry tests for successful flattening, invalid roots,
+  duplicate paths, and hash changes when fallback bytes change.
+- Add adapter-level tests showing that the fallback filesystem reaches both
+  execution construction and inspection construction.
+- Extend checkpoint tests to prove fallback content changes profile-source
+  identity without exposing raw YAML or paths. Use relational comparisons, not
+  a fixed hash literal.
+- Run `go test ./internal/framework/llm ./internal/cli` and `go test ./...`.
+- Run `git diff --check`.
+
+### Completion Criteria
+
+- Generic plumbing can carry application fallback profiles while remaining
+  unaware of D&D IDs or model settings.
+- Inspection, execution, and checkpoint identity use the same fallback asset
+  source.
+
+## Stage 5: Adopt The D&D `dnd-extraction` Fallback
+
+### Goal
+
+Give the D&D module family one stable embedded workload profile that operators
+can replace.
+
+### Work
+
+- Add a D&D-owned embedded PromptKit profile asset with ID `dnd-extraction`
+  under `internal/modules/dnd`. Use the exact baseline defined in
+  `promptkit.md`: OpenRouter, `openai/gpt-5.6-luna`, no explicit reasoning
+  effort, a 240-second timeout, flex service tier, and no selected temperature,
+  token limit, or `top_p`. The omitted reasoning value intentionally allows
+  OpenAI's backend to apply its `medium` default.
+- Register the profile filesystem from the D&D registrar through the generic
+  fallback profile asset boundary. Keep D&D policy out of
+  `internal/framework/llm` and the CLI composition root.
+- Change every maintained D&D LLM prompt definition—including scene chunking,
+  all D&D extractors, and NPC normalization—from the model-named default to
+  `default_profile: dnd-extraction`.
+- Add an integration-level profile-resolution test proving that:
+  - the fallback resolves when no operator source defines the ID;
+  - a valid operator profile with the same ID wins completely; and
+  - an invalid matching operator profile fails rather than falling through.
+- Test through Notarius's assembled production assets and PromptKit boundary;
+  do not duplicate every upstream source-precedence case.
+- Update the implemented profile ownership and prompt-default behavior in
+  `docs/internal/dnd.md`, `docs/internal/llm.md`, and
+  `docs/integrations/pkg-promptkit.md`. Defer the complete operator walkthrough
+  and examples to Stage 10.
+
+### Tests And Validation
+
+- Run focused D&D prompt preparation tests and the production composition
+  tests.
+- Run `go test ./internal/modules/dnd/... ./internal/framework/llm
+  ./internal/cli`.
+- Run `go test ./...`.
+- Verify `rg -n 'default_profile: gemini-2-flash' internal/modules/dnd`
+  returns no matches.
+- Run `git diff --check`.
+
+### Completion Criteria
+
+- All maintained D&D prompts use the application-owned logical profile ID.
+- The fallback works without an operator profile and remains authoritatively
+  overridable by a matching valid operator definition.
+
+## Stage 6: Introduce Module Execution-Class Metadata
+
+### Goal
+
+Make each production module's ability to use an LLM statically discoverable
+without yet changing profile inheritance.
+
+### Work
+
+- Add `ExecutionClass contracts.ExecutionClass` to `pipeline.ModuleSpec` and
+  preserve it through normalization, cloning, catalogs, registries, JSON/debug
+  views, and lookup helpers.
+- In this transitional stage only, allow an omitted execution class to
+  normalize to deterministic so existing test-only fixtures can be migrated in
+  Stage 7 without breaking the repository midway.
+- Explicitly classify every production module:
+  - D&D scene chunking, every D&D extractor, and D&D NPC normalization as
+    `llm_backed`;
+  - all other current production input, chunk, merge, normalize, and output
+    modules as `deterministic`.
+- Update production module specification tests and production catalog tests to
+  assert the semantic class alongside stage, artifact kind, and capabilities.
+- Add catalog lookup support needed by later resolution to retrieve a selected
+  module's execution class by stage and key without constructing it.
+- Do not implement pipeline-level profile inheritance or reject deterministic
+  profiles yet.
+- Update `docs/internal/modules.md` and `docs/internal/dnd.md` to identify
+  execution class as registered module metadata, while noting only implemented
+  uses.
+
+### Tests And Validation
+
+- Run module registration/spec tests across generic, Seriatim, and D&D
+  families.
+- Run `go test ./internal/framework/pipeline ./internal/modules/...`.
+- Run `go test ./...` and `git diff --check`.
+
+### Completion Criteria
+
+- Every production module has an explicit correct execution class.
+- Catalog consumers can retrieve that class without a concrete module
+  instance.
+- Test-only omitted classes remain the only temporary compatibility behavior.
+
+## Stage 7: Enforce Execution Metadata And Remove Runtime Probing
+
+### Goal
+
+Finish the execution-class contract so missing metadata cannot cause future
+profile drift.
+
+### Work
+
+- Update every framework, CLI, and integration test module specification to
+  declare an explicit execution class appropriate to the fake behavior.
+- Change module-spec validation so an empty or unsupported execution class is a
+  registration error. Remove the transitional deterministic default from
+  Stage 6.
+- Replace the chunk runner's special `ChunkExecutionClassProvider` probe with
+  specification-derived behavior. Remove the now-redundant provider interface,
+  implementation methods, and tests when they have no remaining consumer.
+- Ensure chunk producer provenance remains unchanged: it records a non-empty
+  effective binding profile for an LLM-backed chunker, while a deterministic
+  chunker records no profile. A profile selected only through the prompt
+  default remains represented by PromptKit's actual-profile manifest rather
+  than being invented as an explicit chunk binding.
+- Review helper constructors and fixtures for opportunities to set execution
+  class once without obscuring the class under test. Do not introduce an
+  elaborate test-spec framework.
+- Update internal documentation if the removal changes any described runtime
+  mechanics.
+
+### Tests And Validation
+
+- Add or retain focused registration tests for missing and invalid execution
+  classes.
+- Retain chunk-plan provenance tests for LLM-backed and deterministic
+  chunkers.
+- Run `go test ./internal/framework/pipeline ./internal/modules/...`.
+- Run `go test ./...`, `go vet ./...`, and `git diff --check`.
+
+### Completion Criteria
+
+- No registered module specification relies on an implicit execution class.
+- Pipeline metadata, not a concrete runtime type assertion, owns module
+  execution classification.
+
+## Stage 8: Resolve Programmatic Pipeline Profile Defaults
+
+### Goal
+
+Implement profile inheritance and precedence inside the pipeline resolver
+before exposing the field through YAML configuration.
+
+### Work
+
+- Add an optional trimmed `LLMProfile` field to
+  `pipeline.PipelineProfile`. Add a non-empty runtime override field to
+  `pipeline.ResolveOptions` so all precedence decisions occur in the resolver
+  rather than through pre-resolution mutation.
+- After module selection, `--only` filtering, default validator-chain
+  selection, and validator compatibility resolution, apply effective profiles
+  to every selected input, chunk, extract, merge, normalize, output, and
+  validator binding according to the precedence in `promptkit.md`.
+- Apply profiles only when the selected module or validator execution class is
+  `llm_backed`.
+- Reject a binding-specific `llm_profile` on any deterministic module or
+  validator. Do not reject or inspect an unused pipeline default when no
+  selected LLM-backed binding consumes it.
+- Leave an LLM-backed binding empty when no CLI, binding, or pipeline profile is
+  selected so PromptKit can use the prompt's `default_profile`.
+- Store the effective values on resolved bindings before digest construction.
+  Do not add a second inheritance decision to execution.
+- Ensure semantically equivalent repeated binding profiles and one inherited
+  default produce the same resolved pipeline digest. Ensure any changed
+  effective profile changes the digest.
+- Do not modify file configuration or CLI parsing in this stage.
+
+### Tests And Validation
+
+- Add pipeline package tests for the complete precedence matrix:
+  runtime override; binding-specific exception; pipeline default; prompt
+  fallback; and deterministic bindings.
+- Cover default and explicitly configured validator chains, all relevant stage
+  categories, `--only` lane selection, unused defaults, deterministic-profile
+  rejection, and semantic digest equivalence.
+- Prefer table-driven package-level tests over assertions on private traversal
+  helpers.
+- Run `go test ./internal/framework/pipeline` and `go test ./...`.
+- Run `git diff --check`.
+
+### Completion Criteria
+
+- Programmatic pipelines resolve one canonical effective profile policy.
+- Only LLM-backed resolved bindings can contain a profile.
+- Runtime override, binding, pipeline, and prompt precedence is unambiguous and
+  digest-stable.
+
+## Stage 9: Expose Pipeline Defaults Through Configuration And CLI
+
+### Goal
+
+Make the profile-default workflow available to operators while preserving
+validation and override behavior.
+
+### Work
+
+- Add optional `pipelines.<id>.llm_profile` support to the version 4 file
+  configuration model. Use presence-aware decoding so an explicitly set blank
+  value is rejected, while omission remains valid.
+- Preserve the field through file application, configuration cloning,
+  effective configuration, and programmatic profile copies without aliasing or
+  trimming drift.
+- Remove `applyLLMProfileOverride`. Pass the CLI override through the resolver's
+  runtime-override input so deterministic bindings are never populated.
+- Update effective profile-ID collection to cover every selected LLM-backed
+  module stage and LLM-backed validator, including future LLM-backed input and
+  output modules. Do not inspect deterministic or unselected profiles.
+- Ensure `run`, `config validate --pipeline`, resume/checkpoint identity, and
+  relevant dry preflight paths all use the same resolved effective profiles.
+- Preserve `--llm-profile` as the highest-precedence non-empty run-wide
+  override and preserve binding-specific profiles as exceptions when no CLI
+  override is present.
+- Do not increment the configuration version.
+- Update current configuration and CLI contracts in `docs/config.md` and
+  `docs/cli.md` in the same stage. Link to operations for the deployment
+  workflow rather than duplicating it prematurely.
+
+### Tests And Validation
+
+- Add file-config tests for omission, trimming, explicit blank rejection,
+  unknown-key behavior, cloning, and round-trip application.
+- Add effective-config and CLI contract tests for precedence, LLM-only
+  application, inherited-profile inspection failure before factory execution,
+  `--only`, and digest changes.
+- Retain offline operation and do not require credentials for
+  `config validate --pipeline`.
+- Run `go test ./internal/core/config ./internal/framework/pipeline
+  ./internal/cli`.
+- Run `go test ./...`, `go vet ./...`, and `git diff --check`.
+
+### Completion Criteria
+
+- Operators can select `dnd-extraction` once per pipeline.
+- Configuration and CLI paths share the resolver's precedence policy.
+- Unknown effective profiles fail preflight, while deterministic and unused
+  profiles do not cause spurious inspection.
+
+## Stage 10: Complete Operator Documentation, Examples, And Decision Record
+
+### Goal
+
+Make the implemented workflow understandable, copyable, and maintainable
+without duplicating canonical facts.
+
+### Work
+
+- Create an ADR using the next sequential number for the durable decision to
+  use workload-oriented pipeline defaults with operator-overridable application
+  fallback profiles. Record context, decision, alternatives, and consequences;
+  do not turn the ADR into a field reference or implementation log.
+- Complete `docs/config.md` as the canonical owner of profile-source fields,
+  `pipelines.<id>.llm_profile`, validation, and precedence.
+- Complete `docs/operations.md` with an operator workflow that distinguishes
+  Notarius embedded prompts, Notarius fallback profiles, PromptKit built-ins,
+  and deployment filesystem profiles. Include production/development/local use
+  of the same `dnd-extraction` ID, credential handling, absolute-path guidance,
+  and the fact that current relative profile paths use the process working
+  directory rather than the configuration file's directory.
+- Complete `docs/integrations/pkg-promptkit.md` with the v0.5.0 boundary,
+  prepared execution, inspection, fallback and ordinary source precedence,
+  optional provider controls, capacity adaptation, and compatibility policy.
+- Update `docs/internal/configuration.md`, `docs/internal/pipeline.md`,
+  `docs/internal/cli.md`, `docs/internal/llm.md`, `docs/internal/modules.md`, and
+  `docs/internal/dnd.md` only for their owned implementation details. Link to
+  canonical configuration, operations, and upstream format contracts rather
+  than restating them.
+- Keep exactly the existing two D&D configuration examples. Add
+  `llm_profile: dnd-extraction` to the minimal and complete pipelines and remove
+  the now-redundant model-named binding override from the complete example.
+- Add one secret-free maintained operator profile at
+  `examples/profiles/dnd-extraction.yml`. It should be a complete valid profile
+  for the same logical ID and may mirror the embedded baseline; its purpose is
+  to demonstrate file ownership and format, not claim automatic environment
+  detection. Link it from the configuration and operations documentation.
+- If the complete example selects the external profile file, use a path that
+  is valid for the documented repository-root invocation and explicitly note
+  the working-directory rule. Keep the minimal example dependent only on the
+  embedded fallback.
+- Add or extend maintained-example validation so both configuration examples
+  and the profile YAML are checked without generation or credentials.
+- Remove the now-implemented `Pipeline-Level LLM Profile Defaults` section from
+  `docs/roadmap/future.md`. Preserve the unrelated deterministic session and
+  concurrency items.
+- Do not delete `promptkit.md` or this implementation plan during the feature
+  implementation; retire them only after post-implementation review.
+
+### Tests And Validation
+
+- Run maintained example/configuration tests and relevant CLI help/parser
+  tests.
+- Run `go test ./...`.
+- Run `rg -n 'gemini-2-flash' examples docs` and review every remaining match
+  for intentional model-policy or historical context.
+- Run `rg -n 'v0\.3\.0|profileCheckPrompt|applyLLMProfileOverride' .` and resolve
+  stale production or current-documentation matches.
+- Verify all new links and `git diff --check`.
+
+### Completion Criteria
+
+- Every current fact has one canonical documentation owner.
+- Operators can distinguish and deploy all profile layers without reading Go
+  source.
+- Both maintained configurations and the maintained external profile are valid,
+  secret-free, and tested offline.
+- Implemented profile work no longer remains in `future.md`.
+
+## Stage 11: Final Verification And Quality Review
+
+### Goal
+
+Verify the complete migration as one integrated change and correct only defects
+or omissions found during that review.
+
+### Work
+
+- Review the final diff against every acceptance criterion in `promptkit.md`.
+- Confirm provider-specific PromptKit types remain inside the LLM integration
+  boundary and D&D policy remains inside the D&D module family.
+- Confirm execution and inspection receive identical ordinary, fallback, and
+  backend configuration.
+- Confirm no paths, profile YAML, endpoints, credentials, or prepared handle
+  state leak into fingerprints or ordinary diagnostics.
+- Confirm all production module specs have explicit correct execution classes
+  and every resolved deterministic binding is profile-free.
+- Confirm prompt default, pipeline default, binding override, and CLI override
+  behavior through representative assembled configurations.
+- Review tests for redundancy and remove obsolete synthetic-prompt,
+  runtime-probe, exact-hash, or duplicated upstream-behavior tests superseded by
+  stronger contract tests.
+- Perform an optional manual D&D quality comparison if credentials and an
+  evaluation transcript are deliberately supplied. Record no private input or
+  credential material, and do not make this comparison a completion gate.
+
+### Validation Commands
+
+```sh
+gofmt -w <changed-go-files>
+go test ./...
+go test -race ./internal/framework/llm ./internal/core/config ./internal/framework/pipeline ./internal/cli
+go vet ./...
+go build ./cmd/notarius
+git diff --check
+```
+
+Also run focused stale-contract searches:
+
+```sh
+rg -n 'gitea.maximumdirect.net/eric/promptkit v0\.3\.0|PromptKit v0\.3\.0' .
+rg -n 'default_profile: gemini-2-flash|profileCheckPrompt|applyLLMProfileOverride' internal docs examples
+```
+
+Review any matches rather than deleting intentional historical references
+blindly.
+
+### Completion Criteria
+
+- All automated checks pass offline and without real credentials.
+- The implemented behavior matches `promptkit.md` with no known architecture,
+  provenance, checkpoint, profile-precedence, or documentation gap.
+- Any optional live evaluation is clearly separate from correctness testing.
+
+## Open Questions
+
+None. The roadmap decisions are sufficient to implement every stage without an
+additional product or architecture choice.

docs/roadmap/mvp.md

@@ -1,29 +0,0 @@
-# Future Work
-
-Current Notarius behavior is documented in the canonical README, CLI,
-configuration, operations, internal, and integration docs. This roadmap records
-future work only.
-
-## Candidate Product Work
-
-- Additional input adapters, such as Markdown or note-export formats.
-- Additional D&D extractors beyond spell casts.
-- Cross-lane entity normalization.
-- Cross-chunk semantic deduplication.
-- Configurable validator chains with production validator modules.
-- Multiple effective LLM profiles in one run.
-- Parallel execution where it preserves deterministic manifests and diagnostics.
-- Additional output encoders.
-
-## Candidate Operational Work
-
-- Packaged release artifacts for alpha distribution.
-- A documented versioning and release process.
-- Optional generated example output fixtures with a regeneration procedure.
-- Additional diagnostics or reporting views if operator workflows need them.
-
-## Non-Goals To Revisit Deliberately
-
-- A general workflow language.
-- Structural module selection through ad hoc run flags.
-- Storing secrets in config files, diagnostics, manifests, or examples.

docs/roadmap/promptkit.md

@@ -0,0 +1,318 @@
+# PromptKit v0.5 Integration And LLM Profile Policy
+
+## Purpose
+
+This roadmap defines the target state for upgrading Notarius from PromptKit
+v0.3.0 to v0.5.0 and adopting the upstream runtime and profile facilities that
+directly improve Notarius. It also defines the application policy for stable,
+domain-oriented LLM profile names, operator overrides, pipeline inheritance,
+profile validation, provider defaults, checkpoint identity, and documentation.
+
+The ordered work needed to reach this state belongs in
+[the implementation plan](implementation.md). Current behavior remains defined
+by the canonical documentation outside `docs/roadmap/` until the corresponding
+work is implemented.
+
+## Background
+
+Notarius currently pins PromptKit v0.3.0. Its adapter prepares a request once
+for debug material and then independently runs the original request, causing
+PromptKit to prepare the same logical call a second time. The CLI validates an
+explicit profile by preparing a synthetic prompt. PromptKit profile selection
+can be repeated on individual module bindings or replaced for one invocation
+with `--llm-profile`, but a configured pipeline cannot yet declare one inherited
+profile policy.
+
+PromptKit v0.4.0 and v0.5.0 add the upstream boundaries needed to improve these
+areas:
+
+- [v0.4.0](https://gitea.maximumdirect.net/eric/promptkit/src/tag/v0.5.0/docs/releases/v0.4.0.md)
+  adds opaque prepared executions, exact profile and prompt inspection, and a
+  typed backend-capacity error;
+- [v0.5.0](https://gitea.maximumdirect.net/eric/promptkit/src/tag/v0.5.0/docs/releases/v0.5.0.md)
+  adds application fallback profile filesystems and stops sending unset
+  optional sampling controls as framework-selected provider values; and
+- the [v0.5.0 format contract](https://gitea.maximumdirect.net/eric/promptkit/src/tag/v0.5.0/docs/formats.md)
+  defines the resulting profile-source and execution-setting precedence.
+
+A source-compatibility test of the current Notarius repository against
+PromptKit v0.5.0 completed successfully. The work is therefore primarily an
+intentional runtime and configuration migration rather than a repair for a
+breaking Go API change.
+
+## Goals
+
+- Pin and document PromptKit v0.5.0 as Notarius's supported upstream contract.
+- Execute the exact prepared request snapshot whose safe details are recorded
+  in Notarius debug material.
+- Validate configured PromptKit profiles through the upstream inspection API
+  without synthetic prompts, provider calls, or credential-value access.
+- Give Notarius an application-owned, operator-overridable
+  `dnd-extraction` profile fallback.
+- Let a pipeline choose one default LLM profile without repeating that ID on
+  every LLM-backed binding.
+- Apply profile inheritance and run-wide overrides only where the resolved
+  module or validator can use an LLM.
+- Preserve accurate checkpoint invalidation, effective profile provenance,
+  redaction, cancellation, concurrency, and provider-neutral module contracts.
+- Provide operators with one clear deployment pattern for production,
+  development, and local profile definitions.
+
+## Target End State
+
+### PromptKit Runtime Boundary
+
+Notarius depends on PromptKit v0.5.0 and uses its public APIs rather than
+reimplementing source or execution resolution.
+
+For each structured completion, the adapter:
+
+1. builds one PromptKit run request from the provider-neutral Notarius request;
+2. calls `PrepareExecution` once;
+3. immediately arranges an idempotent `Discard` for every unexecuted handle;
+4. obtains credential-redacted `Details` for debug and response metadata; and
+5. calls `RunPrepared` so generation uses that exact frozen snapshot.
+
+The debug prompt and successful result therefore describe the same selected
+profile, rendered messages, input bytes, session, output contract, and effective
+settings even when a filesystem-backed source changes concurrently. PromptKit
+handle types remain private to `internal/framework/llm`.
+
+PromptKit admission failures continue to match Notarius's provider-neutral
+`ErrLLMCapacityExceeded` contract. When PromptKit supplies a `CapacityError`,
+the adapter obtains the normalized backend ID through `errors.As` and may add it
+to safe application-owned diagnostics without parsing upstream error wording.
+The backend ID does not become a provider-specific module contract.
+
+### Optional Provider Controls
+
+Notarius accepts PromptKit v0.5.0's new behavior for `temperature`,
+`max_tokens`, and `top_p`: an unset setting is omitted from compatible provider
+requests and the provider chooses its own default. Notarius does not restore
+PromptKit's former implicit `top_p: 1` value globally.
+
+An operator who requires a particular value specifies it in the selected
+PromptKit profile. The application fallback described below intentionally
+leaves these controls unset. A human-reviewed D&D extraction comparison should
+be performed after the upgrade, but paid or nondeterministic model output is
+not part of the default automated test suite.
+
+### Profile Inspection
+
+Pipeline-aware configuration validation uses `Engine.InspectProfile` for every
+effective explicit profile ID. It verifies that the profile exists, parses and
+validates, resolves its backend and target, and is compatible with the engine's
+registered backends. It does not create a synthetic prompt, load prompt inputs,
+contact a provider, or require credential values to exist in the validation
+process environment.
+
+Credential availability is execution-time state. PromptKit preparation still
+enforces the selected profile's credential contract before generation. This
+keeps `notarius config validate` useful in build and deployment validation
+environments where secrets are deliberately absent.
+
+PromptKit construction for inspection and execution uses one shared internal
+profile-source and backend-option path. The CLI does not expose PromptKit public
+types across the Notarius LLM boundary merely to perform inspection.
+
+`InspectPrompt` is not adopted merely because it exists. It remains available
+for a later, separately defined module-to-prompt interface preflight if a
+concrete validation requirement justifies that additional contract.
+
+### Application And Operator Profile Sources
+
+Notarius embeds one ordinary PromptKit YAML profile with the stable ID
+`dnd-extraction`. It is an application fallback registered through
+`WithFallbackProfileFS`, is owned by the D&D module family, and initially
+preserves the current effective D&D baseline:
+
+- backend: PromptKit's built-in `openrouter` backend;
+- model: `openai/gpt-5.6-luna`;
+- reasoning effort: unset, allowing OpenAI's backend to apply its default of
+  `medium`;
+- generation timeout: 240 seconds;
+- service tier: `flex`; and
+- no application-selected `temperature`, `max_tokens`, or `top_p`.
+
+All maintained D&D LLM prompt definitions use `dnd-extraction` as their
+`default_profile`. The ID communicates workload intent rather than a provider,
+model, or environment. Changing the embedded fallback is an intentional
+Notarius execution-policy change and participates in checkpoint identity.
+
+Effective profile definitions resolve in PromptKit's order:
+
+1. programmatic in-memory profiles used by tests or explicit consumers;
+2. the operator source configured by `promptkit.profile_file` or
+   `promptkit.profile_dir`;
+3. the Notarius application fallback source; and
+4. PromptKit's embedded built-in catalog.
+
+Only an absent ID falls through to the next source. A matching profile is a
+complete definition: fields are not merged with a lower-precedence definition,
+and a malformed matching operator profile fails rather than silently selecting
+the application fallback.
+
+Production, development, and local deployments should normally provide
+different complete definitions for the same `dnd-extraction` ID. An operator
+source is optional because the application fallback keeps the maintained D&D
+workflow usable, but a deployment that needs an intentional model or backend
+policy should configure its own definition.
+
+### Domain Ownership And Asset Assembly
+
+The D&D fallback profile remains under `internal/modules/dnd` and is registered
+by the D&D registrar, consistent with ADR-0004. Generic LLM plumbing knows how
+to collect and flatten application fallback profile filesystems but contains no
+D&D model or policy knowledge.
+
+The shared asset registry detects invalid roots, unreadable sources, and
+duplicate flattened paths. PromptKit remains responsible for strict profile
+YAML parsing, duplicate profile-ID detection, source precedence, and effective
+target resolution. The same assembled fallback source is supplied to runtime
+execution and CLI profile inspection.
+
+### Explicit Module Execution Metadata
+
+Every registered input, chunk, extract, merge, normalize, and output module
+declares one required execution class: `deterministic` or `llm_backed`.
+Validator registrations continue to declare the same distinction through their
+validator specifications.
+
+The registered specification is authoritative for configuration resolution.
+Current production classifications are:
+
+- the D&D scene chunker, all D&D extractors, and the D&D NPC normalizer are
+  LLM-backed;
+- the Seriatim input adapter, generic chunker, all current mergers, all other
+  current normalizers, and the JSON output encoder are deterministic; and
+- current validators retain their declared classifications.
+
+Missing or unsupported execution metadata is a registration error. Explicitly
+assigning `llm_profile` to a deterministic module or validator is a pipeline
+resolution error. The framework does not infer execution class by inspecting
+domain package names or concrete implementation types at runtime.
+
+The module specification replaces the chunk runner's special runtime
+execution-class probe. Effective resolved bindings already express the result:
+only LLM-backed bindings may retain a non-empty profile.
+
+### Pipeline-Level Profile Default
+
+Configuration version 4 gains one optional non-empty pipeline field:
+
+```yaml
+pipelines:
+  dnd-session:
+    llm_profile: dnd-extraction
+```
+
+No configuration-version increment is required because the field is additive
+and existing files remain valid. An explicitly present blank value is invalid.
+
+For every selected LLM-backed module and validator, the effective profile uses
+this precedence:
+
+1. non-empty run-wide `--llm-profile` override;
+2. binding-specific `llm_profile`;
+3. pipeline-level `llm_profile`; and
+4. the prompt definition's `default_profile`, represented by an empty effective
+   Notarius binding profile.
+
+The run-wide override and inherited pipeline default never attach to a
+deterministic binding. Binding-specific exceptions remain available when one
+operation needs a different cost, latency, quality, backend, or reasoning
+policy.
+
+Inheritance is resolved after module and validator selection, including
+`--only` lane filtering, but before effective-pipeline validation, digest
+construction, explicit-profile inspection, checkpoint construction,
+preparation, execution, or provenance capture. Only profiles used by selected
+LLM-backed bindings are inspected. An unused pipeline default in a pipeline
+with no selected LLM-backed work does not require an otherwise unused profile
+to exist.
+
+The resolved pipeline contains effective binding profiles rather than a second
+runtime inheritance mechanism. Two pipelines that differ only by spelling the
+same effective policy once as a pipeline default and once on every LLM-backed
+binding have the same semantic resolved digest. Changing an effective profile
+changes the digest and applicable checkpoint identity.
+
+### Provenance And Checkpoints
+
+The PromptKit profile-source checkpoint fingerprint covers:
+
+- the PromptKit v0.5.0 built-in profile catalog identity;
+- exact application fallback profile asset content; and
+- exact configured operator profile YAML content, when present.
+
+The existing local-backend target fingerprint remains separate and continues
+to exclude scheduling-only concurrency limits. Fingerprints contain hashes and
+stable markers, not profile contents, filesystem paths, endpoints, credentials,
+or other secrets.
+
+Changing the PromptKit version, application fallback, operator profile, or
+effective pipeline profile makes incompatible LLM checkpoints ineligible for
+reuse. The dependency upgrade is expected to invalidate checkpoints produced
+under v0.3.0.
+
+Successful run manifests continue to record only profiles actually selected by
+PromptKit, including their effective model, backend, and reasoning metadata.
+Debug output reports the same effective execution snapshot used for generation.
+
+### Operator Documentation And Examples
+
+Canonical documentation clearly distinguishes:
+
+- Notarius prompt and schema assets embedded in the application;
+- Notarius application fallback profiles embedded in the application;
+- PromptKit's own embedded built-in profiles; and
+- operator profile files on the deployment filesystem.
+
+The configuration reference owns the pipeline field, profile-source fields,
+validation rules, and precedence. Operations owns deployment layout, working
+directory behavior, credentials, and environment-specific profile management.
+The PromptKit integration document owns the pinned upstream contract and
+source-precedence boundary. Internal documents describe asset registration,
+resolution, inspection, prepared execution, fingerprinting, and tests without
+duplicating user-facing field definitions.
+
+The maintained examples continue to include only the minimal and complete D&D
+configurations. They use the stable `dnd-extraction` policy, and one maintained
+PromptKit profile file under `examples/` demonstrates an operator override.
+Examples remain secret-free and are validated without live provider calls.
+
+## Out Of Scope
+
+- Implementing the separate deterministic prompt-session identity roadmap
+  item.
+- Changing the default `concurrency.total_llm` value; PromptKit's retained
+  OpenRouter capacity of 16 remains relevant to that separate item.
+- Adding model evaluation as a deterministic or CI correctness gate.
+- Automatically selecting production, development, or local environments.
+  Deployment configuration chooses the operator profile source.
+- Profile inheritance, partial profile merging, or cross-profile aliases.
+- Exposing PromptKit types to modules, validators, durable output contracts, or
+  public configuration structures.
+- Adopting `InspectPrompt` without a separately justified prompt-interface
+  validation contract.
+
+## Acceptance Criteria
+
+- Notarius builds and its offline test suite passes with PromptKit v0.5.0.
+- Every structured completion executes the exact snapshot used for safe debug
+  prompt details.
+- Profile preflight uses profile inspection and no synthetic prompt.
+- The embedded `dnd-extraction` fallback resolves without an operator source,
+  and a matching valid operator profile replaces it completely.
+- Every production module has explicit, correct execution metadata.
+- Pipeline, binding, CLI, and prompt-default precedence behaves as defined for
+  modules and validators, while deterministic bindings remain profile-free.
+- Effective profiles participate in pipeline digests, profile inspection,
+  checkpoint identity, debug records, and run provenance at the appropriate
+  boundaries.
+- The dependency and application fallback changes invalidate incompatible old
+  checkpoints without exposing profile or credential content.
+- Canonical documentation and maintained examples accurately describe and
+  exercise the implemented operator workflow.
+- Default tests remain deterministic, offline, credential-free, and focused on
+  Notarius-owned behavior rather than duplicating PromptKit's upstream suite.

docs/troubleshooting.md

@@ -1,215 +0,0 @@
-# Troubleshooting
-
-This guide maps common implemented failure modes to inspection steps and fixes.
-For command syntax, see [CLI Reference](cli.md). For YAML fields and
-environment overrides, see [Configuration](config.md). For output and
-diagnostics layout, see [Operations](operations.md).
-
-## Config File Not Found
-
-Symptom:
-
-```text
-notarius: config file not found; pass --config or set NOTARIUS_CONFIG
-```
-
-Fix:
-
-- Pass `--config path/to/config.yml`.
-- Or set `NOTARIUS_CONFIG` to a readable file.
-- Or install a config at `/usr/local/etc/notarius/config.yml`.
-
-If the message says the config path is a directory or is not available, correct
-the path or file permissions.
-
-## Unsupported Or Invalid Config
-
-Symptoms include:
-
-- `unsupported config version`
-- `config version is required`
-- `field <name> not found`
-- `total LLM concurrency must be greater than zero`
-- `diagnostics retention "<value>" is not supported`
-
-Fix:
-
-- Use `version: 1`.
-- Remove unknown YAML fields.
-- Validate with:
-
-```sh
-go run ./cmd/notarius config validate --config path/to/config.yml
-```
-
-## Unknown Pipeline
-
-Symptom:
-
-```text
-notarius: pipeline "..." is not configured
-```
-
-Fix:
-
-- List configured pipeline IDs:
-
-```sh
-go run ./cmd/notarius pipelines list --config path/to/config.yml
-```
-
-- Use one of those IDs in `notarius run <pipeline-id>`.
-- Check indentation under the top-level `pipelines` map.
-
-## Unknown Or Incompatible Module
-
-Symptoms mention a module key, pipeline slot, lane, capability, or `not
-registered`.
-
-Fix:
-
-- Validate the pipeline against the production module catalog:
-
-```sh
-go run ./cmd/notarius config validate \
-  --config path/to/config.yml \
-  --pipeline dnd-session
-```
-
-- Use only implemented production module keys listed in
-  [Configuration](config.md#implemented-production-modules).
-- Check that artifact lanes include an `extract` binding.
-
-## Invalid `--only`
-
-Symptoms include:
-
-- `--only must contain comma-separated non-empty artifact lane IDs`
-- `--only requires --pipeline`
-- `selected artifact lane`
-
-Fix:
-
-- Use comma-separated lane IDs with no empty entries:
-
-```sh
-go run ./cmd/notarius run dnd-session \
-  --config path/to/config.yml \
-  --input path/to/input.json \
-  --only spells
-```
-
-- For `config validate`, include `--pipeline` when using `--only`.
-- Confirm the lane ID exists under `pipelines.<id>.artifacts`.
-
-## Seriatim Input Validation Failure
-
-Symptoms include `seriatim input`, `parse JSON`, `segments must not be empty`,
-or validation errors naming a segment field.
-
-Fix:
-
-- Compare the input to
-  [examples/seriatim-minimal-transcript.json](../examples/seriatim-minimal-transcript.json).
-- Ensure the JSON has a `metadata` object and a non-empty `segments` array.
-- Each segment needs a non-empty `id`, non-empty `speaker`, non-empty `text`,
-  non-negative numeric `start`, and non-negative numeric `end`.
-- Segment IDs must be unique and must not contain leading or trailing
-  whitespace.
-- `end` must be greater than or equal to `start`.
-
-## Missing LLM Base URL Or Model
-
-Symptoms include:
-
-- `LLM profile "default" base URL must not be empty`
-- `LLM profile "default" model must not be empty`
-- `base URL must be valid`
-
-Fix:
-
-- Set `base_url` and `model` in `llm_profiles.default`.
-- Or set `NOTARIUS_LLM_DEFAULT_BASE_URL` and
-  `NOTARIUS_LLM_DEFAULT_MODEL`.
-- If a profile needs authentication, set `api_key_env` in YAML or set
-  `NOTARIUS_LLM_DEFAULT_API_KEY`.
-
-## LLM Profile Override Failure
-
-Symptom:
-
-```text
-notarius: LLM profile override "..." is not configured
-```
-
-Fix:
-
-- Add the profile under `llm_profiles`.
-- Or use an existing profile ID with `--llm-profile`.
-
-Current runs require exactly one distinct effective LLM profile. If a pipeline
-uses several profiles, run with `--llm-profile <id>` or align the bindings in
-configuration.
-
-## Provider HTTP Or Response Failure
-
-Symptoms include:
-
-- `provider request failed`
-- `provider returned status 400`
-- `provider returned status 403`
-- `provider response missing choices`
-- `provider response assistant message content is not valid JSON`
-- `decode structured output`
-
-Fix:
-
-- Confirm the `base_url` points to an OpenAI-compatible endpoint root. Notarius
-  posts to `<base_url>/chat/completions`.
-- Check `model` and provider credentials.
-- Inspect the retained diagnostics `error.log`.
-- For 400 and 403 responses, fix the request configuration or credentials.
-- For 429 and 5xx responses, the client retries according to `max_retries`; if
-  the failure persists, inspect the provider response and adjust capacity,
-  credentials, or model settings.
-- The assistant message content must decode as JSON matching the extractor's
-  structured response schema.
-
-Provider error messages are redacted for configured API key values.
-
-## Output Write Failure
-
-Symptoms include:
-
-- `create output directory`
-- `write output file`
-- `output file name must`
-
-Fix:
-
-- Ensure `--output-dir` points to a directory path or a path that can be
-  created.
-- Check filesystem permissions and available disk space.
-- If diagnostics were retained, inspect `run-report.json`, `run-manifest.json`,
-  and `error.log`.
-
-The CLI rejects unsafe logical output paths before writing files.
-
-## Diagnostics Directory Surprise
-
-Symptom: the diagnostics directory is missing after a successful run.
-
-Fix:
-
-- Check `diagnostics.retention`.
-- With `auto`, successful runs without warnings are removed.
-- Use `diagnostics.retention: always` when every diagnostics run directory
-  should be kept.
-- Use `--diagnostics-dir` to override the configured work directory for a run.
-
-Symptom: diagnostics exist even with `retention: never`.
-
-Explanation:
-
-- Failed runs are retained so that `error.log` and available context can be
-  inspected.

examples/dnd-complete-transcript.json

@@ -0,0 +1,85 @@
+{
+  "metadata": {
+    "id": "session-ravenfall",
+    "title": "The Ravenfall Watchtower"
+  },
+  "segments": [
+    {
+      "id": 1,
+      "start": 0,
+      "end": 14,
+      "speaker": "DM",
+      "text": "Recap: last session, the party learned that Elder Rowan vanished near the Ravenfall watchtower."
+    },
+    {
+      "id": 2,
+      "start": 14,
+      "end": 25,
+      "speaker": "Player",
+      "text": "Out of character, we agree to investigate the watchtower before the next game."
+    },
+    {
+      "id": 3,
+      "start": 25,
+      "end": 39,
+      "speaker": "DM",
+      "text": "Aria and Borin arrive at the ruined Ravenfall watchtower as dusk settles over the road."
+    },
+    {
+      "id": 4,
+      "start": 39,
+      "end": 55,
+      "speaker": "Mira Thorn",
+      "text": "Mira Thorn steps from the doorway and says, \"Elder Rowan warned me that Kesh would return for the relic.\""
+    },
+    {
+      "id": 5,
+      "start": 55,
+      "end": 70,
+      "speaker": "DM",
+      "text": "Mira leads the party to a hidden cache. The party discovers a moonblade and acquires 20 silver pieces."
+    },
+    {
+      "id": 6,
+      "start": 70,
+      "end": 83,
+      "speaker": "Aria",
+      "text": "Aria hands her healing potion to Borin so he can carry it into the tower."
+    },
+    {
+      "id": 7,
+      "start": 83,
+      "end": 96,
+      "speaker": "DM",
+      "text": "Kesh, the goblin captain, orders the raiders to attack. Roll initiative."
+    },
+    {
+      "id": 8,
+      "start": 96,
+      "end": 110,
+      "speaker": "DM",
+      "text": "On Kesh's turn, he strikes Borin with his scimitar. Borin drinks the healing potion on his turn."
+    },
+    {
+      "id": 9,
+      "start": 110,
+      "end": 124,
+      "speaker": "Aria",
+      "text": "Aria casts Cure Wounds on Borin, then invokes Aegis of Emberfall as Kesh closes in."
+    },
+    {
+      "id": 10,
+      "start": 124,
+      "end": 137,
+      "speaker": "DM",
+      "text": "Kesh casts Shield as a reaction against Borin's counterattack, but the party drives the raiders away."
+    },
+    {
+      "id": 11,
+      "start": 137,
+      "end": 150,
+      "speaker": "DM",
+      "text": "After the battle, Aria pays 5 silver pieces to repair the watchtower gate."
+    }
+  ]
+}

examples/dnd-complete.config.yml

@@ -0,0 +1,103 @@
+version: 4
+promptkit:
+  profile_file: ./examples/profiles/dnd-extraction.yml
+concurrency:
+  total_llm: 2
+  stage_workers:
+    extract: 2
+output:
+  directory: ./notarius-output
+cache:
+  chunk_plans:
+    mode: auto
+    directory: ./notarius-cache/chunk-plans
+  checkpoints:
+    enabled: true
+    directory: ./notarius-cache/checkpoints
+debug:
+  directory: ./notarius-debug
+pipelines:
+  dnd-session:
+    llm_profile: dnd-extraction
+    input: seriatim
+    # Stable campaign context is shared by every module that accepts these slots.
+    references:
+      party: ./dnd-party.txt
+      glossary: ./dnd-glossary.txt
+    chunk:
+      module: dnd/scenes
+      retries: 2
+    output:
+      module: json
+      options:
+        include_chunk_map: true
+        evidence_context:
+          enabled: true
+          window_units: 3
+          lanes:
+            - item-events
+            - npcs
+            - spells
+            - combat-turns
+            - npc-interactions
+    steps:
+      # Establish session-wide reference artifacts alongside independent item events.
+      - id: describe-session
+        artifacts:
+          item-events:
+            extract:
+              module: dnd/item-events
+              retries: 2
+            merge: appendorder
+            normalize: dnd/item-events
+          npcs:
+            extract:
+              module: dnd/npcs
+              retries: 2
+            merge: appendorder
+            normalize:
+              module: dnd/npcs
+              retries: 2
+          scene-descriptions:
+            extract:
+              module: dnd/scene-descriptions
+              retries: 2
+            merge: appendorder
+            normalize: dnd/scene-descriptions
+      - id: extract-events
+        # Accepted NPC grounding and scene-description eligibility artifacts are
+        # supplied in memory to their compatible consumers in this step.
+        references:
+          npcs:
+            artifact:
+              step: describe-session
+              lane: npcs
+          scene_descriptions:
+            artifact:
+              step: describe-session
+              lane: scene-descriptions
+        artifacts:
+          spells:
+            extract:
+              module: dnd/spells
+              retries: 2
+              references:
+                spell_catalog: ./dnd-spell-catalog.json
+            merge: appendorder
+            # Stage-local file references are intentionally bound at each stage.
+            normalize:
+              module: dnd/spells
+              references:
+                spell_catalog: ./dnd-spell-catalog.json
+          combat-turns:
+            extract:
+              module: dnd/combat-turns
+              retries: 2
+            merge: appendorder
+            normalize: dnd/combat-turns
+          npc-interactions:
+            extract:
+              module: dnd/npc-interactions
+              retries: 2
+            merge: appendorder
+            normalize: dnd/npc-interactions

examples/dnd-glossary.txt

@@ -0,0 +1,8 @@
+Ravenfall watchtower: a ruined watchtower near the party's current route.
+Mira Thorn: the watchtower's keeper.
+Elder Rowan: a missing local scholar.
+Kesh: a goblin captain leading raiders.
+Moonblade: a blade found in the watchtower's hidden cache.
+Cure Wounds: a healing spell.
+Shield: a defensive reaction spell.
+Aegis of Emberfall: a campaign spell recorded in the supplied catalog overlay.

examples/dnd-minimal.config.yml

@@ -0,0 +1,9 @@
+version: 4
+pipelines:
+  dnd-session:
+    llm_profile: dnd-extraction
+    input: seriatim
+    artifacts:
+      spells:
+        extract: dnd/spells
+        normalize: dnd/spells

examples/dnd-party.txt

@@ -0,0 +1,2 @@
+Aria: party cleric and recurring healer.
+Borin: fighter ally.

examples/dnd-spell-catalog.json

@@ -0,0 +1,21 @@
+{
+  "schema_version": "notarius.dnd.spell-catalog-overlay.v1",
+  "catalogs": [
+    {
+      "id": "notarius.example-campaign",
+      "ruleset": "dnd-5e-2014",
+      "source": {
+        "title": "Notarius example campaign spell names",
+        "version": "1",
+        "url": "",
+        "license": ""
+      },
+      "spells": [
+        {
+          "name": "Aegis of Emberfall",
+          "aliases": ["Emberfall Aegis"]
+        }
+      ]
+    }
+  ]
+}

examples/dnd-spells.config.yml

@@ -1,16 +0,0 @@
-version: 1
-llm_profiles:
-  default:
-    provider: openai-compatible
-    base_url: http://127.0.0.1:1
-    model: fake-model
-pipelines:
-  dnd-session:
-    input: seriatim
-    chunk:
-      module: generic
-      options:
-        max_units: 50
-    artifacts:
-      spells:
-        extract: dnd/spells

examples/profiles/dnd-extraction.yml

@@ -0,0 +1,5 @@
+id: dnd-extraction
+backend: openrouter
+model: openai/gpt-5.6-luna
+timeout_seconds: 240
+service_tier: flex

examples/seriatim-minimal-transcript.json

@@ -5,14 +5,14 @@
   },
   "segments": [
     {
-      "id": "seg-001",
+      "id": 1,
       "start": 0,
       "end": 4,
       "speaker": "Aria",
       "text": "Aria raises her holy symbol and casts Cure Wounds."
     },
     {
-      "id": "seg-002",
+      "id": 2,
       "start": 4,
       "end": 8,
       "speaker": "DM",

go.mod

@@ -1,5 +1,11 @@
 module gitea.maximumdirect.net/eric/notarius
 
-go 1.24.0
+go 1.25.5
 
-require gopkg.in/yaml.v3 v3.0.1
+require (
+	gitea.maximumdirect.net/eric/promptkit v0.5.0
+	github.com/santhosh-tekuri/jsonschema/v6 v6.0.2
+	gopkg.in/yaml.v3 v3.0.1
+)
+
+require golang.org/x/text v0.40.0

go.sum

@@ -1,3 +1,11 @@
+gitea.maximumdirect.net/eric/promptkit v0.5.0 h1:jnpazLyyNhWrB2xzwwtUkNUfktkTdkENTwuSPnKiYrc=
+gitea.maximumdirect.net/eric/promptkit v0.5.0/go.mod h1:R95NM6fbMDGDC0/UomgnSBP6ui2ns+8SZb8bESNvrDQ=
+github.com/dlclark/regexp2 v1.11.0 h1:G/nrcoOa7ZXlpoa/91N3X7mM3r8eIlMBBJZvsz/mxKI=
+github.com/dlclark/regexp2 v1.11.0/go.mod h1:DHkYz0B9wPfa6wondMfaivmHpzrQ3v9q8cnmRbL6yW8=
+github.com/santhosh-tekuri/jsonschema/v6 v6.0.2 h1:KRzFb2m7YtdldCEkzs6KqmJw4nqEVZGK7IN2kJkjTuQ=
+github.com/santhosh-tekuri/jsonschema/v6 v6.0.2/go.mod h1:JXeL+ps8p7/KNMjDQk3TCwPpBy0wYklyWTfbkIzdIFU=
+golang.org/x/text v0.40.0 h1:Ub2Z6/xjgF1WrYQz2nuITOEegKFtiIy+rieRJ5lHZKs=
+golang.org/x/text v0.40.0/go.mod h1:hpnzDAfGV753zIKo+wk3u1bVKCGPbrnF7+7LBF/UHVY=
 gopkg.in/check.v1 v0.0.0-20161208181325-20d25e280405 h1:yhCVgyC4o1eVCa2tZl7eS0r+SDo693bJlVdllGtEeKM=
 gopkg.in/check.v1 v0.0.0-20161208181325-20d25e280405/go.mod h1:Co6ibVJAznAaIkqp8huTwlJQCZ016jof/cbN4VW5Yz0=
 gopkg.in/yaml.v3 v3.0.1 h1:fxVm/GzAzEWqLHuvctI91KS9hhNmmWOoWu0XTYJS7CA=

internal/cli/assembled_spell_pipeline_contract_test.go

@@ -0,0 +1,315 @@
+package cli
+
+import (
+	"context"
+	"encoding/json"
+	"fmt"
+	"reflect"
+	"sort"
+	"strings"
+	"sync"
+	"testing"
+
+	"gitea.maximumdirect.net/eric/notarius/internal/core/artifacts"
+	"gitea.maximumdirect.net/eric/notarius/internal/core/source"
+	"gitea.maximumdirect.net/eric/notarius/internal/framework/contracts"
+	"gitea.maximumdirect.net/eric/notarius/internal/framework/pipeline"
+	"gitea.maximumdirect.net/eric/notarius/internal/modules/dnd"
+	spellnormalize "gitea.maximumdirect.net/eric/notarius/internal/modules/dnd/normalize/spells"
+	spellcatalog "gitea.maximumdirect.net/eric/notarius/internal/modules/dnd/spells/catalog"
+)
+
+const assembledSpellExtractorKey = "test/dnd/spell-casts"
+
+func TestAssembledSpellPipelineNormalizesMergedCasts(t *testing.T) {
+	registries, resolved, extractor := assembledSpellPipeline(t, assembledSpellPipelineOptions{})
+	prepared, err := pipeline.Prepare(resolved, registries, pipeline.ModuleDependencies{})
+	if err != nil {
+		t.Fatalf("Prepare() error = %v, want nil", err)
+	}
+
+	output, err := pipeline.New().Run(context.Background(), pipeline.RunInput{
+		Prepared:       prepared,
+		RawInput:       readRepositoryFile(t, "examples", "seriatim-minimal-transcript.json"),
+		ChunkCacheMode: pipeline.ChunkCacheBypass,
+	})
+	if err != nil {
+		t.Fatalf("Run() error = %v, want nil", err)
+	}
+
+	chunkIndexes := extractor.chunkIndexesSnapshot()
+	sort.Ints(chunkIndexes)
+	if !reflect.DeepEqual(chunkIndexes, []int{0, 1}) {
+		t.Fatalf("extractor chunk indexes = %#v, want two chunk-boundary calls", chunkIndexes)
+	}
+	if output.Manifest.ValidationStatus != "approved" || len(output.Rejected) != 0 || len(output.NormalizeOutputs) != 1 {
+		t.Fatalf("run output = %#v, want approved normalized output without rejections", output)
+	}
+	if output.NormalizeOutputs[0].NormalizerKey != spellnormalize.Key {
+		t.Fatalf("normalized output module = %q, want %q", output.NormalizeOutputs[0].NormalizerKey, spellnormalize.Key)
+	}
+
+	var normalized dnd.SpellList
+	if err := json.Unmarshal(output.NormalizeOutputs[0].Artifact.Content, &normalized); err != nil {
+		t.Fatalf("decode normalized output: %v", err)
+	}
+	if len(normalized.SpellCasts) != 2 {
+		t.Fatalf("normalized casts = %#v, want collapsed duplicate plus distinct evidence", normalized.SpellCasts)
+	}
+	first, distinct := normalized.SpellCasts[0], normalized.SpellCasts[1]
+	if first.Spell != "Cure Wounds" || first.Caster != "  Aria \t" {
+		t.Fatalf("retained cast = %#v, want canonical spell with first occurrence caster", first)
+	}
+	if !reflect.DeepEqual(first.SourceRefs, []source.SourceRef{{SourceID: "session-alpha", StartUnitID: 1, EndUnitID: 1}, {SourceID: "session-alpha", StartUnitID: 2, EndUnitID: 2}}) {
+		t.Fatalf("retained refs = %#v, want sorted complete evidence", first.SourceRefs)
+	}
+	if distinct.Spell != "Cure Wounds" || distinct.Caster != "aria" || !reflect.DeepEqual(distinct.SourceRefs, []source.SourceRef{{SourceID: "session-alpha", StartUnitID: 2, EndUnitID: 2}}) {
+		t.Fatalf("distinct cast = %#v, want separate evidence event", distinct)
+	}
+
+	wantWarningReasons := []string{
+		spellnormalize.ReasonCodeSpellNameCanonicalized,
+		spellnormalize.ReasonCodeSourceReferencesNormalized,
+		spellnormalize.ReasonCodeDuplicateSpellCastCollapsed,
+		"spell_not_near_source",
+	}
+	gotWarningReasons := make([]string, len(output.Warnings))
+	for index, warning := range output.Warnings {
+		gotWarningReasons[index] = warning.ReasonCode
+	}
+	if !reflect.DeepEqual(gotWarningReasons, wantWarningReasons) {
+		t.Fatalf("warnings = %#v, want deterministic normalize and validation warnings", output.Warnings)
+	}
+	if output.Warnings[2].Scope != "spell_casts[0]" || !strings.Contains(output.Warnings[2].Message, "retained input index 0") || !strings.Contains(output.Warnings[2].Message, "removed input indices [1]") {
+		t.Fatalf("duplicate warning = %#v, want retained and removed merged indices", output.Warnings[2])
+	}
+
+	warningsFile := decodeAssembledOutput[struct {
+		Warnings []contracts.Warning `json:"warnings"`
+	}](t, output.OutputFiles, "warnings.json")
+	if !reflect.DeepEqual(warningsFile.Warnings, output.Warnings) {
+		t.Fatalf("warnings file = %#v, run warnings = %#v, want manifest output path to preserve warnings", warningsFile.Warnings, output.Warnings)
+	}
+	manifest := decodeAssembledOutput[artifacts.RunManifest](t, output.OutputFiles, "manifest.json")
+	if len(manifest.ArtifactLanes) != 1 || manifest.ArtifactLanes[0].Normalizer != spellnormalize.Key {
+		t.Fatalf("manifest lanes = %#v, want assembled spell normalizer", manifest.ArtifactLanes)
+	}
+	normalizerMetadata, ok := manifest.ArtifactLanes[0].Metadata["normalizer"].(map[string]any)
+	_, hasOverlayIDs := normalizerMetadata["catalog_overlay_ids"]
+	if !ok || normalizerMetadata["catalog_base_id"] != spellcatalog.SRD5E2014ID || !strings.HasPrefix(stringValue(normalizerMetadata["catalog_digest"]), "sha256:") || !hasOverlayIDs {
+		t.Fatalf("normalizer manifest metadata = %#v, want base ID, digest, and overlay IDs", manifest.ArtifactLanes[0].Metadata)
+	}
+}
+
+func TestAssembledSpellPipelineHonorsNormalizeValidatorOverride(t *testing.T) {
+	registries, resolved, _ := assembledSpellPipeline(t, assembledSpellPipelineOptions{normalizeValidatorOverride: true})
+	var normalizeChain *pipeline.ResolvedValidatorChain
+	for index := range resolved.ValidatorChains {
+		chain := &resolved.ValidatorChains[index]
+		if chain.Stage == pipeline.StageNormalize && chain.ModuleKey == spellnormalize.Key && chain.LaneID == "spells" {
+			normalizeChain = chain
+			break
+		}
+	}
+	if normalizeChain == nil || len(normalizeChain.Validators) != 1 || normalizeChain.Validators[0].Binding.Module != "generic/always_accept" {
+		t.Fatalf("normalize validator chain = %#v, want explicit always-accept override", normalizeChain)
+	}
+
+	prepared, err := pipeline.Prepare(resolved, registries, pipeline.ModuleDependencies{})
+	if err != nil {
+		t.Fatalf("Prepare() error = %v, want nil", err)
+	}
+	output, err := pipeline.New().Run(context.Background(), pipeline.RunInput{
+		Prepared:       prepared,
+		RawInput:       readRepositoryFile(t, "examples", "seriatim-minimal-transcript.json"),
+		ChunkCacheMode: pipeline.ChunkCacheBypass,
+	})
+	if err != nil || output.Manifest.ValidationStatus != "approved" || len(output.Rejected) != 0 || len(output.NormalizeOutputs) != 1 {
+		t.Fatalf("Run() error = %v output = %#v, want approved override run", err, output)
+	}
+	for _, warning := range output.Warnings {
+		if warning.ReasonCode == "spell_not_near_source" {
+			t.Fatalf("warnings = %#v, want explicit validator override to replace default relatedness chain", output.Warnings)
+		}
+	}
+}
+
+func TestAssembledSpellPipelineRejectsUnknownSpellWithoutPromotingAttemptWarning(t *testing.T) {
+	registries, resolved, _ := assembledSpellPipeline(t, assembledSpellPipelineOptions{unknownSpell: true})
+	prepared, err := pipeline.Prepare(resolved, registries, pipeline.ModuleDependencies{})
+	if err != nil {
+		t.Fatalf("Prepare() error = %v, want nil", err)
+	}
+	output, err := pipeline.New().Run(context.Background(), pipeline.RunInput{
+		Prepared:       prepared,
+		RawInput:       readRepositoryFile(t, "examples", "seriatim-minimal-transcript.json"),
+		ChunkCacheMode: pipeline.ChunkCacheBypass,
+	})
+	if err != nil {
+		t.Fatalf("Run() error = %v, want nil", err)
+	}
+	if output.Manifest.ValidationStatus != "rejected" || len(output.NormalizeOutputs) != 0 || len(output.Rejected) != 1 {
+		t.Fatalf("run output = %#v, want one rejected normalize candidate and no normalized output", output)
+	}
+	rejection := output.Rejected[0]
+	if rejection.Stage != string(pipeline.StageNormalize) || rejection.LaneID != "spells" || rejection.ModuleKey != spellnormalize.Key || rejection.ValidatorName != "extract/dnd/spells/catalog" || rejection.ReasonCode != "unknown_spell" {
+		t.Fatalf("rejection = %#v, want durable normalize catalog rejection", rejection)
+	}
+	rejectedFile := decodeAssembledOutput[struct {
+		Rejected []contracts.RejectedOutput `json:"rejected"`
+	}](t, output.OutputFiles, "rejected.json")
+	if !reflect.DeepEqual(rejectedFile.Rejected, output.Rejected) {
+		t.Fatalf("rejected file = %#v, run rejections = %#v, want durable rejection diagnostic", rejectedFile.Rejected, output.Rejected)
+	}
+	for _, warning := range output.Warnings {
+		if warning.ReasonCode == spellnormalize.ReasonCodeSpellNameUnresolved {
+			t.Fatalf("warnings = %#v, want rejected-attempt warning to remain non-durable", output.Warnings)
+		}
+	}
+}
+
+func TestAssembledSpellPipelinePromotesUnknownSpellWarningWhenOverrideAccepts(t *testing.T) {
+	registries, resolved, _ := assembledSpellPipeline(t, assembledSpellPipelineOptions{normalizeValidatorOverride: true, unknownSpell: true})
+	prepared, err := pipeline.Prepare(resolved, registries, pipeline.ModuleDependencies{})
+	if err != nil {
+		t.Fatalf("Prepare() error = %v, want nil", err)
+	}
+	output, err := pipeline.New().Run(context.Background(), pipeline.RunInput{
+		Prepared:       prepared,
+		RawInput:       readRepositoryFile(t, "examples", "seriatim-minimal-transcript.json"),
+		ChunkCacheMode: pipeline.ChunkCacheBypass,
+	})
+	if err != nil {
+		t.Fatalf("Run() error = %v, want nil", err)
+	}
+	if output.Manifest.ValidationStatus != "approved" || len(output.Rejected) != 0 || len(output.NormalizeOutputs) != 1 {
+		t.Fatalf("run output = %#v, want accepted unknown spell with explicit validator override", output)
+	}
+	var normalized dnd.SpellList
+	if err := json.Unmarshal(output.NormalizeOutputs[0].Artifact.Content, &normalized); err != nil {
+		t.Fatalf("decode normalized output: %v", err)
+	}
+	if len(normalized.SpellCasts) != 1 || normalized.SpellCasts[0].Spell != "Mysterious Burst" {
+		t.Fatalf("normalized casts = %#v, want unresolved name preserved", normalized.SpellCasts)
+	}
+	if len(output.Warnings) != 1 || output.Warnings[0].ReasonCode != spellnormalize.ReasonCodeSpellNameUnresolved || output.Warnings[0].Scope != "spell_casts[0]" {
+		t.Fatalf("warnings = %#v, want promoted scoped unresolved-name warning", output.Warnings)
+	}
+	warningsFile := decodeAssembledOutput[struct {
+		Warnings []contracts.Warning `json:"warnings"`
+	}](t, output.OutputFiles, "warnings.json")
+	if !reflect.DeepEqual(warningsFile.Warnings, output.Warnings) {
+		t.Fatalf("warnings file = %#v, run warnings = %#v, want durable unresolved-name warning", warningsFile.Warnings, output.Warnings)
+	}
+}
+
+type assembledSpellPipelineOptions struct {
+	normalizeValidatorOverride bool
+	unknownSpell               bool
+}
+
+func assembledSpellPipeline(t *testing.T, options assembledSpellPipelineOptions) (pipeline.Registries, pipeline.ResolvedPipeline, *assembledSpellExtractor) {
+	t.Helper()
+	components := productionTestComponents(t)
+	extractor := &assembledSpellExtractor{unknownSpell: options.unknownSpell}
+	if err := pipeline.RegisterExtractor[dnd.SpellList](components.registries.Extractors, pipeline.ModuleSpec{
+		Key:            assembledSpellExtractorKey,
+		Stage:          pipeline.StageExtract,
+		ExecutionClass: contracts.ExecutionClassDeterministic,
+		Requires:       []string{"chunks", "source.transcript"},
+		Provides:       []string{"dnd.spell_casts"},
+		ArtifactKind:   dnd.SpellListKind,
+	}, func() (contracts.Extractor[dnd.SpellList], error) {
+		return extractor, nil
+	}); err != nil {
+		t.Fatalf("register assembled extractor: %v", err)
+	}
+
+	normalize := pipeline.Binding(spellnormalize.Key)
+	if options.normalizeValidatorOverride {
+		normalize.Validators = pipeline.ValidatorOverride{
+			Set:        true,
+			Validators: []pipeline.ModuleBinding{pipeline.Binding("generic/always_accept")},
+		}
+	}
+	resolved, err := pipeline.ResolvePipeline(pipeline.PipelineProfile{
+		ID:    "assembled-dnd-spells",
+		Input: pipeline.Binding("seriatim"),
+		Chunk: pipeline.ModuleBinding{Module: "generic", Options: map[string]any{"max_units": 1}},
+		Artifacts: map[string]pipeline.ArtifactLaneProfile{
+			"spells": {Extract: pipeline.Binding(assembledSpellExtractorKey), Normalize: normalize},
+		},
+		Output: pipeline.Binding("json"),
+	}, pipeline.ResolveOptions{}, catalogFromRegistries(components.registries))
+	if err != nil {
+		t.Fatalf("ResolvePipeline() error = %v, want nil", err)
+	}
+	return components.registries, resolved, extractor
+}
+
+type assembledSpellExtractor struct {
+	mu           sync.Mutex
+	chunkIndexes []int
+	unknownSpell bool
+}
+
+func (e *assembledSpellExtractor) Key() string { return assembledSpellExtractorKey }
+
+func (*assembledSpellExtractor) ReferenceSlots() []contracts.ReferenceSlot { return nil }
+
+func (e *assembledSpellExtractor) Extract(ctx context.Context, req contracts.TypedExtractionRequest) (contracts.TypedExtractionResult[dnd.SpellList], error) {
+	if err := ctx.Err(); err != nil {
+		return contracts.TypedExtractionResult[dnd.SpellList]{}, err
+	}
+	if req.Chunk == nil || req.Source == nil {
+		return contracts.TypedExtractionResult[dnd.SpellList]{}, fmt.Errorf("assembled extractor requires source and chunk")
+	}
+	e.mu.Lock()
+	e.chunkIndexes = append(e.chunkIndexes, req.Chunk.Index)
+	e.mu.Unlock()
+	refOne := source.SourceRef{SourceID: req.Source.ID, StartUnitID: 1, EndUnitID: 1}
+	refTwo := source.SourceRef{SourceID: req.Source.ID, StartUnitID: 2, EndUnitID: 2}
+	if e.unknownSpell {
+		if req.Chunk.Index == 0 {
+			return contracts.TypedExtractionResult[dnd.SpellList]{Value: dnd.SpellList{SpellCasts: []dnd.SpellCast{{
+				Caster: "Aria", Spell: "Mysterious Burst", SourceRefs: []source.SourceRef{refOne},
+			}}}}, nil
+		}
+		return contracts.TypedExtractionResult[dnd.SpellList]{Value: dnd.SpellList{SpellCasts: []dnd.SpellCast{}}}, nil
+	}
+	switch req.Chunk.Index {
+	case 0:
+		return contracts.TypedExtractionResult[dnd.SpellList]{Value: dnd.SpellList{SpellCasts: []dnd.SpellCast{{
+			Caster: "  Aria \t", Spell: " cure   wounds ", SourceRefs: []source.SourceRef{refTwo, refOne},
+		}}}}, nil
+	case 1:
+		return contracts.TypedExtractionResult[dnd.SpellList]{Value: dnd.SpellList{SpellCasts: []dnd.SpellCast{
+			{Caster: "aria", Spell: "Cure Wounds", SourceRefs: []source.SourceRef{refOne, refTwo}},
+			{Caster: "aria", Spell: "Cure Wounds", SourceRefs: []source.SourceRef{refTwo}},
+		}}}, nil
+	default:
+		return contracts.TypedExtractionResult[dnd.SpellList]{}, fmt.Errorf("unexpected assembled chunk index %d", req.Chunk.Index)
+	}
+}
+
+func (e *assembledSpellExtractor) chunkIndexesSnapshot() []int {
+	e.mu.Lock()
+	defer e.mu.Unlock()
+	return append([]int(nil), e.chunkIndexes...)
+}
+
+func decodeAssembledOutput[T any](t *testing.T, files []contracts.OutputFile, name string) T {
+	t.Helper()
+	for _, file := range files {
+		if file.Name != name {
+			continue
+		}
+		var value T
+		if err := json.Unmarshal(file.Bytes, &value); err != nil {
+			t.Fatalf("decode %s: %v", name, err)
+		}
+		return value
+	}
+	t.Fatalf("output files = %#v, want %q", files, name)
+	return *new(T)
+}

internal/cli/cache_contract_test.go

@@ -0,0 +1,371 @@
+package cli
+
+import (
+	"bytes"
+	"errors"
+	"fmt"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+
+	"gitea.maximumdirect.net/eric/notarius/internal/framework/chunkplan"
+	"gitea.maximumdirect.net/eric/notarius/internal/framework/pipeline"
+)
+
+func TestRunChunkPlanModePrecedenceAndValidation(t *testing.T) {
+	tests := []struct {
+		name       string
+		fileMode   string
+		envMode    string
+		cliMode    string
+		wantStores int
+	}{
+		{name: "default", wantStores: 1},
+		{name: "file", fileMode: "bypass"},
+		{name: "environment", envMode: "bypass"},
+		{name: "cli", envMode: "refresh", cliMode: "bypass"},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			roots := newStateTestRoots(t)
+			if tt.name == "default" {
+				removeStateTestConfigLine(t, roots.config, "    mode: auto\n")
+			} else if tt.fileMode != "" {
+				replaceStateTestConfigLine(t, roots.config, "    mode: auto\n", "    mode: "+tt.fileMode+"\n")
+			}
+
+			var stores []string
+			opts := newStateTestHarness().options()
+			opts.LookupEnv = func(name string) (string, bool) {
+				if name == "NOTARIUS_CACHE_CHUNK_PLANS_MODE" && tt.envMode != "" {
+					return tt.envMode, true
+				}
+				return "", false
+			}
+			opts.ChunkPlanStoreFactory = func(root string) (pipeline.ChunkPlanStore, error) {
+				stores = append(stores, root)
+				return chunkplan.NewFilesystemStore(root)
+			}
+
+			args := []string{"run", "sample", "--config", roots.config, "--input", roots.input}
+			if tt.cliMode != "" {
+				args = append(args, "--chunk_cache", tt.cliMode)
+			}
+			var stdout, stderr bytes.Buffer
+			if code := RunWithOptions(args, &stdout, &stderr, opts); code != 0 {
+				t.Fatalf("code=%d stdout=%q stderr=%q", code, stdout.String(), stderr.String())
+			}
+			assertStateTestOutput(t, roots.output)
+			if len(stores) != tt.wantStores {
+				t.Fatalf("chunk plan store roots = %v, want %d stores", stores, tt.wantStores)
+			}
+			if tt.wantStores == 1 && stores[0] != roots.plans {
+				t.Fatalf("chunk plan store root = %q, want %q", stores[0], roots.plans)
+			}
+			if tt.wantStores == 0 {
+				assertAbsent(t, roots.plans)
+			}
+		})
+	}
+
+	t.Run("invalid cli syntax is a usage error", func(t *testing.T) {
+		roots := newStateTestRoots(t)
+		var stdout, stderr bytes.Buffer
+		code := RunWithOptions([]string{"run", "sample", "--config", roots.config, "--input", roots.input, "--chunk_cache", "invalid"}, &stdout, &stderr, newStateTestHarness().options())
+		if code != 2 || stdout.Len() != 0 || stderr.Len() == 0 {
+			t.Fatalf("code=%d stdout=%q stderr=%q", code, stdout.String(), stderr.String())
+		}
+		assertNoRunState(t, roots)
+	})
+
+	for _, tt := range []struct {
+		name       string
+		fileConfig bool
+	}{
+		{name: "invalid environment mode"},
+		{name: "invalid file mode", fileConfig: true},
+	} {
+		t.Run(tt.name, func(t *testing.T) {
+			roots := newStateTestRoots(t)
+			opts := newStateTestHarness().options()
+			if tt.fileConfig {
+				replaceStateTestConfigLine(t, roots.config, "    mode: auto\n", "    mode: invalid\n")
+			} else {
+				opts.LookupEnv = func(name string) (string, bool) {
+					if name == "NOTARIUS_CACHE_CHUNK_PLANS_MODE" {
+						return "invalid", true
+					}
+					return "", false
+				}
+			}
+			var stdout, stderr bytes.Buffer
+			code := RunWithOptions([]string{"run", "sample", "--config", roots.config, "--input", roots.input}, &stdout, &stderr, opts)
+			if code != 1 || stdout.Len() != 0 || stderr.Len() == 0 {
+				t.Fatalf("code=%d stdout=%q stderr=%q", code, stdout.String(), stderr.String())
+			}
+			assertNoRunState(t, roots)
+		})
+	}
+}
+
+func TestRunChunkPlanRootSelectionAndFailures(t *testing.T) {
+	t.Run("empty configured root uses the per-user cache root", func(t *testing.T) {
+		roots := newStateTestRoots(t)
+		removeStateTestConfigLine(t, roots.config, fmt.Sprintf("    directory: %q\n", roots.plans))
+		userCache := filepath.Join(t.TempDir(), "user-cache")
+		var stores []string
+		opts := newStateTestHarness().options()
+		opts.UserCacheDir = func() (string, error) { return userCache, nil }
+		opts.ChunkPlanStoreFactory = func(root string) (pipeline.ChunkPlanStore, error) {
+			stores = append(stores, root)
+			return chunkplan.NewFilesystemStore(root)
+		}
+		var stdout, stderr bytes.Buffer
+		code := RunWithOptions([]string{"run", "sample", "--config", roots.config, "--input", roots.input}, &stdout, &stderr, opts)
+		if code != 0 {
+			t.Fatalf("code=%d stdout=%q stderr=%q", code, stdout.String(), stderr.String())
+		}
+		wantRoot := filepath.Join(userCache, "notarius", "chunk-plans")
+		if len(stores) != 1 || stores[0] != wantRoot {
+			t.Fatalf("chunk plan store roots = %v, want [%q]", stores, wantRoot)
+		}
+		assertFile(t, filepath.Join(wantRoot, strings.TrimPrefix(stateTestDigest, "sha256:"), "plan.json"))
+		assertAbsent(t, roots.plans)
+	})
+
+	t.Run("bypass avoids default cache dependencies", func(t *testing.T) {
+		roots := newStateTestRoots(t)
+		removeStateTestConfigLine(t, roots.config, fmt.Sprintf("    directory: %q\n", roots.plans))
+		userCacheCalls := 0
+		storeCalls := 0
+		opts := newStateTestHarness().options()
+		opts.UserCacheDir = func() (string, error) {
+			userCacheCalls++
+			return "", errors.New("user cache must not be resolved")
+		}
+		opts.ChunkPlanStoreFactory = func(string) (pipeline.ChunkPlanStore, error) {
+			storeCalls++
+			return nil, errors.New("chunk plan store must not be constructed")
+		}
+		var stdout, stderr bytes.Buffer
+		code := RunWithOptions([]string{"run", "sample", "--config", roots.config, "--input", roots.input, "--chunk_cache", "bypass"}, &stdout, &stderr, opts)
+		if code != 0 {
+			t.Fatalf("code=%d stdout=%q stderr=%q", code, stdout.String(), stderr.String())
+		}
+		if userCacheCalls != 0 || storeCalls != 0 {
+			t.Fatalf("user cache calls=%d store calls=%d, want none", userCacheCalls, storeCalls)
+		}
+		assertStateTestOutput(t, roots.output)
+		assertAbsent(t, roots.plans)
+	})
+
+	t.Run("user cache resolution failure has context and no output", func(t *testing.T) {
+		roots := newStateTestRoots(t)
+		removeStateTestConfigLine(t, roots.config, fmt.Sprintf("    directory: %q\n", roots.plans))
+		opts := newStateTestHarness().options()
+		opts.UserCacheDir = func() (string, error) { return "", errors.New("cache home unavailable") }
+		var stdout, stderr bytes.Buffer
+		code := RunWithOptions([]string{"run", "sample", "--config", roots.config, "--input", roots.input}, &stdout, &stderr, opts)
+		if code != 1 || !strings.Contains(stderr.String(), "resolve chunk plan root") || !strings.Contains(stderr.String(), "cache home unavailable") {
+			t.Fatalf("code=%d stdout=%q stderr=%q", code, stdout.String(), stderr.String())
+		}
+		assertNoRunState(t, roots)
+	})
+
+	t.Run("store construction failure has context and no output", func(t *testing.T) {
+		roots := newStateTestRoots(t)
+		opts := newStateTestHarness().options()
+		opts.ChunkPlanStoreFactory = func(root string) (pipeline.ChunkPlanStore, error) {
+			return nil, fmt.Errorf("store unavailable")
+		}
+		var stdout, stderr bytes.Buffer
+		code := RunWithOptions([]string{"run", "sample", "--config", roots.config, "--input", roots.input}, &stdout, &stderr, opts)
+		want := fmt.Sprintf("create chunk plan store at %q", roots.plans)
+		if code != 1 || !strings.Contains(stderr.String(), want) || !strings.Contains(stderr.String(), "store unavailable") {
+			t.Fatalf("code=%d stdout=%q stderr=%q", code, stdout.String(), stderr.String())
+		}
+		assertNoRunState(t, roots)
+	})
+
+	t.Run("checkpoint root resolution failure has context and no output", func(t *testing.T) {
+		roots := newStateTestRoots(t)
+		removeStateTestConfigLine(t, roots.config, fmt.Sprintf("    directory: %q\n", roots.checkpoints))
+		opts := newStateTestHarness().options()
+		opts.UserCacheDir = func() (string, error) { return "", errors.New("checkpoint cache unavailable") }
+		result := runStateTest(t, roots, opts, false, true, "bypass")
+		if result.code != 1 || !strings.Contains(result.stderr, "resolve checkpoint root") || !strings.Contains(result.stderr, "checkpoint cache unavailable") {
+			t.Fatalf("code=%d stdout=%q stderr=%q", result.code, result.stdout, result.stderr)
+		}
+		assertNoRunState(t, roots)
+	})
+}
+
+func TestRunAutoReusesPlanWhenRunInputsChange(t *testing.T) {
+	roots := newStateTestRoots(t)
+	data, err := os.ReadFile(roots.config)
+	if err != nil {
+		t.Fatal(err)
+	}
+	configText := replaceRequiredOnce(t, string(data), "    chunk: test/chunk\n", `    chunk:
+      module: test/chunk
+      options:
+        strategy: first
+`)
+	configText = replaceRequiredOnce(t, configText, "    output: test/output\n", `      other:
+        extract: test/extract
+        merge: test/merge
+        normalize: test/normalize
+    output: test/output
+`)
+	if err := os.WriteFile(roots.config, []byte(configText), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	referencePath := filepath.Join(filepath.Dir(roots.input), "reference.txt")
+	if err := os.WriteFile(referencePath, []byte("reference content"), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	harness := newStateTestHarness()
+	var firstStdout, firstStderr bytes.Buffer
+	first := RunWithOptions([]string{"run", "sample", "--config", roots.config, "--input", roots.input}, &firstStdout, &firstStderr, harness.options())
+	if first != 0 {
+		t.Fatalf("first run code=%d stdout=%q stderr=%q", first, firstStdout.String(), firstStderr.String())
+	}
+	configData, err := os.ReadFile(roots.config)
+	if err != nil {
+		t.Fatal(err)
+	}
+	configText = replaceRequiredOnce(t, string(configData), "strategy: first", "strategy: second")
+	if err := os.WriteFile(roots.config, []byte(configText), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	var stdout, stderr bytes.Buffer
+	second := RunWithOptions([]string{
+		"run", "sample", "--config", roots.config, "--input", roots.input,
+		"--only", "items", "--reference", "chunk.cache-reference=" + referencePath,
+	}, &stdout, &stderr, harness.options())
+	if second != 0 {
+		t.Fatalf("second run code=%d stdout=%q stderr=%q", second, stdout.String(), stderr.String())
+	}
+	harness.mu.Lock()
+	chunkCalls := harness.chunkCalls
+	harness.mu.Unlock()
+	if chunkCalls != 1 {
+		t.Fatalf("chunk calls across changed run inputs = %d, want 1", chunkCalls)
+	}
+	assertFile(t, filepath.Join(roots.plans, strings.TrimPrefix(stateTestDigest, "sha256:"), "plan.json"))
+	assertAnyFile(t, roots.output)
+}
+
+func TestRunResumeSelectsConfiguredOrPerUserCheckpointRoot(t *testing.T) {
+	for _, configured := range []bool{true, false} {
+		name := "per-user root"
+		if configured {
+			name = "configured root"
+		}
+		t.Run(name, func(t *testing.T) {
+			roots := newStateTestRoots(t)
+			if !configured {
+				removeStateTestConfigLine(t, roots.config, fmt.Sprintf("    directory: %q\n", roots.checkpoints))
+			}
+			userCache := filepath.Join(t.TempDir(), "user-cache")
+			userCacheCalls := 0
+			opts := newStateTestHarness().options()
+			opts.UserCacheDir = func() (string, error) {
+				userCacheCalls++
+				return userCache, nil
+			}
+			result := runStateTest(t, roots, opts, false, true, "bypass")
+			if result.code != 0 {
+				t.Fatalf("code=%d stdout=%q stderr=%q", result.code, result.stdout, result.stderr)
+			}
+			wantRoot := roots.checkpoints
+			wantCalls := 0
+			if !configured {
+				wantRoot = filepath.Join(userCache, "notarius", "checkpoints")
+				wantCalls = 1
+			}
+			if userCacheCalls != wantCalls {
+				t.Fatalf("user cache calls = %d, want %d", userCacheCalls, wantCalls)
+			}
+			assertAnyFile(t, wantRoot)
+			if !configured {
+				assertAbsent(t, roots.checkpoints)
+			}
+			assertStateTestOutput(t, roots.output)
+		})
+	}
+
+	t.Run("disabled avoids checkpoint root resolution", func(t *testing.T) {
+		roots := newStateTestRoots(t)
+		replaceStateTestConfigLine(t, roots.config, "    enabled: true\n", "    enabled: false\n")
+		removeStateTestConfigLine(t, roots.config, fmt.Sprintf("    directory: %q\n", roots.checkpoints))
+		opts := newStateTestHarness().options()
+		opts.UserCacheDir = func() (string, error) { return "", errors.New("checkpoint cache must not be resolved") }
+		result := runStateTest(t, roots, opts, false, false, "bypass")
+		if result.code != 0 {
+			t.Fatalf("code=%d stdout=%q stderr=%q", result.code, result.stdout, result.stderr)
+		}
+		assertStateTestOutput(t, roots.output)
+		assertAbsent(t, roots.checkpoints)
+	})
+
+	t.Run("resume requires enabled checkpoint recording", func(t *testing.T) {
+		roots := newStateTestRoots(t)
+		replaceStateTestConfigLine(t, roots.config, "    enabled: true\n", "    enabled: false\n")
+		result := runStateTest(t, roots, newStateTestHarness().options(), true, true, "bypass")
+		if result.code != 1 || !strings.Contains(result.stderr, "--resume requires cache.checkpoints.enabled: true") {
+			t.Fatalf("code=%d stdout=%q stderr=%q", result.code, result.stdout, result.stderr)
+		}
+		assertNoRunState(t, roots)
+	})
+}
+
+func TestConfigCommandsDoNotResolveRunState(t *testing.T) {
+	for _, args := range [][]string{
+		{"config", "validate", "--config"},
+		{"pipelines", "list", "--config"},
+	} {
+		name := strings.Join(args[:2], "-")
+		t.Run(name, func(t *testing.T) {
+			roots := newStateTestRoots(t)
+			opts := newStateTestHarness().options()
+			opts.UserCacheDir = func() (string, error) { return "", errors.New("state root must not be resolved") }
+			opts.ChunkPlanStoreFactory = func(string) (pipeline.ChunkPlanStore, error) {
+				return nil, errors.New("chunk plan store must not be constructed")
+			}
+			command := append([]string(nil), args...)
+			command = append(command, roots.config)
+			var stdout, stderr bytes.Buffer
+			code := RunWithOptions(command, &stdout, &stderr, opts)
+			if code != 0 {
+				t.Fatalf("code=%d stdout=%q stderr=%q", code, stdout.String(), stderr.String())
+			}
+			assertNoRunState(t, roots)
+		})
+	}
+}
+
+func replaceStateTestConfigLine(t *testing.T, path, old, new string) {
+	t.Helper()
+	data, err := os.ReadFile(path)
+	if err != nil {
+		t.Fatal(err)
+	}
+	text := replaceRequiredOnce(t, string(data), old, new)
+	if err := os.WriteFile(path, []byte(text), 0o600); err != nil {
+		t.Fatal(err)
+	}
+}
+
+func removeStateTestConfigLine(t *testing.T, path, line string) {
+	replaceStateTestConfigLine(t, path, line, "")
+}
+
+func assertNoRunState(t *testing.T, roots stateTestRoots) {
+	t.Helper()
+	assertAbsent(t, roots.output)
+	assertAbsent(t, roots.plans)
+	assertAbsent(t, roots.checkpoints)
+	assertAbsent(t, roots.debug)
+}

internal/cli/catalog.go

@@ -3,50 +3,55 @@ package cli
 import (
 	"context"
 	"fmt"
-	"strings"
 
 	"gitea.maximumdirect.net/eric/notarius/internal/core/artifacts"
 	"gitea.maximumdirect.net/eric/notarius/internal/core/config"
 	"gitea.maximumdirect.net/eric/notarius/internal/framework/contracts"
 	"gitea.maximumdirect.net/eric/notarius/internal/framework/llm"
 	"gitea.maximumdirect.net/eric/notarius/internal/framework/pipeline"
-	"gitea.maximumdirect.net/eric/notarius/internal/modules/chunk/generic"
+	dndregister "gitea.maximumdirect.net/eric/notarius/internal/modules/dnd/register"
-	"gitea.maximumdirect.net/eric/notarius/internal/modules/extract/dnd/spells"
+	genericregister "gitea.maximumdirect.net/eric/notarius/internal/modules/generic/register"
-	"gitea.maximumdirect.net/eric/notarius/internal/modules/input/seriatim"
+	seriatimregister "gitea.maximumdirect.net/eric/notarius/internal/modules/seriatim/register"
-	"gitea.maximumdirect.net/eric/notarius/internal/modules/merge/appendorder"
-	"gitea.maximumdirect.net/eric/notarius/internal/modules/normalize/noop"
-	jsonoutput "gitea.maximumdirect.net/eric/notarius/internal/modules/output/json"
 )
 
-func productionRegistries() (pipeline.Registries, error) {
+type productionComponents struct {
+	registries pipeline.Registries
+	assets     *llm.AssetRegistry
+}
+
+func newProductionComponents() (productionComponents, error) {
 	registries := pipeline.Registries{
-		Inputs:      pipeline.NewInputAdapterRegistry(),
+		Inputs:           pipeline.NewInputAdapterRegistry(),
-		Chunkers:    pipeline.NewChunkerRegistry(),
+		Chunkers:         pipeline.NewChunkerRegistry(),
-		Extractors:  pipeline.NewExtractorRegistry(),
+		ArtifactCodecs:   pipeline.NewArtifactCodecRegistry(),
-		Mergers:     pipeline.NewMergerRegistry(),
+		ArtifactEvidence: pipeline.NewArtifactEvidenceRegistry(),
-		Normalizers: pipeline.NewNormalizerRegistry(),
+		Extractors:       pipeline.NewExtractorRegistry(),
-		Validators:  pipeline.NewValidatorRegistry(),
+		Mergers:          pipeline.NewMergerRegistry(),
-		Outputs:     pipeline.NewOutputEncoderRegistry(),
+		Normalizers:      pipeline.NewNormalizerRegistry(),
+		Validators:       pipeline.NewValidatorRegistry(),
+		ValidatorChains:  pipeline.NewValidatorChainRegistry(),
+		Outputs:          pipeline.NewOutputEncoderRegistry(),
 	}
-	if err := seriatim.Register(registries.Inputs); err != nil {
+	assets := llm.NewAssetRegistry()
-		return pipeline.Registries{}, fmt.Errorf("register seriatim input: %w", err)
+	registrars := []struct {
+		name     string
+		register func(pipeline.Registries, *llm.AssetRegistry) error
+	}{
+		{name: "generic", register: genericregister.Register},
+		{name: "seriatim", register: seriatimregister.Register},
+		{name: "dnd", register: dndregister.Register},
 	}
-	if err := generic.Register(registries.Chunkers); err != nil {
+	for _, registrar := range registrars {
-		return pipeline.Registries{}, fmt.Errorf("register generic chunker: %w", err)
+		if err := registrar.register(registries, assets); err != nil {
+			return productionComponents{}, fmt.Errorf("register %s module family: %w", registrar.name, err)
+		}
 	}
-	if err := spells.Register(registries.Extractors); err != nil {
+	return productionComponents{registries: registries, assets: assets}, nil
-		return pipeline.Registries{}, fmt.Errorf("register dnd spells extractor: %w", err)
+}
-	}
+
-	if err := appendorder.Register(registries.Mergers); err != nil {
+func productionRegistries() (pipeline.Registries, error) {
-		return pipeline.Registries{}, fmt.Errorf("register appendorder merger: %w", err)
+	components, err := newProductionComponents()
-	}
+	return components.registries, err
-	if err := noop.Register(registries.Normalizers); err != nil {
-		return pipeline.Registries{}, fmt.Errorf("register noop normalizer: %w", err)
-	}
-	if err := jsonoutput.Register(registries.Outputs); err != nil {
-		return pipeline.Registries{}, fmt.Errorf("register json output encoder: %w", err)
-	}
-	return registries, nil
 }
 
 func productionCatalog() (pipeline.ModuleCatalog, error) {
@@ -57,6 +62,11 @@ func productionCatalog() (pipeline.ModuleCatalog, error) {
 	return catalogFromRegistries(registries), nil
 }
 
+func productionPromptAssets() (*llm.AssetRegistry, error) {
+	components, err := newProductionComponents()
+	return components.assets, err
+}
+
 func effectiveCatalog(opts Options) (pipeline.ModuleCatalog, error) {
 	if !isEmptyCatalog(opts.Catalog) {
 		return opts.Catalog, nil
@@ -79,94 +89,88 @@ func effectiveRegistries(opts Options) (pipeline.Registries, error) {
 
 func catalogFromRegistries(registries pipeline.Registries) pipeline.ModuleCatalog {
 	return pipeline.ModuleCatalog{
-		Inputs:      registries.Inputs,
+		Inputs:           registries.Inputs,
-		Chunkers:    registries.Chunkers,
+		Chunkers:         registries.Chunkers,
-		Extractors:  registries.Extractors,
+		ArtifactCodecs:   registries.ArtifactCodecs,
-		Mergers:     registries.Mergers,
+		ArtifactEvidence: registries.ArtifactEvidence,
-		Normalizers: registries.Normalizers,
+		Extractors:       registries.Extractors,
-		Validators:  registries.Validators,
+		Mergers:          registries.Mergers,
-		Outputs:     registries.Outputs,
+		Normalizers:      registries.Normalizers,
+		Validators:       registries.Validators,
+		ValidatorChains:  registries.ValidatorChains,
+		Outputs:          registries.Outputs,
 	}
 }
 
 func registriesFromCatalog(catalog pipeline.ModuleCatalog) pipeline.Registries {
 	return pipeline.Registries{
-		Inputs:      catalog.Inputs,
+		Inputs:           catalog.Inputs,
-		Chunkers:    catalog.Chunkers,
+		Chunkers:         catalog.Chunkers,
-		Extractors:  catalog.Extractors,
+		ArtifactCodecs:   catalog.ArtifactCodecs,
-		Mergers:     catalog.Mergers,
+		ArtifactEvidence: catalog.ArtifactEvidence,
-		Normalizers: catalog.Normalizers,
+		Extractors:       catalog.Extractors,
-		Validators:  catalog.Validators,
+		Mergers:          catalog.Mergers,
-		Outputs:     catalog.Outputs,
+		Normalizers:      catalog.Normalizers,
+		Validators:       catalog.Validators,
+		ValidatorChains:  catalog.ValidatorChains,
+		Outputs:          catalog.Outputs,
 	}
 }
 
 func isEmptyCatalog(catalog pipeline.ModuleCatalog) bool {
 	return catalog.Inputs == nil &&
 		catalog.Chunkers == nil &&
+		catalog.ArtifactCodecs == nil &&
+		catalog.ArtifactEvidence == nil &&
 		catalog.Extractors == nil &&
 		catalog.Mergers == nil &&
 		catalog.Normalizers == nil &&
 		catalog.Validators == nil &&
+		catalog.ValidatorChains == nil &&
 		catalog.Outputs == nil
 }
 
 func isEmptyRegistries(registries pipeline.Registries) bool {
 	return registries.Inputs == nil &&
 		registries.Chunkers == nil &&
+		registries.ArtifactCodecs == nil &&
+		registries.ArtifactEvidence == nil &&
 		registries.Extractors == nil &&
 		registries.Mergers == nil &&
 		registries.Normalizers == nil &&
 		registries.Validators == nil &&
+		registries.ValidatorChains == nil &&
 		registries.Outputs == nil
 }
 
-func productionLLMClientFactory(ctx context.Context, cfg config.Config, profileID string) (contracts.StructuredLLMClient, []artifacts.LLMProfileManifest, error) {
+func productionLLMClientFactoryWithAssets(assets *llm.AssetRegistry) LLMClientFactory {
+	return func(ctx context.Context, cfg config.Config, profileID string, overrides LLMRuntimeOverrides) (contracts.StructuredLLMClient, []artifacts.LLMProfileManifest, error) {
+		return buildProductionLLMClient(ctx, cfg, profileID, overrides, assets)
+	}
+}
+
+func buildProductionLLMClient(ctx context.Context, cfg config.Config, profileID string, overrides LLMRuntimeOverrides, assets *llm.AssetRegistry) (contracts.StructuredLLMClient, []artifacts.LLMProfileManifest, error) {
 	if err := ctx.Err(); err != nil {
 		return nil, nil, err
 	}
-	trimmedID := strings.TrimSpace(profileID)
+	if assets == nil {
-	if trimmedID == "" {
+		return nil, nil, fmt.Errorf("production asset registry must not be nil")
-		trimmedID = pipeline.DefaultLLMProfile
 	}
-
+	recorder := llm.NewLLMProfileRecorder()
-	profile, ok := cfg.LLMProfile(trimmedID)
+	client, err := llm.NewPromptKitClient(llm.PromptKitClientConfig{
-	if !ok {
+		ProfileDir:      cfg.PromptKit.ProfileDir,
-		return nil, nil, fmt.Errorf("LLM profile %q is not configured", trimmedID)
+		ProfileFile:     cfg.PromptKit.ProfileFile,
-	}
+		LocalBackend:    mapPromptKitLocalBackend(cfg.PromptKit.LocalBackend),
-	clientCfg, err := cfg.OpenAICompatibleClientConfig(trimmedID)
+		Assets:          assets,
+		Recorder:        recorder,
+		ReasoningEffort: overrides.ReasoningEffort,
+	})
 	if err != nil {
-		return nil, nil, err
+		return nil, nil, fmt.Errorf("create PromptKit-backed LLM client: %w", err)
 	}
-	client, err := llm.NewOpenAICompatibleClient(clientCfg)
+	scheduler, err := llm.NewScheduler(cfg.Concurrency.TotalLLM)
 	if err != nil {
-		return nil, nil, fmt.Errorf("create LLM client for profile %q: %w", trimmedID, err)
+		return nil, nil, fmt.Errorf("create LLM scheduler: %w", err)
 	}
-
+	return llm.NewScheduledClient(client, scheduler), nil, nil
-	scheduler, err := llm.NewScheduler(effectiveLLMConcurrency(cfg, profile))
-	if err != nil {
-		return nil, nil, fmt.Errorf("create LLM scheduler for profile %q: %w", trimmedID, err)
-	}
-	provider := strings.TrimSpace(profile.Provider)
-	if provider == "" {
-		provider = "openai-compatible"
-	}
-	metadata := []artifacts.LLMProfileManifest{
-		{
-			ID:       trimmedID,
-			Provider: provider,
-			Model:    strings.TrimSpace(profile.Model),
-		},
-	}
-	return llm.NewScheduledClient(client, scheduler), metadata, nil
-}
-
-func effectiveLLMConcurrency(cfg config.Config, profile config.LLMProfile) int {
-	if profile.MaxConcurrency > 0 {
-		return profile.MaxConcurrency
-	}
-	if cfg.Concurrency.TotalLLM > 0 {
-		return cfg.Concurrency.TotalLLM
-	}
-	return 1
 }

internal/cli/command_contract_test.go

@@ -0,0 +1,257 @@
+package cli
+
+import (
+	"bytes"
+	"encoding/json"
+	"fmt"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+)
+
+func TestCommandHelpSpellingsWriteUsageToStdout(t *testing.T) {
+	tests := [][]string{nil, {"help"}, {"--help"}, {"-h"}}
+	for _, args := range tests {
+		name := "no arguments"
+		if len(args) > 0 {
+			name = args[0]
+		}
+		t.Run(name, func(t *testing.T) {
+			var stdout, stderr bytes.Buffer
+			code := RunWithOptions(args, &stdout, &stderr, commandContractOptions(t))
+			if code != 0 || !strings.Contains(stdout.String(), "Usage:") || stderr.Len() != 0 {
+				t.Fatalf("code=%d stdout=%q stderr=%q", code, stdout.String(), stderr.String())
+			}
+		})
+	}
+}
+
+func TestCommandSyntaxErrorsUseStderrAndExitTwo(t *testing.T) {
+	tests := []struct {
+		name string
+		args []string
+		want string
+	}{
+		{name: "unknown command", args: []string{"unknown"}, want: "unknown command"},
+		{name: "missing config subcommand", args: []string{"config"}, want: "config requires a subcommand"},
+		{name: "unknown pipelines subcommand", args: []string{"pipelines", "unknown"}, want: "unknown pipelines subcommand"},
+		{name: "malformed run flag", args: []string{"run", "demo", "--chunk_cache", "invalid"}, want: "not supported"},
+		{name: "unknown flag", args: []string{"config", "validate", "--unknown"}, want: "flag provided but not defined"},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			var stdout, stderr bytes.Buffer
+			code := RunWithOptions(tt.args, &stdout, &stderr, commandContractOptions(t))
+			if code != 2 || !strings.Contains(stderr.String(), tt.want) || stdout.Len() != 0 {
+				t.Fatalf("code=%d stdout=%q stderr=%q", code, stdout.String(), stderr.String())
+			}
+		})
+	}
+}
+
+func TestConfigDiscoveryPrefersExplicitPathThenEnvironment(t *testing.T) {
+	explicit := writeCommandConfig(t, "explicit", "alpha")
+	environment := writeCommandConfig(t, "environment", "beta")
+	lookup := func(name string) (string, bool) {
+		if name == "NOTARIUS_CONFIG" {
+			return environment, true
+		}
+		return "", false
+	}
+
+	var stdout, stderr bytes.Buffer
+	code := RunWithOptions([]string{"pipelines", "list", "--config", explicit}, &stdout, &stderr, commandContractOptionsWithLookup(t, lookup))
+	if code != 0 || stdout.String() != "alpha\nexplicit\n" || stderr.Len() != 0 {
+		t.Fatalf("explicit config: code=%d stdout=%q stderr=%q", code, stdout.String(), stderr.String())
+	}
+
+	stdout.Reset()
+	stderr.Reset()
+	code = RunWithOptions([]string{"pipelines", "list"}, &stdout, &stderr, commandContractOptionsWithLookup(t, lookup))
+	if code != 0 || stdout.String() != "beta\nenvironment\n" || stderr.Len() != 0 {
+		t.Fatalf("environment config: code=%d stdout=%q stderr=%q", code, stdout.String(), stderr.String())
+	}
+}
+
+func TestConfigDiscoveryUsesCompiledDefaultOnlyWhenAvailable(t *testing.T) {
+	info, statErr := os.Stat(defaultConfigPath)
+	if statErr != nil && !os.IsNotExist(statErr) {
+		t.Fatalf("stat compiled default config: %v", statErr)
+	}
+	if statErr == nil && !info.Mode().IsRegular() {
+		t.Skipf("compiled default config has unexpected host state: %s", info.Mode())
+	}
+
+	path, err := discoverConfigPath("", commandContractOptions(t))
+	if statErr == nil {
+		if err != nil || path != defaultConfigPath {
+			t.Fatalf("discoverConfigPath() = %q, %v; want compiled default", path, err)
+		}
+		return
+	}
+	if err == nil || !strings.Contains(err.Error(), "config file not found") {
+		t.Fatalf("discoverConfigPath() error = %v, want documented not-found context", err)
+	}
+}
+
+func TestConfigLoadingFailuresReturnOneWithPathContext(t *testing.T) {
+	missing := filepath.Join(t.TempDir(), "missing.yml")
+	var stdout, stderr bytes.Buffer
+	code := RunWithOptions([]string{"config", "validate", "--config", missing}, &stdout, &stderr, commandContractOptions(t))
+	if code != 1 || !strings.Contains(stderr.String(), missing) || stdout.Len() != 0 {
+		t.Fatalf("missing config: code=%d stdout=%q stderr=%q", code, stdout.String(), stderr.String())
+	}
+
+	malformed := filepath.Join(t.TempDir(), "malformed.yml")
+	if err := os.WriteFile(malformed, []byte("version: [\n"), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	stdout.Reset()
+	stderr.Reset()
+	code = RunWithOptions([]string{"config", "validate", "--config", malformed}, &stdout, &stderr, commandContractOptions(t))
+	if code != 1 || !strings.Contains(stderr.String(), malformed) || !strings.Contains(stderr.String(), "parse config file") || stdout.Len() != 0 {
+		t.Fatalf("malformed config: code=%d stdout=%q stderr=%q", code, stdout.String(), stderr.String())
+	}
+}
+
+func TestConfigValidateResolvesPipelineAndChecksSelection(t *testing.T) {
+	configPath := writeResolvableCommandConfig(t)
+	options := commandContractOptions(t)
+
+	var stdout, stderr bytes.Buffer
+	code := RunWithOptions([]string{"config", "validate", "--config", configPath, "--pipeline", "demo", "--only", "spells"}, &stdout, &stderr, options)
+	if code != 0 || !strings.Contains(stdout.String(), "valid for pipeline \"demo\"") || stderr.Len() != 0 {
+		t.Fatalf("valid resolution: code=%d stdout=%q stderr=%q", code, stdout.String(), stderr.String())
+	}
+
+	stdout.Reset()
+	stderr.Reset()
+	code = RunWithOptions([]string{"config", "validate", "--config", configPath, "--pipeline", "missing"}, &stdout, &stderr, options)
+	if code != 1 || !strings.Contains(stderr.String(), "pipeline \"missing\"") {
+		t.Fatalf("unknown pipeline: code=%d stdout=%q stderr=%q", code, stdout.String(), stderr.String())
+	}
+
+	stdout.Reset()
+	stderr.Reset()
+	code = RunWithOptions([]string{"config", "validate", "--config", configPath, "--pipeline", "demo", "--only", "missing"}, &stdout, &stderr, options)
+	if code != 1 || !strings.Contains(stderr.String(), "lane \"missing\"") {
+		t.Fatalf("unknown lane: code=%d stdout=%q stderr=%q", code, stdout.String(), stderr.String())
+	}
+
+	stdout.Reset()
+	stderr.Reset()
+	code = RunWithOptions([]string{"config", "validate", "--config", configPath, "--only", "spells"}, &stdout, &stderr, options)
+	if code != 2 || !strings.Contains(stderr.String(), "--only requires --pipeline") {
+		t.Fatalf("missing pipeline for only: code=%d stdout=%q stderr=%q", code, stdout.String(), stderr.String())
+	}
+
+	stdout.Reset()
+	stderr.Reset()
+	code = RunWithOptions([]string{"config", "validate", "--config", configPath, "--pipeline", "demo", "--only", "spells,,other"}, &stdout, &stderr, options)
+	if code != 2 || !strings.Contains(stderr.String(), "--only must contain") {
+		t.Fatalf("malformed only: code=%d stdout=%q stderr=%q", code, stdout.String(), stderr.String())
+	}
+}
+
+func TestConfigValidatePipelineDefaultProfileIsOffline(t *testing.T) {
+	configPath := writeCommandConfigContent(t, `version: 4
+pipelines:
+  demo:
+    llm_profile: dnd-extraction
+    input: seriatim
+    artifacts:
+      spells:
+        extract: dnd/spells
+`)
+	var stdout, stderr bytes.Buffer
+	code := RunWithOptions([]string{"config", "validate", "--config", configPath, "--pipeline", "demo"}, &stdout, &stderr, Options{})
+	if code != 0 || !strings.Contains(stdout.String(), "valid for pipeline \"demo\"") || stderr.Len() != 0 {
+		t.Fatalf("code=%d stdout=%q stderr=%q", code, stdout.String(), stderr.String())
+	}
+}
+
+func TestPipelinesListSortsNormalizedIDsInTextAndJSON(t *testing.T) {
+	configPath := writeCommandConfig(t, " zeta ", "alpha")
+	options := commandContractOptions(t)
+	var stdout, stderr bytes.Buffer
+	code := RunWithOptions([]string{"pipelines", "list", "--config", configPath}, &stdout, &stderr, options)
+	if code != 0 || stdout.String() != "alpha\nzeta\n" || stderr.Len() != 0 {
+		t.Fatalf("text list: code=%d stdout=%q stderr=%q", code, stdout.String(), stderr.String())
+	}
+
+	stdout.Reset()
+	stderr.Reset()
+	code = RunWithOptions([]string{"pipelines", "list", "--config", configPath, "--json"}, &stdout, &stderr, options)
+	var payload struct {
+		Pipelines []string `json:"pipelines"`
+	}
+	if err := json.Unmarshal(stdout.Bytes(), &payload); err != nil {
+		t.Fatalf("JSON list = %q: %v", stdout.String(), err)
+	}
+	if code != 0 || len(payload.Pipelines) != 2 || payload.Pipelines[0] != "alpha" || payload.Pipelines[1] != "zeta" || stderr.Len() != 0 {
+		t.Fatalf("JSON list: code=%d payload=%#v stderr=%q", code, payload, stderr.String())
+	}
+}
+
+func TestRemovedStructuralFlagsAndRuntimeFailuresKeepExitClasses(t *testing.T) {
+	configPath := writeResolvableCommandConfig(t)
+	options := commandContractOptions(t)
+
+	var stdout, stderr bytes.Buffer
+	code := RunWithOptions([]string{"run", "demo", "--input", "missing-input", "--config", configPath, "--diagnostics-dir", t.TempDir()}, &stdout, &stderr, options)
+	if code != 2 || !strings.Contains(stderr.String(), "flag provided but not defined") {
+		t.Fatalf("removed flag: code=%d stdout=%q stderr=%q", code, stdout.String(), stderr.String())
+	}
+
+	stdout.Reset()
+	stderr.Reset()
+	code = RunWithOptions([]string{"run", "missing", "--input", "missing-input", "--config", configPath, "--chunk_cache", "bypass"}, &stdout, &stderr, options)
+	if code != 1 || !strings.Contains(stderr.String(), "pipeline \"missing\"") || stdout.Len() != 0 {
+		t.Fatalf("valid-runtime failure: code=%d stdout=%q stderr=%q", code, stdout.String(), stderr.String())
+	}
+}
+
+func commandContractOptions(t *testing.T) Options {
+	return commandContractOptionsWithLookup(t, emptyLookup)
+}
+
+func commandContractOptionsWithLookup(t *testing.T, lookup func(string) (string, bool)) Options {
+	t.Helper()
+	components, err := newProductionComponents()
+	if err != nil {
+		t.Fatal(err)
+	}
+	return Options{
+		Catalog:    catalogFromRegistries(components.registries),
+		Registries: components.registries,
+		LookupEnv:  lookup,
+	}
+}
+
+func writeCommandConfig(t *testing.T, firstID, secondID string) string {
+	t.Helper()
+	content := fmt.Sprintf("version: 4\npipelines:\n  %q:\n    input: seriatim\n  %q:\n    input: seriatim\n", firstID, secondID)
+	return writeCommandConfigContent(t, content)
+}
+
+func writeResolvableCommandConfig(t *testing.T) string {
+	t.Helper()
+	return writeCommandConfigContent(t, `version: 4
+pipelines:
+  demo:
+    input: seriatim
+    artifacts:
+      spells:
+        extract: dnd/spells
+`)
+}
+
+func writeCommandConfigContent(t *testing.T, content string) string {
+	t.Helper()
+	path := filepath.Join(t.TempDir(), "config.yml")
+	if err := os.WriteFile(path, []byte(content), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	return path
+}

internal/cli/contract_test_helpers_test.go

@@ -0,0 +1,16 @@
+package cli
+
+import (
+	"strings"
+	"testing"
+)
+
+func replaceRequiredOnce(t *testing.T, input, old, replacement string) string {
+	t.Helper()
+	if count := strings.Count(input, old); count != 1 {
+		t.Fatalf("replacement marker %q occurs %d times, want exactly once", old, count)
+	}
+	return strings.Replace(input, old, replacement, 1)
+}
+
+func emptyLookup(string) (string, bool) { return "", false }

internal/cli/dnd_combat_contract_test.go

@@ -0,0 +1,208 @@
+package cli
+
+import (
+	"reflect"
+	"strings"
+	"testing"
+
+	"gitea.maximumdirect.net/eric/notarius/internal/core/config"
+	"gitea.maximumdirect.net/eric/notarius/internal/framework/contracts"
+	"gitea.maximumdirect.net/eric/notarius/internal/framework/pipeline"
+	"gitea.maximumdirect.net/eric/notarius/internal/modules/dnd"
+	combatextract "gitea.maximumdirect.net/eric/notarius/internal/modules/dnd/extract/combatturns"
+	combatnormalize "gitea.maximumdirect.net/eric/notarius/internal/modules/dnd/normalize/combatturns"
+)
+
+func TestProductionCombatConfigurationResolvesTypedLane(t *testing.T) {
+	components := productionTestComponents(t)
+	cfg := productionCombatContractConfig()
+	effective, err := cfg.Resolve(config.ResolveInput{PipelineID: "dnd-combat", Catalog: catalogFromRegistries(components.registries)})
+	if err != nil {
+		t.Fatalf("Resolve() error = %v, want nil", err)
+	}
+	if effective.ResolvedPipeline.Chunk.Module != pipeline.DefaultChunkModule {
+		t.Fatalf("chunk module = %q, want %q", effective.ResolvedPipeline.Chunk.Module, pipeline.DefaultChunkModule)
+	}
+	if len(effective.ResolvedPipeline.Steps[0].ArtifactLanes) != 1 {
+		t.Fatalf("artifact lanes = %#v, want one combat lane", effective.ResolvedPipeline.Steps[0].ArtifactLanes)
+	}
+	lane := effective.ResolvedPipeline.Steps[0].ArtifactLanes[0]
+	if lane.ID != "combat" || lane.ArtifactKind != dnd.CombatTurnListKind || lane.Extract.Module != combatextract.Key || lane.Extract.Retries != 2 || lane.Merge.Module != pipeline.DefaultMergeModule || lane.Normalize.Module != combatnormalize.Key {
+		t.Fatalf("resolved combat lane = %#v, want typed production composition", lane)
+	}
+
+	catalog := catalogFromRegistries(components.registries)
+	extractSpec, ok := catalog.Extractors.Spec(combatextract.Key)
+	if !ok || !reflect.DeepEqual(extractSpec.Requires, []string{"chunks", "source.transcript"}) || !reflect.DeepEqual(extractSpec.Provides, []string{"dnd.combat_turns"}) {
+		t.Fatalf("combat extractor spec = %#v, want source and artifact capabilities", extractSpec)
+	}
+	normalizeSpec, ok := catalog.Normalizers.SpecForArtifact(combatnormalize.Key, dnd.CombatTurnListKind)
+	if !ok || !reflect.DeepEqual(normalizeSpec.Requires, []string{"merged"}) || !reflect.DeepEqual(normalizeSpec.Provides, []string{"normalized"}) {
+		t.Fatalf("combat normalizer spec = %#v, want merged/normalized capabilities", normalizeSpec)
+	}
+	mergeSpec, ok := catalog.Mergers.SpecForArtifact(pipeline.DefaultMergeModule, dnd.CombatTurnListKind)
+	if !ok || !reflect.DeepEqual(mergeSpec.Provides, []string{"merged"}) {
+		t.Fatalf("combat merger spec = %#v, want merged capability", mergeSpec)
+	}
+	codecSpec, ok := catalog.ArtifactCodecs.Spec(dnd.CombatTurnListKind)
+	if !ok || codecSpec.Schema.ID != "notarius.dnd.combat_turns" || codecSpec.Schema.Version != "v1" {
+		t.Fatalf("combat codec spec = %#v, want compatible durable schema", codecSpec)
+	}
+	if !hasReferenceSlot(extractSpec.ReferenceSlots, "npcs") || !hasReferenceSlot(extractSpec.ReferenceSlots, "scene_descriptions") || !hasReferenceSlot(normalizeSpec.ReferenceSlots, "npcs") {
+		t.Fatalf("combat reference slots = %#v / %#v, want extraction scene and NPC slots plus normalization NPC slot", extractSpec.ReferenceSlots, normalizeSpec.ReferenceSlots)
+	}
+	sceneSlot := referenceSlot(extractSpec.ReferenceSlots, "scene_descriptions")
+	if !sceneSlot.Required || !reflect.DeepEqual(sceneSlot.AcceptedMediaTypes, []string{"application/json"}) || !reflect.DeepEqual(sceneSlot.AcceptedArtifactKinds, []contracts.ArtifactKind{dnd.SceneDescriptionListKind}) || sceneSlot.MaxBytes != 1048576 {
+		t.Fatalf("scene description slot = %#v, want required approved scene artifact", sceneSlot)
+	}
+
+	wantExtractChain := []pipeline.ModuleBinding{
+		pipeline.Binding("generic/valid_json"),
+		pipeline.Binding("extract/dnd/combat-turns/shape"),
+		pipeline.Binding("extract/dnd/combat-turns/source_refs"),
+		pipeline.Binding("generic/valid_json_schema"),
+		pipeline.Binding("extract/dnd/combat-turns/source_relatedness"),
+	}
+	wantNormalizeChain := []pipeline.ModuleBinding{
+		pipeline.Binding("generic/valid_json"),
+		pipeline.Binding("extract/dnd/combat-turns/shape"),
+		pipeline.Binding("normalize/dnd/combat-turns/invariants"),
+		pipeline.Binding("extract/dnd/combat-turns/source_refs"),
+		pipeline.Binding("generic/valid_json_schema"),
+		pipeline.Binding("extract/dnd/combat-turns/source_relatedness"),
+	}
+	if got := validatorChain(effective.ResolvedPipeline, pipeline.StageExtract, combatextract.Key); !reflect.DeepEqual(got, wantExtractChain) {
+		t.Fatalf("combat extract chain = %#v, want %#v", got, wantExtractChain)
+	}
+	if got := validatorChain(effective.ResolvedPipeline, pipeline.StageNormalize, combatnormalize.Key); !reflect.DeepEqual(got, wantNormalizeChain) {
+		t.Fatalf("combat normalize chain = %#v, want %#v", got, wantNormalizeChain)
+	}
+	if got := validatorChain(effective.ResolvedPipeline, pipeline.StageMerge, pipeline.DefaultMergeModule); len(got) != 0 {
+		t.Fatalf("combat merge chain = %#v, want empty", got)
+	}
+
+	bound, err := cfg.Resolve(config.ResolveInput{
+		PipelineID: "dnd-combat",
+		Catalog:    catalog,
+		ReferenceOverrides: []pipeline.ReferenceBinding{
+			{Stage: pipeline.StageExtract, LaneID: "combat", SlotName: "npcs", Source: "npc-run/lanes/npcs.json", BindingSource: contracts.ReferenceBindingSourceCLI},
+			{Stage: pipeline.StageNormalize, LaneID: "combat", SlotName: "npcs", Source: "npc-run/lanes/npcs.json", BindingSource: contracts.ReferenceBindingSourceCLI},
+		},
+	})
+	if err != nil {
+		t.Fatalf("Resolve(bound references) error = %v, want nil", err)
+	}
+	boundLane := bound.ResolvedPipeline.Steps[0].ArtifactLanes[0]
+	if len(boundLane.ExtractReferences.Bindings) != 2 || len(boundLane.NormalizeReferences.Bindings) != 1 || !hasReferenceBinding(boundLane.ExtractReferences.Bindings, "npcs") || !hasReferenceBinding(boundLane.ExtractReferences.Bindings, "scene_descriptions") || !hasReferenceBinding(boundLane.NormalizeReferences.Bindings, "npcs") {
+		t.Fatalf("bound combat references = %#v / %#v, want extraction scene and NPC bindings plus normalization NPC binding", boundLane.ExtractReferences, boundLane.NormalizeReferences)
+	}
+}
+
+func TestProductionCombatConfigurationRequiresSceneDescriptions(t *testing.T) {
+	components := productionTestComponents(t)
+	cfg := productionCombatContractConfig()
+	profile := cfg.Pipelines["dnd-combat"]
+	profile.References = nil
+	cfg.Pipelines["dnd-combat"] = profile
+	if _, err := cfg.Resolve(config.ResolveInput{PipelineID: "dnd-combat", Catalog: catalogFromRegistries(components.registries)}); err == nil || !strings.Contains(err.Error(), "scene_descriptions") || !strings.Contains(err.Error(), "required") {
+		t.Fatalf("Resolve() error = %v, want required scene reference failure", err)
+	}
+}
+
+func TestProductionCombatConfigurationRejectsLooseOptionsAndLaneValidators(t *testing.T) {
+	components := productionTestComponents(t)
+	resolve := func(mutate func(*pipeline.PipelineProfile)) error {
+		cfg := productionCombatContractConfig()
+		profile := cfg.Pipelines["dnd-combat"]
+		mutate(&profile)
+		cfg.Pipelines["dnd-combat"] = profile
+		_, err := cfg.Resolve(config.ResolveInput{PipelineID: "dnd-combat", Catalog: catalogFromRegistries(components.registries)})
+		return err
+	}
+	if err := resolve(func(profile *pipeline.PipelineProfile) {
+		lane := profile.Artifacts["combat"]
+		lane.Extract.Options = map[string]any{"unexpected": true}
+		profile.Artifacts["combat"] = lane
+	}); err == nil || !strings.Contains(err.Error(), "unknown option") {
+		t.Fatalf("unknown extractor option error = %v, want strict option rejection", err)
+	}
+	if err := resolve(func(profile *pipeline.PipelineProfile) {
+		lane := profile.Artifacts["combat"]
+		lane.Normalize.Options = map[string]any{"unexpected": true}
+		profile.Artifacts["combat"] = lane
+	}); err == nil || !strings.Contains(err.Error(), "unknown option") {
+		t.Fatalf("unknown normalizer option error = %v, want strict option rejection", err)
+	}
+	if err := resolve(func(profile *pipeline.PipelineProfile) {
+		lane := profile.Artifacts["combat"]
+		lane.Validators = []pipeline.ModuleBinding{pipeline.Binding("generic/always_accept")}
+		profile.Artifacts["combat"] = lane
+	}); err == nil || !strings.Contains(err.Error(), "artifact lane level") {
+		t.Fatalf("lane-level validator error = %v, want invalid placement rejection", err)
+	}
+}
+
+func TestProductionCombatConfigurationResolvesTypedUnconditionalValidators(t *testing.T) {
+	components := productionTestComponents(t)
+	cfg := productionCombatContractConfig()
+	profile := cfg.Pipelines["dnd-combat"]
+	lane := profile.Artifacts["combat"]
+	lane.Extract.Validators = pipeline.ValidatorOverride{Set: true, Validators: []pipeline.ModuleBinding{pipeline.Binding("generic/always_accept")}}
+	lane.Normalize.Validators = pipeline.ValidatorOverride{Set: true, Validators: []pipeline.ModuleBinding{pipeline.Binding("generic/always_reject")}}
+	profile.Artifacts["combat"] = lane
+	cfg.Pipelines["dnd-combat"] = profile
+	effective, err := cfg.Resolve(config.ResolveInput{PipelineID: "dnd-combat", Catalog: catalogFromRegistries(components.registries)})
+	if err != nil {
+		t.Fatalf("Resolve() error = %v, want typed unconditional validators to resolve", err)
+	}
+	if got := validatorChain(effective.ResolvedPipeline, pipeline.StageExtract, combatextract.Key); !reflect.DeepEqual(got, []pipeline.ModuleBinding{pipeline.Binding("generic/always_accept")}) {
+		t.Fatalf("extract override chain = %#v, want typed always-accept", got)
+	}
+	if got := validatorChain(effective.ResolvedPipeline, pipeline.StageNormalize, combatnormalize.Key); !reflect.DeepEqual(got, []pipeline.ModuleBinding{pipeline.Binding("generic/always_reject")}) {
+		t.Fatalf("normalize override chain = %#v, want typed always-reject", got)
+	}
+}
+
+func productionCombatContractConfig() config.Config {
+	cfg := config.Default()
+	cfg.Pipelines["dnd-combat"] = pipeline.PipelineProfile{
+		ID:         "dnd-combat",
+		Input:      pipeline.Binding("seriatim"),
+		Chunk:      pipeline.Binding(pipeline.DefaultChunkModule),
+		References: map[string]pipeline.ReferenceSource{"scene_descriptions": pipeline.ExternalReference("scenes.json")},
+		Artifacts: map[string]pipeline.ArtifactLaneProfile{
+			"combat": {
+				Extract:   pipeline.ModuleBinding{Module: combatextract.Key, Retries: 2},
+				Normalize: pipeline.Binding(combatnormalize.Key),
+			},
+		},
+	}
+	return cfg
+}
+
+func hasReferenceSlot(slots []contracts.ReferenceSlot, name string) bool {
+	for _, slot := range slots {
+		if slot.Name == name {
+			return true
+		}
+	}
+	return false
+}
+
+func hasReferenceBinding(bindings []pipeline.ReferenceBinding, name string) bool {
+	for _, binding := range bindings {
+		if binding.SlotName == name {
+			return true
+		}
+	}
+	return false
+}
+
+func referenceSlot(slots []contracts.ReferenceSlot, name string) contracts.ReferenceSlot {
+	for _, slot := range slots {
+		if slot.Name == name {
+			return slot
+		}
+	}
+	return contracts.ReferenceSlot{}
+}

internal/cli/dnd_interactions_contract_test.go

@@ -0,0 +1,135 @@
+package cli
+
+import (
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+
+	"gitea.maximumdirect.net/eric/notarius/internal/framework/pipeline"
+	"gitea.maximumdirect.net/eric/notarius/internal/modules/dnd"
+	interactioncodec "gitea.maximumdirect.net/eric/notarius/internal/modules/dnd/codec/npcinteractions"
+	interactionextract "gitea.maximumdirect.net/eric/notarius/internal/modules/dnd/extract/npcinteractions"
+	npcextract "gitea.maximumdirect.net/eric/notarius/internal/modules/dnd/extract/npcs"
+	interactionnormalize "gitea.maximumdirect.net/eric/notarius/internal/modules/dnd/normalize/npcinteractions"
+	npcnormalize "gitea.maximumdirect.net/eric/notarius/internal/modules/dnd/normalize/npcs"
+)
+
+func TestProductionNPCInteractionPipelineResolvesAndPrepares(t *testing.T) {
+	components := productionTestComponents(t)
+	resolved, err := pipeline.ResolvePipeline(npcInteractionProfile(pipeline.GeneratedReference("npcs", "npcs")), pipeline.ResolveOptions{}, catalogFromRegistries(components.registries))
+	if err != nil {
+		t.Fatalf("ResolvePipeline() error = %v", err)
+	}
+	if len(resolved.Steps) != 2 || len(resolved.Steps[1].ArtifactLanes) != 1 {
+		t.Fatalf("resolved pipeline = %#v", resolved)
+	}
+	lane := resolved.Steps[1].ArtifactLanes[0]
+	if lane.ArtifactKind != dnd.NPCInteractionListKind || lane.Extract.Module != interactionextract.Key || lane.Normalize.Module != interactionnormalize.Key {
+		t.Fatalf("interaction lane = %#v", lane)
+	}
+	for _, bindings := range [][]pipeline.ReferenceBinding{lane.ExtractReferences.Bindings, lane.NormalizeReferences.Bindings} {
+		if len(bindings) != 1 || bindings[0].SlotName != "npcs" || bindings[0].Artifact == nil || bindings[0].Artifact.Step != "npcs" || bindings[0].Artifact.Lane != "npcs" {
+			t.Fatalf("generated bindings = %#v", bindings)
+		}
+	}
+	if _, err := pipeline.Prepare(resolved, components.registries, pipeline.ModuleDependencies{LLM: &productionFakeLLMClient{}}); err != nil {
+		t.Fatalf("Prepare() error = %v", err)
+	}
+
+	catalog := catalogFromRegistries(components.registries)
+	codecSpec, ok := catalog.ArtifactCodecs.Spec(dnd.NPCInteractionListKind)
+	if !ok || codecSpec.Schema.ID != interactioncodec.SchemaID || codecSpec.Schema.Version != interactioncodec.SchemaVersion {
+		t.Fatalf("NPC interaction codec spec = %#v", codecSpec)
+	}
+}
+
+func TestProductionNPCInteractionReferencesRequireEarlierCompatibleProducer(t *testing.T) {
+	components := productionTestComponents(t)
+	catalog := catalogFromRegistries(components.registries)
+	laterProfile := npcInteractionProfile(pipeline.GeneratedReference("npcs", "npcs"))
+	laterProfile.Steps[0].ID = "seed"
+	laterProfile.Steps[0].Artifacts["seed"] = laterProfile.Steps[0].Artifacts["npcs"]
+	delete(laterProfile.Steps[0].Artifacts, "npcs")
+	laterProfile.Steps = append(laterProfile.Steps, pipeline.PipelineStepProfile{ID: "future", Artifacts: map[string]pipeline.ArtifactLaneProfile{
+		"npcs": {Extract: pipeline.Binding(npcextract.Key), Normalize: pipeline.Binding(npcnormalize.Key)},
+	}})
+	laterProfile.Steps[1].References["npcs"] = pipeline.GeneratedReference("future", "npcs")
+	tests := []struct {
+		name    string
+		profile pipeline.PipelineProfile
+		want    string
+	}{
+		{name: "missing", profile: npcInteractionProfile(pipeline.ReferenceSource{}), want: "source must not be empty"},
+		{name: "same step", profile: npcInteractionProfile(pipeline.GeneratedReference("interactions", "interactions")), want: "earlier step"},
+		{name: "later step", profile: laterProfile, want: "earlier step"},
+		{name: "wrong artifact kind", profile: npcInteractionProfile(pipeline.GeneratedReference("npcs", "npcs")), want: "does not accept artifact kind"},
+	}
+	tests[3].profile.Steps[0].Artifacts["npcs"] = pipeline.ArtifactLaneProfile{Extract: pipeline.Binding("dnd/spells")}
+	for _, test := range tests {
+		t.Run(test.name, func(t *testing.T) {
+			_, err := pipeline.ResolvePipeline(test.profile, pipeline.ResolveOptions{}, catalog)
+			if err == nil || !strings.Contains(err.Error(), test.want) {
+				t.Fatalf("ResolvePipeline() error = %v, want %q", err, test.want)
+			}
+		})
+	}
+}
+
+func TestProductionNPCInteractionReferencesRejectIncompatibleExternalRegistries(t *testing.T) {
+	components := productionTestComponents(t)
+	catalog := catalogFromRegistries(components.registries)
+	root := t.TempDir()
+	for _, test := range []struct {
+		name    string
+		file    string
+		content string
+		prepare bool
+		want    string
+	}{
+		{name: "media type", file: "registry.txt", content: "not JSON", want: "media type"},
+		{name: "artifact schema", file: "registry.json", content: `{"npcs":[{"name":"missing required fields"}]}`, prepare: true, want: "NPC registry"},
+	} {
+		t.Run(test.name, func(t *testing.T) {
+			path := filepath.Join(root, test.file)
+			if err := os.WriteFile(path, []byte(test.content), 0o600); err != nil {
+				t.Fatal(err)
+			}
+			resolved, err := pipeline.ResolvePipeline(npcInteractionProfile(pipeline.ExternalReference(path)), pipeline.ResolveOptions{}, catalog)
+			if err != nil {
+				t.Fatalf("ResolvePipeline() error = %v", err)
+			}
+			materialized, _, err := pipeline.MaterializeReferences(resolved, catalog, pipeline.ReferenceMaterializationOptions{})
+			if !test.prepare {
+				if err == nil || !strings.Contains(err.Error(), test.want) {
+					t.Fatalf("MaterializeReferences() error = %v, want %q", err, test.want)
+				}
+				return
+			}
+			if err != nil {
+				t.Fatalf("MaterializeReferences() error = %v", err)
+			}
+			if _, err := pipeline.Prepare(materialized, components.registries, pipeline.ModuleDependencies{LLM: &productionFakeLLMClient{}}); err == nil || !strings.Contains(err.Error(), test.want) {
+				t.Fatalf("Prepare() error = %v, want %q", err, test.want)
+			}
+		})
+	}
+}
+
+func npcInteractionProfile(reference pipeline.ReferenceSource) pipeline.PipelineProfile {
+	profile := pipeline.PipelineProfile{
+		ID:     "dnd-npc-interactions",
+		Input:  pipeline.Binding("seriatim"),
+		Chunk:  pipeline.ModuleBinding{Module: "generic", Options: map[string]any{"max_units": 1}},
+		Output: pipeline.Binding("json"),
+		Steps: []pipeline.PipelineStepProfile{
+			{ID: "npcs", Artifacts: map[string]pipeline.ArtifactLaneProfile{
+				"npcs": {Extract: pipeline.Binding(npcextract.Key), Normalize: pipeline.Binding(npcnormalize.Key)},
+			}},
+			{ID: "interactions", References: map[string]pipeline.ReferenceSource{"npcs": reference}, Artifacts: map[string]pipeline.ArtifactLaneProfile{
+				"interactions": {Extract: pipeline.Binding(interactionextract.Key), Normalize: pipeline.Binding(interactionnormalize.Key)},
+			}},
+		},
+	}
+	return profile
+}

internal/cli/dnd_npc_contract_test.go

@@ -0,0 +1,151 @@
+package cli
+
+import (
+	"reflect"
+	"strings"
+	"testing"
+
+	"gitea.maximumdirect.net/eric/notarius/internal/core/config"
+	"gitea.maximumdirect.net/eric/notarius/internal/framework/pipeline"
+	"gitea.maximumdirect.net/eric/notarius/internal/modules/dnd"
+	npccodec "gitea.maximumdirect.net/eric/notarius/internal/modules/dnd/codec/npcs"
+	npcextract "gitea.maximumdirect.net/eric/notarius/internal/modules/dnd/extract/npcs"
+	npcnormalize "gitea.maximumdirect.net/eric/notarius/internal/modules/dnd/normalize/npcs"
+)
+
+func TestProductionNPCConfigurationResolvesTypedLane(t *testing.T) {
+	components := productionTestComponents(t)
+	catalog := catalogFromRegistries(components.registries)
+	cfg := productionNPCContractConfig()
+	effective, err := cfg.Resolve(config.ResolveInput{PipelineID: "dnd-session", Catalog: catalog})
+	if err != nil {
+		t.Fatalf("Resolve() error = %v, want nil", err)
+	}
+	if effective.ResolvedPipeline.Chunk.Module != pipeline.DefaultChunkModule {
+		t.Fatalf("chunk module = %q, want %q", effective.ResolvedPipeline.Chunk.Module, pipeline.DefaultChunkModule)
+	}
+	if len(effective.ResolvedPipeline.Steps[0].ArtifactLanes) != 1 {
+		t.Fatalf("artifact lanes = %#v, want one NPC lane", effective.ResolvedPipeline.Steps[0].ArtifactLanes)
+	}
+	lane := effective.ResolvedPipeline.Steps[0].ArtifactLanes[0]
+	if lane.ID != "npcs" || lane.ArtifactKind != dnd.NPCListKind || lane.Extract.Module != npcextract.Key || lane.Extract.Retries != 2 || lane.Merge.Module != pipeline.DefaultMergeModule || lane.Normalize.Module != npcnormalize.Key {
+		t.Fatalf("resolved NPC lane = %#v, want typed production composition", lane)
+	}
+	if len(lane.ExtractReferences.Bindings) != 0 || len(lane.NormalizeReferences.Bindings) != 0 {
+		t.Fatalf("unbound NPC references = %#v / %#v, want none", lane.ExtractReferences, lane.NormalizeReferences)
+	}
+
+	extractSpec, ok := catalog.Extractors.Spec(npcextract.Key)
+	if !ok || !reflect.DeepEqual(extractSpec.Requires, []string{"chunks", "source.transcript"}) || !reflect.DeepEqual(extractSpec.Provides, []string{"dnd.npcs"}) {
+		t.Fatalf("NPC extractor spec = %#v, want source and artifact capabilities", extractSpec)
+	}
+	mergeSpec, ok := catalog.Mergers.SpecForArtifact(pipeline.DefaultMergeModule, dnd.NPCListKind)
+	if !ok || !reflect.DeepEqual(mergeSpec.Provides, []string{"merged"}) {
+		t.Fatalf("NPC merger spec = %#v, want merged capability", mergeSpec)
+	}
+	normalizeSpec, ok := catalog.Normalizers.SpecForArtifact(npcnormalize.Key, dnd.NPCListKind)
+	if !ok || !reflect.DeepEqual(normalizeSpec.Requires, []string{"merged"}) || !reflect.DeepEqual(normalizeSpec.Provides, []string{"normalized"}) {
+		t.Fatalf("NPC normalizer spec = %#v, want merged/normalized capabilities", normalizeSpec)
+	}
+	codecSpec, ok := catalog.ArtifactCodecs.Spec(dnd.NPCListKind)
+	if !ok || codecSpec.Kind != dnd.NPCListKind || codecSpec.Schema.ID != npccodec.SchemaID || codecSpec.Schema.Version != npccodec.SchemaVersion {
+		t.Fatalf("NPC codec spec = %#v, want typed v1 durable schema", codecSpec)
+	}
+
+	wantExtractChain := []pipeline.ModuleBinding{
+		pipeline.Binding("generic/valid_json"),
+		pipeline.Binding("extract/dnd/npcs/shape"),
+		pipeline.Binding("extract/dnd/npcs/source_refs"),
+		pipeline.Binding("generic/valid_json_schema"),
+		pipeline.Binding("extract/dnd/npcs/source_relatedness"),
+	}
+	wantNormalizeChain := []pipeline.ModuleBinding{
+		pipeline.Binding("generic/valid_json"),
+		pipeline.Binding("extract/dnd/npcs/shape"),
+		pipeline.Binding("normalize/dnd/npcs/identity"),
+		pipeline.Binding("extract/dnd/npcs/source_refs"),
+		pipeline.Binding("generic/valid_json_schema"),
+		pipeline.Binding("extract/dnd/npcs/source_relatedness"),
+	}
+	if got := validatorChain(effective.ResolvedPipeline, pipeline.StageExtract, npcextract.Key); !reflect.DeepEqual(got, wantExtractChain) {
+		t.Fatalf("NPC extract chain = %#v, want %#v", got, wantExtractChain)
+	}
+	if got := validatorChain(effective.ResolvedPipeline, pipeline.StageNormalize, npcnormalize.Key); !reflect.DeepEqual(got, wantNormalizeChain) {
+		t.Fatalf("NPC normalize chain = %#v, want %#v", got, wantNormalizeChain)
+	}
+	if got := validatorChain(effective.ResolvedPipeline, pipeline.StageMerge, pipeline.DefaultMergeModule); len(got) != 0 {
+		t.Fatalf("NPC merge chain = %#v, want empty", got)
+	}
+}
+
+func TestProductionNPCConfigurationValidatesOptionsReferencesAndPlacement(t *testing.T) {
+	components := productionTestComponents(t)
+	resolve := func(mutate func(*pipeline.PipelineProfile)) error {
+		cfg := productionNPCContractConfig()
+		profile := cfg.Pipelines["dnd-session"]
+		mutate(&profile)
+		cfg.Pipelines["dnd-session"] = profile
+		_, err := cfg.Resolve(config.ResolveInput{PipelineID: "dnd-session", Catalog: catalogFromRegistries(components.registries)})
+		return err
+	}
+
+	if err := resolve(func(profile *pipeline.PipelineProfile) {
+		lane := profile.Artifacts["npcs"]
+		lane.Extract.Options = map[string]any{"unexpected": true}
+		profile.Artifacts["npcs"] = lane
+	}); err == nil || !strings.Contains(err.Error(), "unknown option") {
+		t.Fatalf("unknown extractor option error = %v, want strict option rejection", err)
+	}
+	if err := resolve(func(profile *pipeline.PipelineProfile) {
+		lane := profile.Artifacts["npcs"]
+		lane.Normalize.Options = map[string]any{"unexpected": true}
+		profile.Artifacts["npcs"] = lane
+	}); err == nil || !strings.Contains(err.Error(), "unknown option") {
+		t.Fatalf("unknown normalizer option error = %v, want strict option rejection", err)
+	}
+	if err := resolve(func(profile *pipeline.PipelineProfile) {
+		profile.References = pipeline.ExternalReferenceMap(map[string]string{
+			"players":  "players.txt",
+			"party":    "party.txt",
+			"glossary": "glossary.txt",
+		})
+	}); err != nil {
+		t.Fatalf("optional NPC references error = %v, want resolution success", err)
+	}
+	if err := resolve(func(profile *pipeline.PipelineProfile) {
+		lane := profile.Artifacts["npcs"]
+		lane.Validators = []pipeline.ModuleBinding{pipeline.Binding("normalize/dnd/npcs/identity")}
+		profile.Artifacts["npcs"] = lane
+	}); err == nil || !strings.Contains(err.Error(), "artifact lane level") {
+		t.Fatalf("lane-level validator error = %v, want invalid placement rejection", err)
+	}
+}
+
+func productionNPCContractConfig() config.Config {
+	cfg := config.Default()
+	cfg.Pipelines["dnd-session"] = pipeline.PipelineProfile{
+		ID:    "dnd-session",
+		Input: pipeline.Binding("seriatim"),
+		Chunk: pipeline.Binding(pipeline.DefaultChunkModule),
+		Artifacts: map[string]pipeline.ArtifactLaneProfile{
+			"npcs": {
+				Extract:   pipeline.ModuleBinding{Module: npcextract.Key, Retries: 2},
+				Normalize: pipeline.Binding(npcnormalize.Key),
+			},
+		},
+	}
+	return cfg
+}
+
+func validatorChain(resolved pipeline.ResolvedPipeline, stage pipeline.ModuleStage, module string) []pipeline.ModuleBinding {
+	for _, chain := range resolved.ValidatorChains {
+		if chain.Stage == stage && chain.ModuleKey == module {
+			bindings := make([]pipeline.ModuleBinding, len(chain.Validators))
+			for index, validator := range chain.Validators {
+				bindings[index] = validator.Binding
+			}
+			return bindings
+		}
+	}
+	return nil
+}

internal/cli/dnd_scene_descriptions_contract_test.go

@@ -0,0 +1,147 @@
+package cli
+
+import (
+	"context"
+	"encoding/json"
+	"fmt"
+	"reflect"
+	"strings"
+	"sync"
+	"testing"
+
+	"gitea.maximumdirect.net/eric/notarius/internal/core/artifacts"
+	"gitea.maximumdirect.net/eric/notarius/internal/core/config"
+	"gitea.maximumdirect.net/eric/notarius/internal/core/source"
+	"gitea.maximumdirect.net/eric/notarius/internal/framework/contracts"
+	"gitea.maximumdirect.net/eric/notarius/internal/framework/pipeline"
+	"gitea.maximumdirect.net/eric/notarius/internal/modules/dnd"
+	scenecodec "gitea.maximumdirect.net/eric/notarius/internal/modules/dnd/codec/scenedescriptions"
+	sceneextract "gitea.maximumdirect.net/eric/notarius/internal/modules/dnd/extract/scenedescriptions"
+	scenenormalize "gitea.maximumdirect.net/eric/notarius/internal/modules/dnd/normalize/scenedescriptions"
+)
+
+func TestProductionSceneDescriptionWorkflow(t *testing.T) {
+	components := productionTestComponents(t)
+	cfg := config.Default()
+	cfg.Pipelines["scene-descriptions"] = pipeline.PipelineProfile{
+		ID:     "scene-descriptions",
+		Input:  pipeline.Binding("seriatim"),
+		Chunk:  pipeline.ModuleBinding{Module: "generic", Options: map[string]any{"max_units": 1}},
+		Output: pipeline.Binding("json"),
+		Artifacts: map[string]pipeline.ArtifactLaneProfile{
+			"scene-descriptions": {
+				Extract:   pipeline.ModuleBinding{Module: sceneextract.Key, LLMProfile: "scene-description-profile"},
+				Normalize: pipeline.Binding(scenenormalize.Key),
+			},
+		},
+	}
+	effective, err := cfg.Resolve(config.ResolveInput{PipelineID: "scene-descriptions", Catalog: catalogFromRegistries(components.registries)})
+	if err != nil {
+		t.Fatalf("Resolve() error = %v", err)
+	}
+	lane := effective.ResolvedPipeline.Steps[0].ArtifactLanes[0]
+	if lane.ArtifactKind != dnd.SceneDescriptionListKind || lane.Extract.Module != sceneextract.Key || lane.Merge.Module != pipeline.DefaultMergeModule || lane.Normalize.Module != scenenormalize.Key {
+		t.Fatalf("resolved lane = %#v, want production scene-description composition", lane)
+	}
+	if len(lane.ExtractReferences.Bindings) != 0 || len(lane.NormalizeReferences.Bindings) != 0 {
+		t.Fatalf("resolved references = %#v / %#v, want no generated or required references", lane.ExtractReferences, lane.NormalizeReferences)
+	}
+
+	llmClient := &sceneDescriptionLLM{}
+	prepared, err := pipeline.Prepare(effective.ResolvedPipeline, components.registries, pipeline.ModuleDependencies{LLM: llmClient})
+	if err != nil {
+		t.Fatalf("Prepare() error = %v", err)
+	}
+	output, err := pipeline.New().Run(context.Background(), pipeline.RunInput{
+		Prepared:       prepared,
+		RawInput:       readRepositoryFile(t, "examples", "seriatim-minimal-transcript.json"),
+		ChunkCacheMode: pipeline.ChunkCacheBypass,
+	})
+	if err != nil {
+		t.Fatalf("Run() error = %v", err)
+	}
+	if output.Manifest.ValidationStatus != "approved" || len(output.Rejected) != 0 || len(output.NormalizeOutputs) != 1 {
+		t.Fatalf("run output = %#v, want one approved normalized artifact", output)
+	}
+	wantProfiles := []artifacts.LLMProfileManifest{{
+		ID:       "scene-description-profile",
+		Provider: "promptkit",
+		Model:    "deterministic",
+	}}
+	if !reflect.DeepEqual(output.Manifest.LLMProfiles, wantProfiles) {
+		t.Fatalf("manifest LLM profiles = %#v, want %#v", output.Manifest.LLMProfiles, wantProfiles)
+	}
+	normalizedOutput := output.NormalizeOutputs[0]
+	if normalizedOutput.NormalizerKey != scenenormalize.Key || normalizedOutput.Artifact.Kind != dnd.SceneDescriptionListKind || normalizedOutput.Artifact.Schema.ID != scenecodec.SchemaID || normalizedOutput.Artifact.Schema.Name != scenecodec.SchemaName || normalizedOutput.Artifact.Schema.Version != scenecodec.SchemaVersion {
+		t.Fatalf("normalized output = %#v, want registered durable scene-description schema", normalizedOutput)
+	}
+
+	var value dnd.SceneDescriptionList
+	if err := json.Unmarshal(normalizedOutput.Artifact.Content, &value); err != nil {
+		t.Fatalf("decode normalized artifact: %v", err)
+	}
+	want := dnd.SceneDescriptionList{Scenes: []dnd.SceneDescription{
+		{ID: "chunk-000001", SourceRef: source.SourceRef{SourceID: "session-alpha", StartUnitID: 1, EndUnitID: 1}, Kind: dnd.SceneKindNarrative, Title: "Aria casts Cure Wounds", Summary: "Aria casts Cure Wounds."},
+		{ID: "chunk-000002", SourceRef: source.SourceRef{SourceID: "session-alpha", StartUnitID: 2, EndUnitID: 2}, Kind: dnd.SceneKindCombat, Title: "Bandit mage casts Shield", Summary: "The bandit mage casts Shield."},
+	}}
+	if !reflect.DeepEqual(value, want) {
+		t.Fatalf("normalized scene descriptions = %#v, want %#v", value, want)
+	}
+	durable := decodeAssembledOutput[dnd.SceneDescriptionList](t, output.OutputFiles, "lanes/scene-descriptions.json")
+	if !reflect.DeepEqual(durable, want) {
+		t.Fatalf("durable output payload = %#v, want %#v", durable, want)
+	}
+	if len(output.Warnings) != 0 {
+		t.Fatalf("warnings = %#v, want grounded descriptions without warnings", output.Warnings)
+	}
+}
+
+type sceneDescriptionLLM struct {
+	mu      sync.Mutex
+	profile *artifacts.LLMProfileManifest
+}
+
+func (client *sceneDescriptionLLM) CompleteStructured(ctx context.Context, req contracts.StructuredCompletionRequest, out any) (contracts.StructuredCompletionResponse, error) {
+	if err := ctx.Err(); err != nil {
+		return contracts.StructuredCompletionResponse{}, err
+	}
+	if req.PromptID != sceneextract.PromptID {
+		return contracts.StructuredCompletionResponse{}, fmt.Errorf("unexpected prompt %q", req.PromptID)
+	}
+	transcript := string(req.Inputs["transcript"].Content)
+	var content string
+	switch {
+	case strings.Contains(transcript, "Cure Wounds"):
+		content = `{"kind":"narrative","title":"  Aria casts Cure Wounds  ","summary":"  Aria casts Cure Wounds.  "}`
+	case strings.Contains(transcript, "Shield"):
+		content = `{"kind":"combat","title":"Bandit mage casts Shield","summary":"The bandit mage casts Shield."}`
+	default:
+		return contracts.StructuredCompletionResponse{}, fmt.Errorf("unexpected transcript material %q", transcript)
+	}
+	if err := json.Unmarshal([]byte(content), out); err != nil {
+		return contracts.StructuredCompletionResponse{}, fmt.Errorf("populate structured response: %w", err)
+	}
+	profile := artifacts.LLMProfileManifest{
+		ID:       req.ProfileID,
+		Provider: "promptkit",
+		Model:    "deterministic",
+	}
+	client.mu.Lock()
+	client.profile = &profile
+	client.mu.Unlock()
+	return contracts.StructuredCompletionResponse{
+		Content:   []byte(content),
+		Provider:  profile.Provider,
+		Model:     profile.Model,
+		ProfileID: profile.ID,
+	}, nil
+}
+
+func (client *sceneDescriptionLLM) LLMProfileManifests() []artifacts.LLMProfileManifest {
+	client.mu.Lock()
+	defer client.mu.Unlock()
+	if client.profile == nil {
+		return nil
+	}
+	return []artifacts.LLMProfileManifest{*client.profile}
+}

internal/cli/example_contract_test.go

@@ -0,0 +1,266 @@
+package cli
+
+import (
+	"context"
+	"encoding/json"
+	"os"
+	"path/filepath"
+	"sort"
+	"strings"
+	"testing"
+
+	"gitea.maximumdirect.net/eric/notarius/internal/core/artifacts"
+	"gitea.maximumdirect.net/eric/notarius/internal/core/config"
+	"gitea.maximumdirect.net/eric/notarius/internal/core/debugbundle"
+	"gitea.maximumdirect.net/eric/notarius/internal/framework/contracts"
+	"gitea.maximumdirect.net/eric/notarius/internal/framework/pipeline"
+	"gitea.maximumdirect.net/eric/notarius/internal/modules/dnd"
+	spellnormalize "gitea.maximumdirect.net/eric/notarius/internal/modules/dnd/normalize/spells"
+	"gitea.maximumdirect.net/eric/notarius/internal/modules/seriatim/input/transcript"
+)
+
+func TestMaintainedExamplesLoadResolveAndList(t *testing.T) {
+	components := productionTestComponents(t)
+	for _, example := range maintainedExampleFiles(t) {
+		t.Run(example.name, func(t *testing.T) {
+			cfg := loadMaintainedExample(t, example.path)
+			raw, err := os.ReadFile(example.transcriptPath)
+			if err != nil {
+				t.Fatalf("read maintained transcript %q: %v", example.transcriptPath, err)
+			}
+			document, err := transcript.New().Parse(context.Background(), contracts.ParseRequest{
+				Path: example.transcriptPath,
+				Raw:  raw,
+			})
+			if err != nil {
+				t.Fatalf("parse maintained transcript %q: %v", example.transcriptPath, err)
+			}
+			if len(document.Units) == 0 {
+				t.Fatalf("maintained transcript %q has no parsed units", example.transcriptPath)
+			}
+			for _, pipelineID := range example.pipelineIDs {
+				effective, err := cfg.Resolve(resolveInputForMaintainedExample(components, pipelineID))
+				if err != nil {
+					t.Fatalf("resolve maintained example %q: %v", pipelineID, err)
+				}
+				materialized, _, err := pipeline.MaterializeReferences(effective.ResolvedPipeline, catalogFromRegistries(components.registries), pipeline.ReferenceMaterializationOptions{
+					ConfigPath: example.path,
+					WorkingDir: filepath.Dir(example.path),
+				})
+				if err != nil {
+					t.Fatalf("materialize maintained example references for %q: %v", pipelineID, err)
+				}
+				if example.name == "complete" {
+					if got := exampleStepLaneIDs(materialized); strings.Join(got, "|") != "describe-session:item-events,npcs,scene-descriptions|extract-events:combat-turns,npc-interactions,spells" {
+						t.Fatalf("complete example steps and lanes = %v, want every D&D extractor in the documented two-step composition", got)
+					}
+					spellLane := referenceContractLane(t, materialized, "spells")
+					if len(spellLane.ExtractReferences.ReferenceSet.Slots["spell_catalog"].Items) != 1 ||
+						len(spellLane.NormalizeReferences.ReferenceSet.Slots["spell_catalog"].Items) != 1 {
+						t.Fatalf("complete example spell catalog reference was not materialized: %#v", spellLane)
+					}
+					itemEventLane := referenceContractLane(t, materialized, "item-events")
+					for _, references := range []pipeline.ResolvedReferenceTarget{itemEventLane.ExtractReferences, itemEventLane.NormalizeReferences} {
+						if _, found := references.ReferenceSet.Slots["npcs"]; found {
+							t.Fatalf("item event lane unexpectedly depends on generated NPCs: %#v", itemEventLane)
+						}
+						if _, found := references.ReferenceSet.Slots["scene_descriptions"]; found {
+							t.Fatalf("item event lane unexpectedly depends on generated scene descriptions: %#v", itemEventLane)
+						}
+					}
+				}
+			}
+			var stdout, stderr strings.Builder
+			code := RunWithOptions([]string{"pipelines", "list", "--config", example.path}, &stdout, &stderr, productionOptionsFromComponents(components))
+			if code != 0 || stdout.String() != strings.Join(example.pipelineIDs, "\n")+"\n" || stderr.Len() != 0 {
+				t.Fatalf("pipelines list: code=%d stdout=%q stderr=%q", code, stdout.String(), stderr.String())
+			}
+		})
+	}
+}
+
+func TestMaintainedConfigurationExampleSet(t *testing.T) {
+	entries, err := os.ReadDir(repositoryPath("examples"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	var names []string
+	for _, entry := range entries {
+		if !entry.IsDir() && strings.HasSuffix(entry.Name(), ".config.yml") {
+			names = append(names, entry.Name())
+		}
+	}
+	sort.Strings(names)
+	if got := strings.Join(names, ","); got != "dnd-complete.config.yml,dnd-minimal.config.yml" {
+		t.Fatalf("maintained configuration examples = %q, want only the minimal and complete D&D examples", got)
+	}
+
+	profileEntries, err := os.ReadDir(repositoryPath("examples", "profiles"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	names = names[:0]
+	for _, entry := range profileEntries {
+		if !entry.IsDir() && strings.HasSuffix(entry.Name(), ".yml") {
+			names = append(names, entry.Name())
+		}
+	}
+	sort.Strings(names)
+	if got := strings.Join(names, ","); got != "dnd-extraction.yml" {
+		t.Fatalf("maintained operator profiles = %q, want dnd-extraction.yml", got)
+	}
+}
+
+func TestMaintainedExamplesValidateEffectiveProfilesOffline(t *testing.T) {
+	t.Chdir(repositoryPath())
+	t.Setenv("OPENROUTER_API_KEY", "")
+	for _, example := range maintainedExampleFiles(t) {
+		t.Run(example.name, func(t *testing.T) {
+			var stdout, stderr strings.Builder
+			code := RunWithOptions([]string{
+				"config", "validate", "--config", example.path, "--pipeline", "dnd-session",
+			}, &stdout, &stderr, Options{})
+			if code != 0 || stderr.Len() != 0 || !strings.Contains(stdout.String(), `valid for pipeline "dnd-session"`) {
+				t.Fatalf("code=%d stdout=%q stderr=%q", code, stdout.String(), stderr.String())
+			}
+		})
+	}
+}
+
+func exampleStepLaneIDs(resolved pipeline.ResolvedPipeline) []string {
+	result := make([]string, 0, len(resolved.Steps))
+	for _, step := range resolved.Steps {
+		laneIDs := make([]string, 0, len(step.ArtifactLanes))
+		for _, lane := range step.ArtifactLanes {
+			laneIDs = append(laneIDs, lane.ID)
+		}
+		sort.Strings(laneIDs)
+		result = append(result, step.ID+":"+strings.Join(laneIDs, ","))
+	}
+	return result
+}
+
+func TestMaintainedMinimalInvocationProducesJSONBundle(t *testing.T) {
+	outputRoot := filepath.Join(t.TempDir(), "output")
+	fake := &productionFakeLLMClient{}
+	options := productionRunOptions(t, fake)
+	var stdout, stderr strings.Builder
+	code := RunWithOptions([]string{
+		"run", "dnd-session",
+		"--config", repositoryPath("examples", "dnd-minimal.config.yml"),
+		"--input", repositoryPath("examples", "seriatim-minimal-transcript.json"),
+		"--only", "spells", "--chunk_cache", "bypass", "--output-dir", outputRoot,
+	}, &stdout, &stderr, options)
+	if code != 0 {
+		t.Fatalf("code=%d stdout=%q stderr=%q", code, stdout.String(), stderr.String())
+	}
+	if !strings.Contains(stdout.String(), `pipeline "dnd-session"`) || !strings.Contains(stdout.String(), "outputs=1 rejected=0") {
+		t.Fatalf("stdout=%q, want completed pipeline and counts", stdout.String())
+	}
+
+	runRoot := filepath.Join(outputRoot, productionRunID)
+	index := readProductionJSON[exampleOutputIndex](t, filepath.Join(runRoot, "index.json"))
+	if index.ManifestFile != "manifest.json" || index.RejectedFile != "rejected.json" || index.WarningsFile != "warnings.json" || len(index.OutputFiles) != 1 {
+		t.Fatalf("index = %#v, want one spells output and fixed companion files", index)
+	}
+	entry := index.OutputFiles[0]
+	if entry.LaneID != "spells" || entry.File != "lanes/spells.json" || entry.MediaType != "application/json" || entry.SchemaID != "notarius.dnd.spells" || entry.SchemaVersion != "v1" {
+		t.Fatalf("index output entry = %#v, want spells JSON contract", entry)
+	}
+
+	manifest := readProductionJSON[artifacts.RunManifest](t, filepath.Join(runRoot, "manifest.json"))
+	if manifest.PipelineID != "dnd-session" || manifest.InputModule != "seriatim" || manifest.Chunker != "generic" || manifest.OutputEncoder != "json" || manifest.ValidationStatus != "approved" || manifest.ChunkPlan == nil || manifest.ChunkPlan.Action != "bypassed" {
+		t.Fatalf("manifest = %#v, want approved minimal run", manifest)
+	}
+	if len(manifest.ArtifactLanes) != 1 {
+		t.Fatalf("manifest lanes = %#v, want exactly spells", manifest.ArtifactLanes)
+	}
+	lane := manifest.ArtifactLanes[0]
+	if lane.ID != "spells" || lane.Extractor != "dnd/spells" || lane.Merger != "appendorder" || lane.Normalizer != spellnormalize.Key {
+		t.Fatalf("manifest lane = %#v, want production spells composition", lane)
+	}
+	if len(manifest.References) != 0 {
+		t.Fatalf("base-only manifest references = %#v, want no overlay provenance", manifest.References)
+	}
+	extractorMetadata, ok := lane.Metadata["extractor"].(map[string]any)
+	if !ok || len(stringValues(extractorMetadata["catalog_overlay_ids"])) != 0 {
+		t.Fatalf("base-only extractor metadata = %#v, want no overlay IDs", lane.Metadata)
+	}
+
+	artifact := readProductionJSON[dnd.SpellList](t, filepath.Join(runRoot, entry.File))
+	if len(artifact.SpellCasts) != 1 || artifact.SpellCasts[0].Spell != "Cure Wounds" || artifact.SpellCasts[0].SourceRefs[0].SourceID != "session-alpha" {
+		t.Fatalf("artifact = %#v, want one source-linked Cure Wounds cast", artifact)
+	}
+	rejected := readProductionJSON[struct {
+		Rejected []json.RawMessage `json:"rejected"`
+	}](t, filepath.Join(runRoot, "rejected.json"))
+	if len(rejected.Rejected) != 0 {
+		t.Fatalf("rejected = %#v, want empty rejection list", rejected.Rejected)
+	}
+	warnings := readProductionJSON[struct {
+		Warnings []json.RawMessage `json:"warnings"`
+	}](t, filepath.Join(runRoot, "warnings.json"))
+	if len(warnings.Warnings) != 0 {
+		t.Fatalf("warnings = %#v, want empty warning list", warnings.Warnings)
+	}
+}
+
+func TestMaintainedMalformedInputOnlyRecordsDebugFailureWhenRequested(t *testing.T) {
+	malformed := filepath.Join(t.TempDir(), "malformed.json")
+	if err := os.WriteFile(malformed, []byte("{not valid json"), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	for _, debug := range []bool{false, true} {
+		name := "without debug"
+		if debug {
+			name = "with debug"
+		}
+		t.Run(name, func(t *testing.T) {
+			outputRoot := filepath.Join(t.TempDir(), "output")
+			debugRoot := filepath.Join(t.TempDir(), "debug")
+			options := productionRunOptions(t, &productionFakeLLMClient{})
+			args := []string{
+				"run", "dnd-session",
+				"--config", repositoryPath("examples", "dnd-minimal.config.yml"),
+				"--input", malformed, "--chunk_cache", "bypass", "--output-dir", outputRoot,
+			}
+			if debug {
+				args = append(args, "--debug", "--debug-dir", debugRoot)
+			}
+			var stdout, stderr strings.Builder
+			code := RunWithOptions(args, &stdout, &stderr, options)
+			if code != 1 || stdout.Len() != 0 || !strings.Contains(stderr.String(), "parse input") {
+				t.Fatalf("code=%d stdout=%q stderr=%q", code, stdout.String(), stderr.String())
+			}
+			assertAbsent(t, outputRoot)
+			if !debug {
+				assertAbsent(t, debugRoot)
+				return
+			}
+			bundle := onlyChildDir(t, debugRoot)
+			report := readProductionJSON[debugbundle.RunReport](t, filepath.Join(bundle, "summary", "run-report.json"))
+			if report.Succeeded || report.PipelineID != "dnd-session" {
+				t.Fatalf("failure report = %#v, want failed dnd-session report", report)
+			}
+		})
+	}
+}
+
+type exampleOutputIndex struct {
+	ManifestFile string                    `json:"manifest_file"`
+	OutputFiles  []exampleOutputIndexEntry `json:"output_files"`
+	RejectedFile string                    `json:"rejected_file"`
+	WarningsFile string                    `json:"warnings_file"`
+}
+
+type exampleOutputIndexEntry struct {
+	LaneID        string `json:"lane_id"`
+	MediaType     string `json:"media_type"`
+	File          string `json:"file"`
+	SchemaID      string `json:"schema_id"`
+	SchemaVersion string `json:"schema_version"`
+}
+
+func resolveInputForMaintainedExample(components productionComponents, pipelineID string) config.ResolveInput {
+	return config.ResolveInput{PipelineID: pipelineID, Catalog: catalogFromRegistries(components.registries)}
+}

internal/cli/npc_registry_contract_test.go

@@ -0,0 +1,80 @@
+package cli
+
+import (
+	"context"
+	"errors"
+	"fmt"
+	"io/fs"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+
+	"gitea.maximumdirect.net/eric/notarius/internal/core/artifacts"
+	"gitea.maximumdirect.net/eric/notarius/internal/core/config"
+	"gitea.maximumdirect.net/eric/notarius/internal/framework/contracts"
+	"gitea.maximumdirect.net/eric/notarius/internal/framework/pipeline"
+)
+
+func TestOversizedNPCRegistryFailsBeforeRuntimeAndCheckpointConstruction(t *testing.T) {
+	components := productionTestComponents(t)
+	npcPath := filepath.Join(t.TempDir(), "npcs.json")
+	if err := os.WriteFile(npcPath, []byte(strings.Repeat("x", 1048577)), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	checkpointRoot := filepath.Join(t.TempDir(), "checkpoints")
+	content := fmt.Sprintf(`version: 4
+cache:
+  chunk_plans:
+    mode: bypass
+  checkpoints:
+    enabled: true
+    directory: %q
+pipelines:
+  dnd-session:
+    input: seriatim
+    artifacts:
+      spells:
+        extract:
+          module: dnd/spells
+          references:
+            npcs: %q
+        normalize: dnd/spells
+`, checkpointRoot, npcPath)
+	configPath := filepath.Join(t.TempDir(), "config.yml")
+	if err := os.WriteFile(configPath, []byte(content), 0o600); err != nil {
+		t.Fatal(err)
+	}
+
+	llmConstructed := false
+	chunkStoreConstructed := false
+	options := Options{
+		Catalog:    catalogFromRegistries(components.registries),
+		Registries: components.registries,
+		LLMClientFactory: func(context.Context, config.Config, string, LLMRuntimeOverrides) (contracts.StructuredLLMClient, []artifacts.LLMProfileManifest, error) {
+			llmConstructed = true
+			return nil, nil, errors.New("LLM client must not be constructed")
+		},
+		ChunkPlanStoreFactory: func(string) (pipeline.ChunkPlanStore, error) {
+			chunkStoreConstructed = true
+			return nil, errors.New("chunk-plan store must not be constructed")
+		},
+	}
+	var stdout, stderr strings.Builder
+	code := RunWithOptions([]string{
+		"run", "dnd-session", "--config", configPath,
+		"--input", repositoryPath("examples", "seriatim-minimal-transcript.json"),
+		"--chunk_cache", "bypass", "--output-dir", t.TempDir(),
+	}, &stdout, &stderr, options)
+	for _, fragment := range []string{`pipeline "dnd-session"`, `reference slot "npcs"`, "1048577 bytes", "limit 1048576"} {
+		if code == 0 || !strings.Contains(stderr.String(), fragment) {
+			t.Fatalf("RunWithOptions() code = %d stderr = %q, want context fragment %q", code, stderr.String(), fragment)
+		}
+	}
+	if llmConstructed || chunkStoreConstructed {
+		t.Fatalf("runtime construction = LLM %t, chunk store %t; want materialization failure first", llmConstructed, chunkStoreConstructed)
+	}
+	if _, err := os.Stat(checkpointRoot); !errors.Is(err, fs.ErrNotExist) {
+		t.Fatalf("checkpoint root stat error = %v, want no checkpoint allocation", err)
+	}
+}

internal/cli/promptkit_profiles.go

@@ -0,0 +1,46 @@
+package cli
+
+import (
+	"context"
+	"fmt"
+
+	"gitea.maximumdirect.net/eric/notarius/internal/core/config"
+	"gitea.maximumdirect.net/eric/notarius/internal/framework/llm"
+)
+
+func validateExplicitPromptKitProfiles(ctx context.Context, cfg config.Config, profileIDs []string, assets *llm.AssetRegistry) error {
+	if len(profileIDs) == 0 {
+		return nil
+	}
+	inspector, err := llm.NewPromptKitProfileInspector(llm.PromptKitProfileInspectorConfig{
+		Source: promptKitProfileSourceConfig(cfg),
+		Assets: assets,
+	})
+	if err != nil {
+		return fmt.Errorf("load PromptKit profiles: %w", err)
+	}
+	for _, profileID := range profileIDs {
+		if _, err := inspector.InspectProfile(ctx, profileID); err != nil {
+			return err
+		}
+	}
+	return nil
+}
+
+func promptKitProfileSourceConfig(cfg config.Config) llm.PromptKitProfileSourceConfig {
+	return llm.PromptKitProfileSourceConfig{
+		ProfileDir:   cfg.PromptKit.ProfileDir,
+		ProfileFile:  cfg.PromptKit.ProfileFile,
+		LocalBackend: mapPromptKitLocalBackend(cfg.PromptKit.LocalBackend),
+	}
+}
+
+func mapPromptKitLocalBackend(cfg *config.PromptKitLocalBackendConfig) *llm.PromptKitLocalBackendConfig {
+	if cfg == nil {
+		return nil
+	}
+	return &llm.PromptKitLocalBackendConfig{
+		Endpoint:         cfg.Endpoint,
+		ConcurrencyLimit: cfg.ConcurrencyLimit,
+	}
+}

internal/cli/promptkit_profiles_test.go

@@ -0,0 +1,159 @@
+package cli
+
+import (
+	"context"
+	"errors"
+	"net/http"
+	"net/http/httptest"
+	"os"
+	"path/filepath"
+	"strings"
+	"sync/atomic"
+	"testing"
+	"testing/fstest"
+
+	"gitea.maximumdirect.net/eric/notarius/internal/core/config"
+	"gitea.maximumdirect.net/eric/notarius/internal/framework/llm"
+)
+
+func TestExplicitPromptKitProfileValidationInspectsProfilesWithoutGeneration(t *testing.T) {
+	var providerCalls atomic.Int32
+	server := httptest.NewServer(http.HandlerFunc(func(http.ResponseWriter, *http.Request) {
+		providerCalls.Add(1)
+	}))
+	defer server.Close()
+
+	writeProfile := func(t *testing.T, name, content string) string {
+		t.Helper()
+		profilePath := filepath.Join(t.TempDir(), name+".yaml")
+		if err := os.WriteFile(profilePath, []byte(content), 0o600); err != nil {
+			t.Fatal(err)
+		}
+		return profilePath
+	}
+	localProfile := "id: local-profile\nbackend: local\nmodel: local-model\n"
+	credentialProfile := `id: credential-profile
+endpoint: ` + server.URL + `/v1
+model: credential-model
+api_key_env: NOTARIUS_PROMPTKIT_PROFILE_INSPECTION_TEST_KEY
+`
+	t.Setenv("NOTARIUS_PROMPTKIT_PROFILE_INSPECTION_TEST_KEY", "")
+
+	tests := []struct {
+		name         string
+		profilePath  string
+		profileID    string
+		profileDir   bool
+		localBackend bool
+		canceled     bool
+		wantErr      []string
+		rejectErr    []string
+	}{
+		{
+			name:         "configured local backend",
+			profilePath:  writeProfile(t, "local-profile", localProfile),
+			profileID:    "local-profile",
+			profileDir:   true,
+			localBackend: true,
+		},
+		{
+			name:        "missing local backend registration",
+			profilePath: writeProfile(t, "local-profile", localProfile),
+			profileID:   "local-profile",
+			wantErr:     []string{`PromptKit profile "local-profile" is invalid or unreadable`},
+		},
+		{
+			name:         "absent profile",
+			profilePath:  writeProfile(t, "local-profile", localProfile),
+			profileID:    "absent-profile",
+			localBackend: true,
+			wantErr:      []string{`PromptKit profile "absent-profile" is not configured`},
+		},
+		{
+			name:        "malformed profile",
+			profilePath: writeProfile(t, "malformed-profile", "id: malformed-profile\nbackend: [\n"),
+			profileID:   "malformed-profile",
+			wantErr:     []string{`PromptKit profile "malformed-profile" is invalid or unreadable`},
+			rejectErr:   []string{"malformed-profile.yaml", "backend: ["},
+		},
+		{
+			name:        "invalid profile source",
+			profilePath: filepath.Join(t.TempDir(), "missing-profile.yaml"),
+			profileID:   "missing-profile",
+			wantErr:     []string{"load PromptKit profiles", "profile configuration is invalid or unreadable"},
+		},
+		{
+			name:        "credential environment intentionally unset",
+			profilePath: writeProfile(t, "credential-profile", credentialProfile),
+			profileID:   "credential-profile",
+		},
+		{
+			name:         "canceled inspection",
+			profilePath:  writeProfile(t, "local-profile", localProfile),
+			profileID:    "local-profile",
+			localBackend: true,
+			canceled:     true,
+			wantErr:      []string{"context canceled"},
+		},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			cfg := config.Default()
+			if tt.profileDir {
+				cfg.PromptKit.ProfileDir = filepath.Dir(tt.profilePath)
+			} else {
+				cfg.PromptKit.ProfileFile = tt.profilePath
+			}
+			if tt.localBackend {
+				cfg.PromptKit.LocalBackend = &config.PromptKitLocalBackendConfig{
+					Endpoint:         server.URL + "/v1",
+					ConcurrencyLimit: 2,
+				}
+			}
+			ctx := context.Background()
+			if tt.canceled {
+				var cancel context.CancelFunc
+				ctx, cancel = context.WithCancel(ctx)
+				cancel()
+			}
+			err := validateExplicitPromptKitProfiles(ctx, cfg, []string{tt.profileID}, nil)
+			if len(tt.wantErr) == 0 {
+				if err != nil {
+					t.Fatalf("validateExplicitPromptKitProfiles() error = %v, want nil", err)
+				}
+				return
+			}
+			if err == nil {
+				t.Fatal("validateExplicitPromptKitProfiles() error = nil, want failure")
+			}
+			if tt.canceled && !errors.Is(err, context.Canceled) {
+				t.Fatalf("canceled inspection error = %v, want context canceled", err)
+			}
+			for _, want := range tt.wantErr {
+				if !strings.Contains(err.Error(), want) {
+					t.Fatalf("validation error = %q, want %q", err, want)
+				}
+			}
+			for _, rejected := range append(tt.rejectErr, tt.profilePath) {
+				if rejected != "" && strings.Contains(err.Error(), rejected) {
+					t.Fatalf("validation error = %q, must not expose %q", err, rejected)
+				}
+			}
+		})
+	}
+	if providerCalls.Load() != 0 {
+		t.Fatalf("provider calls during profile inspection = %d, want 0", providerCalls.Load())
+	}
+}
+
+func TestExplicitPromptKitProfileValidationUsesFallbackAssets(t *testing.T) {
+	assets := llm.NewAssetRegistry()
+	if err := assets.RegisterFallbackProfileFS(fstest.MapFS{
+		"profiles/fallback.yaml": {Data: []byte("id: fallback-profile\nendpoint: http://promptkit.test/v1\nmodel: fallback-model\n")},
+	}, "profiles"); err != nil {
+		t.Fatalf("RegisterFallbackProfileFS() error = %v, want nil", err)
+	}
+	if err := validateExplicitPromptKitProfiles(context.Background(), config.Default(), []string{"fallback-profile"}, assets); err != nil {
+		t.Fatalf("validateExplicitPromptKitProfiles() error = %v, want nil", err)
+	}
+}

internal/cli/recompute_execution_contract_test.go

@@ -0,0 +1,323 @@
+package cli
+
+import (
+	"bytes"
+	"context"
+	"encoding/json"
+	"fmt"
+	"os"
+	"path/filepath"
+	"reflect"
+	"sort"
+	"strings"
+	"sync"
+	"testing"
+
+	"gitea.maximumdirect.net/eric/notarius/internal/framework/contracts"
+	"gitea.maximumdirect.net/eric/notarius/internal/framework/pipeline"
+)
+
+func TestRecomputeStepRecoversThroughFilesystemCheckpoints(t *testing.T) {
+	tests := []struct {
+		name             string
+		invalidateOutput bool
+		wantCode         int
+	}{
+		{name: "accepted producer is hydrated", wantCode: 0},
+		{name: "invalid producer stops dependents", invalidateOutput: true, wantCode: 1},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			roots := newRecomputeTestRoots(t)
+			harness := newRecomputeTestHarness()
+			fresh := runRecomputeCommand(roots, harness.options(), false)
+			if fresh.code != 0 {
+				t.Fatalf("fresh run code=%d stderr=%q", fresh.code, fresh.stderr)
+			}
+			removeCheckpointLaneStage(t, roots.checkpoints, "extract", "first", "producer")
+			removeCheckpointLaneStage(t, roots.checkpoints, "merge", "first", "producer")
+			if tt.invalidateOutput {
+				path := findCheckpointFile(t, roots.checkpoints, "normalize", "first", "producer", "output.json")
+				if err := os.WriteFile(path, []byte("{"), 0o600); err != nil {
+					t.Fatal(err)
+				}
+			}
+			harness.resetCalls()
+
+			resumed := runRecomputeCommand(roots, harness.options(), true)
+			if resumed.code != tt.wantCode {
+				t.Fatalf("resumed code=%d stdout=%q stderr=%q", resumed.code, resumed.stdout, resumed.stderr)
+			}
+			events := readLatestCheckpointEvents(t, roots.debug)
+			if tt.invalidateOutput {
+				if harness.callsFor("test/extract/middle") != 0 || harness.callsFor("test/extract/dependent") != 0 {
+					t.Fatalf("dependent calls after invalid producer = %#v", harness.callsSnapshot())
+				}
+				if !strings.Contains(resumed.stderr, string(pipeline.CheckpointReasonDecodeFailed)) {
+					t.Fatalf("stderr=%q, want stable checkpoint reason", resumed.stderr)
+				}
+				assertNormalizeDecisionSequence(t, events, []checkpointDecisionExpectation{{"first", "producer", pipeline.CheckpointDecisionExecuted, pipeline.CheckpointReasonDecodeFailed}})
+				return
+			}
+
+			if got := harness.callsSnapshot(); !reflect.DeepEqual(got, map[string]int{"test/extract/dependent": 1, "test/extract/middle": 1}) {
+				t.Fatalf("resumed extractor calls = %#v", got)
+			}
+			outputPath := filepath.Join(latestChildDir(t, roots.output), "result.json")
+			data, err := os.ReadFile(outputPath)
+			if err != nil {
+				t.Fatal(err)
+			}
+			if string(data) != "[\"producer\",\"unrelated\",\"middle\",\"dependent\"]\n" {
+				t.Fatalf("ordered output = %q", data)
+			}
+			assertNormalizeDecisionSequence(t, events, []checkpointDecisionExpectation{
+				{"first", "producer", pipeline.CheckpointDecisionReused, pipeline.CheckpointReasonAcceptedArtifactReused},
+				{"first", "unrelated", pipeline.CheckpointDecisionReused, pipeline.CheckpointReasonReused},
+				{"second", "middle", pipeline.CheckpointDecisionForcedRecompute, pipeline.CheckpointReasonRecomputeStep},
+				{"third", "dependent", pipeline.CheckpointDecisionForcedRecompute, pipeline.CheckpointReasonRecomputeStep},
+			})
+		})
+	}
+}
+
+type checkpointDecisionExpectation struct {
+	step, lane string
+	action     pipeline.CheckpointDecisionCategory
+	reason     pipeline.CheckpointReasonCode
+}
+
+func assertNormalizeDecisionSequence(t *testing.T, events []pipeline.CheckpointEvent, want []checkpointDecisionExpectation) {
+	t.Helper()
+	var got []checkpointDecisionExpectation
+	for _, event := range events {
+		if event.Stage == string(pipeline.StageNormalize) {
+			got = append(got, checkpointDecisionExpectation{event.StepID, event.LaneID, event.Action, event.ReasonCode})
+		}
+	}
+	if !reflect.DeepEqual(got, want) {
+		t.Fatalf("normalize decisions = %#v, want %#v", got, want)
+	}
+}
+
+type recomputeTestHarness struct {
+	base  *stateTestHarness
+	mu    sync.Mutex
+	calls map[string]int
+}
+
+func newRecomputeTestHarness() *recomputeTestHarness {
+	return &recomputeTestHarness{base: newStateTestHarness(), calls: make(map[string]int)}
+}
+
+func (h *recomputeTestHarness) options() Options {
+	opts := h.base.options()
+	for _, key := range []string{"test/extract/producer", "test/extract/unrelated", "test/extract/middle", "test/extract/dependent"} {
+		moduleKey := key
+		spec := pipeline.ModuleSpec{
+			Key: moduleKey, Stage: pipeline.StageExtract, ExecutionClass: contracts.ExecutionClassDeterministic, Requires: []string{"chunks"}, Provides: []string{"artifact"}, ArtifactKind: stateTestArtifactKind,
+			ReferenceSlots: []contracts.ReferenceSlot{{Name: "upstream", AcceptedMediaTypes: []string{"application/json"}, AcceptedArtifactKinds: []contracts.ArtifactKind{stateTestArtifactKind}}},
+		}
+		if err := pipeline.RegisterExtractor(opts.Registries.Extractors, spec, func() (contracts.Extractor[stateTestArtifact], error) {
+			return recomputeTestExtractor{key: moduleKey, harness: h}, nil
+		}); err != nil {
+			panic(err)
+		}
+	}
+	if err := opts.Registries.Outputs.RegisterWithSpec(pipeline.ModuleSpec{Key: "test/recompute-output", Stage: pipeline.StageOutput, ExecutionClass: contracts.ExecutionClassDeterministic, Requires: []string{"normalized"}, Provides: []string{"output"}}, func() (contracts.OutputEncoder, error) {
+		return recomputeTestOutput{}, nil
+	}); err != nil {
+		panic(err)
+	}
+	opts.Catalog = catalogFromRegistries(opts.Registries)
+	return opts
+}
+
+func (h *recomputeTestHarness) record(key string) {
+	h.mu.Lock()
+	defer h.mu.Unlock()
+	h.calls[key]++
+}
+
+func (h *recomputeTestHarness) resetCalls() {
+	h.mu.Lock()
+	defer h.mu.Unlock()
+	h.calls = make(map[string]int)
+}
+
+func (h *recomputeTestHarness) callsFor(key string) int {
+	h.mu.Lock()
+	defer h.mu.Unlock()
+	return h.calls[key]
+}
+
+func (h *recomputeTestHarness) callsSnapshot() map[string]int {
+	h.mu.Lock()
+	defer h.mu.Unlock()
+	result := make(map[string]int, len(h.calls))
+	for key, value := range h.calls {
+		result[key] = value
+	}
+	return result
+}
+
+type recomputeTestExtractor struct {
+	key     string
+	harness *recomputeTestHarness
+}
+
+func (e recomputeTestExtractor) Key() string { return e.key }
+func (e recomputeTestExtractor) ReferenceSlots() []contracts.ReferenceSlot {
+	return []contracts.ReferenceSlot{{Name: "upstream", AcceptedMediaTypes: []string{"application/json"}, AcceptedArtifactKinds: []contracts.ArtifactKind{stateTestArtifactKind}}}
+}
+func (e recomputeTestExtractor) Extract(context.Context, contracts.TypedExtractionRequest) (contracts.TypedExtractionResult[stateTestArtifact], error) {
+	e.harness.record(e.key)
+	return contracts.TypedExtractionResult[stateTestArtifact]{Value: stateTestArtifact{Value: e.key}}, nil
+}
+
+type recomputeTestOutput struct{}
+
+func (recomputeTestOutput) Key() string { return "test/recompute-output" }
+func (recomputeTestOutput) Encode(_ context.Context, req contracts.OutputRequest) (contracts.OutputResult, error) {
+	lanes := make([]string, 0, len(req.NormalizeOutputs))
+	for _, output := range req.NormalizeOutputs {
+		lanes = append(lanes, output.LaneID)
+	}
+	data, err := json.Marshal(lanes)
+	if err != nil {
+		return contracts.OutputResult{}, err
+	}
+	return contracts.OutputResult{Files: []contracts.OutputFile{{Name: "result.json", Bytes: append(data, '\n')}}}, nil
+}
+
+func newRecomputeTestRoots(t *testing.T) stateTestRoots {
+	t.Helper()
+	roots := newStateTestRoots(t)
+	config := fmt.Sprintf(`version: 4
+output:
+  directory: %q
+cache:
+  chunk_plans:
+    directory: %q
+    mode: bypass
+  checkpoints:
+    enabled: true
+    directory: %q
+debug:
+  directory: %q
+pipelines:
+  sample:
+    input: test/input
+    chunk: test/chunk
+    steps:
+      - id: first
+        artifacts:
+          producer:
+            extract: test/extract/producer
+            merge: test/merge
+            normalize: test/normalize
+          unrelated:
+            extract: test/extract/unrelated
+            merge: test/merge
+            normalize: test/normalize
+      - id: second
+        references:
+          upstream:
+            artifact:
+              step: first
+              lane: producer
+        artifacts:
+          middle:
+            extract: test/extract/middle
+            merge: test/merge
+            normalize: test/normalize
+      - id: third
+        references:
+          upstream:
+            artifact:
+              step: second
+              lane: middle
+        artifacts:
+          dependent:
+            extract: test/extract/dependent
+            merge: test/merge
+            normalize: test/normalize
+    output: test/recompute-output
+`, roots.output, roots.plans, roots.checkpoints, roots.debug)
+	if err := os.WriteFile(roots.config, []byte(config), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	return roots
+}
+
+func runRecomputeCommand(roots stateTestRoots, opts Options, recompute bool) stateTestResult {
+	args := []string{"run", "sample", "--config", roots.config, "--input", roots.input, "--chunk_cache", "bypass"}
+	if recompute {
+		args = append(args, "--resume", "--recompute-step", "second", "--debug")
+	}
+	var stdout, stderr bytes.Buffer
+	return stateTestResult{code: RunWithOptions(args, &stdout, &stderr, opts), stdout: stdout.String(), stderr: stderr.String()}
+}
+
+func removeCheckpointLaneStage(t *testing.T, root, stage, step, lane string) {
+	t.Helper()
+	dir := filepath.Dir(findCheckpointFile(t, root, stage, step, lane, "manifest.json"))
+	if err := os.RemoveAll(dir); err != nil {
+		t.Fatal(err)
+	}
+}
+
+func findCheckpointFile(t *testing.T, root, stage, step, lane, name string) string {
+	t.Helper()
+	want := filepath.Join(stage, step, lane, name)
+	var matches []string
+	err := filepath.WalkDir(root, func(path string, entry os.DirEntry, err error) error {
+		if err != nil {
+			return err
+		}
+		if !entry.IsDir() && strings.HasSuffix(path, want) {
+			matches = append(matches, path)
+		}
+		return nil
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(matches) != 1 {
+		t.Fatalf("checkpoint files ending in %q = %v", want, matches)
+	}
+	return matches[0]
+}
+
+func latestChildDir(t *testing.T, root string) string {
+	t.Helper()
+	entries, err := os.ReadDir(root)
+	if err != nil {
+		t.Fatal(err)
+	}
+	var dirs []string
+	for _, entry := range entries {
+		if entry.IsDir() {
+			dirs = append(dirs, filepath.Join(root, entry.Name()))
+		}
+	}
+	if len(dirs) == 0 {
+		t.Fatal("no child directory")
+	}
+	sort.Strings(dirs)
+	return dirs[len(dirs)-1]
+}
+
+func readLatestCheckpointEvents(t *testing.T, root string) []pipeline.CheckpointEvent {
+	t.Helper()
+	var events []pipeline.CheckpointEvent
+	data, err := os.ReadFile(filepath.Join(latestChildDir(t, root), "summary", "checkpoint-events.json"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	if err := json.Unmarshal(data, &events); err != nil {
+		t.Fatal(err)
+	}
+	return events
+}

internal/cli/recompute_policy_test.go

@@ -0,0 +1,56 @@
+package cli
+
+import (
+	"testing"
+
+	"gitea.maximumdirect.net/eric/notarius/internal/framework/pipeline"
+)
+
+func TestRecomputePolicyIncludesDependentLanesAndReusablePredecessors(t *testing.T) {
+	producer := pipeline.ResolvedArtifactLane{StepID: "first", ID: "producer"}
+	unrelated := pipeline.ResolvedArtifactLane{StepID: "first", ID: "unrelated"}
+	consumer := pipeline.ResolvedArtifactLane{
+		StepID: "second", ID: "consumer",
+		ExtractReferences: pipeline.ResolvedReferenceTarget{Bindings: []pipeline.ReferenceBinding{{Artifact: &pipeline.ArtifactReference{Step: "first", Lane: "producer"}}}},
+	}
+	downstream := pipeline.ResolvedArtifactLane{
+		StepID: "third", ID: "downstream",
+		ExtractReferences: pipeline.ResolvedReferenceTarget{Bindings: []pipeline.ReferenceBinding{{Artifact: &pipeline.ArtifactReference{Step: "second", Lane: "consumer"}}}},
+	}
+	independent := pipeline.ResolvedArtifactLane{StepID: "third", ID: "independent"}
+	resolved := pipeline.ResolvedPipeline{Steps: []pipeline.ResolvedPipelineStep{
+		{ID: "first", ArtifactLanes: []pipeline.ResolvedArtifactLane{producer, unrelated}},
+		{ID: "second", ArtifactLanes: []pipeline.ResolvedArtifactLane{consumer}},
+		{ID: "third", ArtifactLanes: []pipeline.ResolvedArtifactLane{downstream, independent}},
+	}}
+
+	policy, err := recomputePolicy(resolved, "second")
+	if err != nil {
+		t.Fatal(err)
+	}
+	if _, ok := policy.ForcedLanes[pipeline.CheckpointLaneKey("second", "consumer")]; !ok {
+		t.Fatal("selected lane was not forced")
+	}
+	if _, ok := policy.ForcedLanes[pipeline.CheckpointLaneKey("third", "downstream")]; !ok {
+		t.Fatal("transitive dependent lane was not forced")
+	}
+	if _, ok := policy.ForcedLanes[pipeline.CheckpointLaneKey("first", "producer")]; ok {
+		t.Fatal("predecessor was implicitly forced")
+	}
+	if _, ok := policy.RequireReusableLanes[pipeline.CheckpointLaneKey("first", "producer")]; !ok {
+		t.Fatal("required predecessor was not marked reusable")
+	}
+	if _, ok := policy.ForcedLanes[pipeline.CheckpointLaneKey("first", "unrelated")]; ok {
+		t.Fatal("unrelated lane was forced")
+	}
+	if _, ok := policy.ForcedLanes[pipeline.CheckpointLaneKey("third", "independent")]; ok {
+		t.Fatal("unrelated later lane was forced")
+	}
+}
+
+func TestRecomputePolicyRejectsUnknownStep(t *testing.T) {
+	_, err := recomputePolicy(pipeline.ResolvedPipeline{Steps: []pipeline.ResolvedPipelineStep{{ID: "known"}}}, "missing")
+	if err == nil {
+		t.Fatal("unknown step was accepted")
+	}
+}

internal/cli/reference_contract_test.go

@@ -0,0 +1,456 @@
+package cli
+
+import (
+	"bytes"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+
+	"gitea.maximumdirect.net/eric/notarius/internal/core/config"
+	"gitea.maximumdirect.net/eric/notarius/internal/framework/contracts"
+	"gitea.maximumdirect.net/eric/notarius/internal/framework/pipeline"
+)
+
+func TestReferenceSelectorsParseAndApplyAllDocumentedForms(t *testing.T) {
+	tests := []struct {
+		name      string
+		selector  string
+		only      []string
+		wantStage pipeline.ModuleStage
+		wantLane  string
+		wantSlot  string
+	}{
+		{name: "flat", selector: "alpha-slot", wantStage: pipeline.StageExtract, wantLane: "alpha", wantSlot: "alpha-slot"},
+		{name: "chunk", selector: "chunk.chunk-slot", wantStage: pipeline.StageChunk, wantSlot: "chunk-slot"},
+		{name: "merge", selector: "merge.alpha-merge", only: []string{"alpha"}, wantStage: pipeline.StageMerge, wantLane: "alpha", wantSlot: "alpha-merge"},
+		{name: "lane", selector: "alpha.alpha-slot", wantStage: pipeline.StageExtract, wantLane: "alpha", wantSlot: "alpha-slot"},
+		{name: "lane extract", selector: "alpha.extract.alpha-slot", wantStage: pipeline.StageExtract, wantLane: "alpha", wantSlot: "alpha-slot"},
+		{name: "lane merge", selector: "alpha.merge.alpha-merge", wantStage: pipeline.StageMerge, wantLane: "alpha", wantSlot: "alpha-merge"},
+		{name: "lane normalize", selector: "alpha.normalize.alpha-normalize", wantStage: pipeline.StageNormalize, wantLane: "alpha", wantSlot: "alpha-normalize"},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			cfg := referenceContractConfig()
+			catalog := referenceContractCatalog(t, true, true)
+			selector, err := parseReferenceSelector(tt.selector, "--reference")
+			if err != nil {
+				t.Fatal(err)
+			}
+			overrides, _, err := resolveCLIReferenceRequests(cfg, "demo", tt.only, catalog, []cliReferenceRequest{{Selector: selector, Source: "reference.txt"}}, nil)
+			if err != nil {
+				t.Fatalf("resolve selector: %v", err)
+			}
+			if len(overrides) != 1 {
+				t.Fatalf("overrides = %#v, want one binding", overrides)
+			}
+			got := overrides[0]
+			if got.Stage != tt.wantStage || got.LaneID != tt.wantLane || got.SlotName != tt.wantSlot || got.BindingSource != contracts.ReferenceBindingSourceCLI {
+				t.Fatalf("binding = %#v, want %s/%s/%s from CLI", got, tt.wantStage, tt.wantLane, tt.wantSlot)
+			}
+		})
+	}
+}
+
+func TestReferenceSelectorsRejectAmbiguityWithSpecificSuggestions(t *testing.T) {
+	cfg := referenceContractConfig()
+	catalog := referenceContractCatalog(t, true, true)
+	for _, tt := range []struct {
+		name     string
+		selector string
+		want     []string
+	}{
+		{name: "flat shared slot", selector: "shared", want: []string{"alpha.extract.shared", "beta.extract.shared"}},
+		{name: "lane shared slot", selector: "alpha.shared", want: []string{"alpha.extract.shared", "alpha.merge.shared", "alpha.normalize.shared"}},
+		{name: "all mergers", selector: "merge.shared", want: []string{"alpha.merge.shared", "beta.merge.shared"}},
+	} {
+		t.Run(tt.name, func(t *testing.T) {
+			selector, err := parseReferenceSelector(tt.selector, "--reference")
+			if err != nil {
+				t.Fatal(err)
+			}
+			_, _, err = resolveCLIReferenceRequests(cfg, "demo", nil, catalog, []cliReferenceRequest{{Selector: selector, Source: "reference.txt"}}, nil)
+			if err == nil {
+				t.Fatal("resolve selector succeeded, want ambiguity error")
+			}
+			for _, fragment := range tt.want {
+				if !strings.Contains(err.Error(), fragment) {
+					t.Fatalf("error = %q, want suggestion %q", err, fragment)
+				}
+			}
+		})
+	}
+}
+
+func TestReferenceSelectorsRespectSelectedLanesBeforeMaterialization(t *testing.T) {
+	cfg := referenceContractConfig()
+	catalog := referenceContractCatalog(t, true, true)
+	for _, tt := range []struct {
+		name     string
+		selector string
+		want     string
+	}{
+		{name: "unselected lane", selector: "beta.extract.beta-slot", want: `reference lane "beta" is not selected`},
+		{name: "unknown lane", selector: "missing.extract.beta-slot", want: `reference lane "missing" is not selected`},
+	} {
+		t.Run(tt.name, func(t *testing.T) {
+			selector, err := parseReferenceSelector(tt.selector, "--reference")
+			if err != nil {
+				t.Fatal(err)
+			}
+			_, _, err = resolveCLIReferenceRequests(cfg, "demo", []string{"alpha"}, catalog, []cliReferenceRequest{{Selector: selector, Source: filepath.Join(t.TempDir(), "missing.txt")}}, nil)
+			if err == nil || !strings.Contains(err.Error(), tt.want) || strings.Contains(err.Error(), "missing.txt") {
+				t.Fatalf("error = %v, want selection failure before file access", err)
+			}
+		})
+	}
+}
+
+func TestReferenceSyntaxErrorsReturnTwo(t *testing.T) {
+	tests := []struct {
+		name string
+		args []string
+	}{
+		{name: "reference missing value", args: []string{"run", "demo", "--config", "missing.yml", "--input", "input.txt", "--reference"}},
+		{name: "reference missing selector", args: []string{"run", "demo", "--config", "missing.yml", "--input", "input.txt", "--reference", "=path.txt"}},
+		{name: "reference missing separator", args: []string{"run", "demo", "--config", "missing.yml", "--input", "input.txt", "--reference", "slot"}},
+		{name: "reference missing path", args: []string{"run", "demo", "--config", "missing.yml", "--input", "input.txt", "--reference", "slot="}},
+		{name: "reference excess segments", args: []string{"run", "demo", "--config", "missing.yml", "--input", "input.txt", "--reference", "a.b.c.d=path.txt"}},
+		{name: "unbind with path", args: []string{"run", "demo", "--config", "missing.yml", "--input", "input.txt", "--without-reference", "slot=path.txt"}},
+		{name: "unbind excess segments", args: []string{"run", "demo", "--config", "missing.yml", "--input", "input.txt", "--without-reference", "a.b.c.d"}},
+		{name: "unbind missing value", args: []string{"run", "demo", "--config", "missing.yml", "--input", "input.txt", "--without-reference"}},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			var stdout, stderr bytes.Buffer
+			code := RunWithOptions(tt.args, &stdout, &stderr, Options{LookupEnv: emptyLookup})
+			if code != 2 || stdout.Len() != 0 || stderr.Len() == 0 {
+				t.Fatalf("code=%d stdout=%q stderr=%q", code, stdout.String(), stderr.String())
+			}
+		})
+	}
+}
+
+func TestReferenceOverridesUseFinalExactTargetBinding(t *testing.T) {
+	cfg := referenceContractConfig()
+	catalog := referenceContractCatalog(t, true, true)
+	alphaShared, err := parseReferenceSelector("alpha.extract.shared", "--reference")
+	if err != nil {
+		t.Fatal(err)
+	}
+	betaShared, err := parseReferenceSelector("beta.extract.shared", "--reference")
+	if err != nil {
+		t.Fatal(err)
+	}
+	overrides, unbinds, err := resolveCLIReferenceRequests(cfg, "demo", nil, catalog, []cliReferenceRequest{
+		{Selector: alphaShared, Source: "alpha-first.txt"},
+		{Selector: alphaShared, Source: "alpha-final.txt"},
+		{Selector: betaShared, Source: "beta-only.txt"},
+	}, nil)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(unbinds) != 0 {
+		t.Fatalf("unbinds = %#v, want none", unbinds)
+	}
+	effective, err := cfg.Resolve(config.ResolveInput{PipelineID: "demo", Catalog: catalog, ReferenceOverrides: overrides})
+	if err != nil {
+		t.Fatalf("resolve pipeline: %v", err)
+	}
+	alpha := referenceContractLane(t, effective.ResolvedPipeline, "alpha")
+	beta := referenceContractLane(t, effective.ResolvedPipeline, "beta")
+	if source := referenceContractBindingSource(alpha.ExtractReferences.Bindings, "shared"); source != "alpha-final.txt" {
+		t.Fatalf("alpha shared source = %q, want final exact-target override", source)
+	}
+	if source := referenceContractBindingSource(beta.ExtractReferences.Bindings, "shared"); source != "beta-only.txt" {
+		t.Fatalf("beta shared source = %q, want target-specific override", source)
+	}
+}
+
+func TestReferenceUnbindsRemoveOptionalAndProtectRequiredSlots(t *testing.T) {
+	cfg := referenceContractConfig()
+	catalog := referenceContractCatalog(t, true, true)
+	optional, err := parseReferenceSelector("alpha.extract.alpha-slot", "--without-reference")
+	if err != nil {
+		t.Fatal(err)
+	}
+	_, without, err := resolveCLIReferenceRequests(cfg, "demo", nil, catalog, nil, []cliReferenceUnbindRequest{{Selector: optional}})
+	if err != nil {
+		t.Fatal(err)
+	}
+	effective, err := cfg.Resolve(config.ResolveInput{PipelineID: "demo", Catalog: catalog, ReferenceUnbinds: without})
+	if err != nil {
+		t.Fatalf("optional unbind: %v", err)
+	}
+	if binding := referenceContractFindBinding(referenceContractLane(t, effective.ResolvedPipeline, "alpha").ExtractReferences.Bindings, "alpha-slot"); binding != nil {
+		t.Fatalf("optional binding after unbind = %#v, want absent", binding)
+	}
+
+	for _, tt := range []struct {
+		name     string
+		selector string
+	}{
+		{name: "chunk", selector: "chunk.required-chunk"},
+		{name: "extract", selector: "alpha.extract.required-extract"},
+		{name: "merge", selector: "alpha.merge.required-merge"},
+		{name: "normalize", selector: "alpha.normalize.required-normalize"},
+	} {
+		t.Run(tt.name, func(t *testing.T) {
+			selector, err := parseReferenceSelector(tt.selector, "--without-reference")
+			if err != nil {
+				t.Fatal(err)
+			}
+			_, unbinds, err := resolveCLIReferenceRequests(cfg, "demo", nil, catalog, nil, []cliReferenceUnbindRequest{{Selector: selector}})
+			if err != nil {
+				t.Fatal(err)
+			}
+			_, err = cfg.Resolve(config.ResolveInput{PipelineID: "demo", Catalog: catalog, ReferenceUnbinds: unbinds})
+			if err == nil || !strings.Contains(err.Error(), "required reference slot") {
+				t.Fatalf("resolve error = %v, want required-slot failure", err)
+			}
+		})
+	}
+}
+
+func TestReferenceMaterializationSeparatesCLIAndConfigPathOrigins(t *testing.T) {
+	configDir := t.TempDir()
+	workingDir := t.TempDir()
+	cfg := referenceContractConfig()
+	configPath := filepath.Join(configDir, "config.yml")
+	if err := os.WriteFile(configPath, []byte("version: 4\n"), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.WriteFile(filepath.Join(configDir, "required.txt"), []byte("config reference"), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.WriteFile(filepath.Join(configDir, "optional.txt"), []byte("optional reference"), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.WriteFile(filepath.Join(workingDir, "cli-reference.txt"), []byte("CLI reference"), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	catalog := referenceContractCatalog(t, true, true)
+	selector, err := parseReferenceSelector("alpha.extract.alpha-slot", "--reference")
+	if err != nil {
+		t.Fatal(err)
+	}
+	overrides, unbinds, err := resolveCLIReferenceRequests(cfg, "demo", nil, catalog, []cliReferenceRequest{{Selector: selector, Source: "cli-reference.txt"}}, nil)
+	if err != nil {
+		t.Fatal(err)
+	}
+	effective, err := cfg.Resolve(config.ResolveInput{PipelineID: "demo", Catalog: catalog, ReferenceOverrides: overrides, ReferenceUnbinds: unbinds})
+	if err != nil {
+		t.Fatalf("resolve pipeline: %v", err)
+	}
+	materialized, _, err := pipeline.MaterializeReferences(effective.ResolvedPipeline, catalog, pipeline.ReferenceMaterializationOptions{ConfigPath: configPath, WorkingDir: workingDir})
+	if err != nil {
+		t.Fatalf("materialize references: %v", err)
+	}
+	alpha := referenceContractLane(t, materialized, "alpha")
+	cliItem := alpha.ExtractReferences.ReferenceSet.Slots["alpha-slot"].Items[0]
+	if string(cliItem.Content) != "CLI reference" || cliItem.BindingSource != contracts.ReferenceBindingSourceCLI || cliItem.Origin.URI != referenceContractFileURI(filepath.Join(workingDir, "cli-reference.txt")) {
+		t.Fatalf("CLI materialization = %#v, want working-directory provenance", cliItem)
+	}
+	configItem := alpha.ExtractReferences.ReferenceSet.Slots["required-extract"].Items[0]
+	if string(configItem.Content) != "config reference" || configItem.BindingSource != contracts.ReferenceBindingSourceConfig || configItem.Origin.URI != referenceContractFileURI(filepath.Join(configDir, "required.txt")) {
+		t.Fatalf("config materialization = %#v, want config-directory provenance", configItem)
+	}
+}
+
+func TestReferenceTargetLookupUsesArtifactVariantsAndReportsMissingContext(t *testing.T) {
+	cfg := referenceContractConfig()
+	full := referenceContractCatalog(t, true, true)
+	targets, err := selectedReferenceTargets(cfg, "demo", nil, full)
+	if err != nil {
+		t.Fatalf("select reference targets: %v", err)
+	}
+	var alphaMerge, betaMerge selectedReferenceTarget
+	for _, target := range targets {
+		if target.stage == pipeline.StageMerge && target.laneID == "alpha" {
+			alphaMerge = target
+		}
+		if target.stage == pipeline.StageMerge && target.laneID == "beta" {
+			betaMerge = target
+		}
+	}
+	if _, ok := alphaMerge.slots["alpha-merge"]; !ok {
+		t.Fatalf("alpha merger slots = %#v, want alpha artifact variant", alphaMerge.slots)
+	}
+	if _, ok := betaMerge.slots["beta-merge"]; !ok {
+		t.Fatalf("beta merger slots = %#v, want beta artifact variant", betaMerge.slots)
+	}
+	if _, ok := betaMerge.slots["alpha-merge"]; ok {
+		t.Fatalf("beta merger slots = %#v, must not use alpha variant", betaMerge.slots)
+	}
+
+	missingMerger := referenceContractCatalog(t, false, true)
+	_, err = selectedReferenceTargets(cfg, "demo", nil, missingMerger)
+	if err == nil || !strings.Contains(err.Error(), "merger") || !strings.Contains(err.Error(), string(referenceContractKindBeta)) {
+		t.Fatalf("missing merger error = %v, want artifact variant context", err)
+	}
+	missingNormalizer := referenceContractCatalog(t, true, false)
+	_, err = selectedReferenceTargets(cfg, "demo", nil, missingNormalizer)
+	if err == nil || !strings.Contains(err.Error(), "normalizer") || !strings.Contains(err.Error(), string(referenceContractKindBeta)) {
+		t.Fatalf("missing normalizer error = %v, want artifact variant context", err)
+	}
+	missingExtractor := referenceContractCatalog(t, true, true)
+	missingExtractor.Extractors = pipeline.NewExtractorRegistry()
+	_, err = selectedReferenceTargets(cfg, "demo", nil, missingExtractor)
+	if err == nil || !strings.Contains(err.Error(), `lane "alpha" extract module`) || !strings.Contains(err.Error(), "not registered") {
+		t.Fatalf("missing extractor error = %v, want lane/module context", err)
+	}
+}
+
+const (
+	referenceContractKindAlpha contracts.ArtifactKind = "reference/alpha"
+	referenceContractKindBeta  contracts.ArtifactKind = "reference/beta"
+)
+
+func referenceContractConfig() config.Config {
+	cfg := config.Default()
+	cfg.Pipelines = map[string]pipeline.PipelineProfile{
+		"demo": {
+			ID:     "demo",
+			Input:  pipeline.Binding("reference/input"),
+			Chunk:  pipeline.Binding("reference/chunk"),
+			Output: pipeline.Binding("reference/output"),
+			Artifacts: map[string]pipeline.ArtifactLaneProfile{
+				"alpha": {
+					Extract:    pipeline.Binding("reference/extract-alpha"),
+					Merge:      pipeline.Binding("reference/shared-merge"),
+					Normalize:  pipeline.Binding("reference/shared-normalize"),
+					References: pipeline.ExternalReferenceMap(map[string]string{"required-extract": "required.txt"}),
+				},
+				"beta": {
+					Extract:    pipeline.Binding("reference/extract-beta"),
+					Merge:      pipeline.Binding("reference/shared-merge"),
+					Normalize:  pipeline.Binding("reference/shared-normalize"),
+					References: pipeline.ExternalReferenceMap(map[string]string{"required-extract": "required.txt"}),
+				},
+			},
+		},
+	}
+	profile := cfg.Pipelines["demo"]
+	profile.Chunk.References = pipeline.ExternalReferenceMap(map[string]string{"required-chunk": "required.txt"})
+	alpha := profile.Artifacts["alpha"]
+	alpha.Extract.References = pipeline.ExternalReferenceMap(map[string]string{"required-extract": "required.txt", "alpha-slot": "optional.txt"})
+	alpha.Merge.References = pipeline.ExternalReferenceMap(map[string]string{"required-merge": "required.txt"})
+	alpha.Normalize.References = pipeline.ExternalReferenceMap(map[string]string{"required-normalize": "required.txt"})
+	profile.Artifacts["alpha"] = alpha
+	beta := profile.Artifacts["beta"]
+	beta.Extract.References = pipeline.ExternalReferenceMap(map[string]string{"required-extract": "required.txt"})
+	beta.Merge.References = pipeline.ExternalReferenceMap(map[string]string{"required-merge": "required.txt"})
+	beta.Normalize.References = pipeline.ExternalReferenceMap(map[string]string{"required-normalize": "required.txt"})
+	profile.Artifacts["beta"] = beta
+	cfg.Pipelines["demo"] = profile
+	return cfg
+}
+
+func referenceContractCatalog(t *testing.T, includeBetaMerger, includeBetaNormalizer bool) pipeline.ModuleCatalog {
+	t.Helper()
+	registries := pipeline.Registries{
+		Inputs:          pipeline.NewInputAdapterRegistry(),
+		Chunkers:        pipeline.NewChunkerRegistry(),
+		ArtifactCodecs:  pipeline.NewArtifactCodecRegistry(),
+		Extractors:      pipeline.NewExtractorRegistry(),
+		Mergers:         pipeline.NewMergerRegistry(),
+		Normalizers:     pipeline.NewNormalizerRegistry(),
+		Validators:      pipeline.NewValidatorRegistry(),
+		ValidatorChains: pipeline.NewValidatorChainRegistry(),
+		Outputs:         pipeline.NewOutputEncoderRegistry(),
+	}
+	register := func(err error) {
+		if err != nil {
+			t.Fatal(err)
+		}
+	}
+	register(registries.Inputs.RegisterBuilderWithSpec(pipeline.ModuleSpec{Key: "reference/input", Stage: pipeline.StageInput, ExecutionClass: contracts.ExecutionClassDeterministic, Provides: []string{"source"}}, func(map[string]any) error { return nil }, func(pipeline.BuildRequest) (contracts.InputAdapter, error) { return stateTestInput{}, nil }))
+	register(registries.Chunkers.RegisterWithSpec(pipeline.ModuleSpec{Key: "reference/chunk", Stage: pipeline.StageChunk, ExecutionClass: contracts.ExecutionClassDeterministic, Requires: []string{"source"}, Provides: []string{"chunks"}, ReferenceSlots: []contracts.ReferenceSlot{{Name: "chunk-slot"}, {Name: "required-chunk", Required: true}}}, func() (contracts.Chunker, error) { return stateTestChunker{}, nil }))
+	register(pipeline.RegisterArtifactCodec(registries.ArtifactCodecs, referenceContractCodecA{}))
+	register(pipeline.RegisterArtifactCodec(registries.ArtifactCodecs, referenceContractCodecB{}))
+	register(pipeline.RegisterExtractor(registries.Extractors, pipeline.ModuleSpec{Key: "reference/extract-alpha", Stage: pipeline.StageExtract, ExecutionClass: contracts.ExecutionClassDeterministic, Requires: []string{"chunks"}, Provides: []string{"artifact"}, ArtifactKind: referenceContractKindAlpha, ReferenceSlots: []contracts.ReferenceSlot{{Name: "shared"}, {Name: "alpha-slot"}, {Name: "required-extract", Required: true}}}, func() (contracts.Extractor[stateTestArtifact], error) { return stateTestExtractor{}, nil }))
+	register(pipeline.RegisterExtractor(registries.Extractors, pipeline.ModuleSpec{Key: "reference/extract-beta", Stage: pipeline.StageExtract, ExecutionClass: contracts.ExecutionClassDeterministic, Requires: []string{"chunks"}, Provides: []string{"artifact"}, ArtifactKind: referenceContractKindBeta, ReferenceSlots: []contracts.ReferenceSlot{{Name: "shared"}, {Name: "beta-slot"}, {Name: "required-extract", Required: true}}}, func() (contracts.Extractor[stateTestArtifact], error) { return stateTestExtractor{}, nil }))
+	register(pipeline.RegisterMerger(registries.Mergers, pipeline.ModuleSpec{Key: "reference/shared-merge", Stage: pipeline.StageMerge, ExecutionClass: contracts.ExecutionClassDeterministic, Requires: []string{"artifact"}, Provides: []string{"merged"}, ArtifactKind: referenceContractKindAlpha, ReferenceSlots: []contracts.ReferenceSlot{{Name: "shared"}, {Name: "alpha-merge"}, {Name: "required-merge", Required: true}}}, func() (contracts.Merger[stateTestArtifact], error) { return stateTestMerger{}, nil }))
+	if includeBetaMerger {
+		register(pipeline.RegisterMerger(registries.Mergers, pipeline.ModuleSpec{Key: "reference/shared-merge", Stage: pipeline.StageMerge, ExecutionClass: contracts.ExecutionClassDeterministic, Requires: []string{"artifact"}, Provides: []string{"merged"}, ArtifactKind: referenceContractKindBeta, ReferenceSlots: []contracts.ReferenceSlot{{Name: "shared"}, {Name: "beta-merge"}, {Name: "required-merge", Required: true}}}, func() (contracts.Merger[stateTestArtifact], error) { return stateTestMerger{}, nil }))
+	}
+	register(pipeline.RegisterNormalizer(registries.Normalizers, pipeline.ModuleSpec{Key: "reference/shared-normalize", Stage: pipeline.StageNormalize, ExecutionClass: contracts.ExecutionClassDeterministic, Requires: []string{"merged"}, Provides: []string{"normalized"}, ArtifactKind: referenceContractKindAlpha, ReferenceSlots: []contracts.ReferenceSlot{{Name: "shared"}, {Name: "alpha-normalize"}, {Name: "required-normalize", Required: true}}}, func() (contracts.Normalizer[stateTestArtifact], error) { return stateTestNormalizer{}, nil }))
+	if includeBetaNormalizer {
+		register(pipeline.RegisterNormalizer(registries.Normalizers, pipeline.ModuleSpec{Key: "reference/shared-normalize", Stage: pipeline.StageNormalize, ExecutionClass: contracts.ExecutionClassDeterministic, Requires: []string{"merged"}, Provides: []string{"normalized"}, ArtifactKind: referenceContractKindBeta, ReferenceSlots: []contracts.ReferenceSlot{{Name: "shared"}, {Name: "beta-normalize"}, {Name: "required-normalize", Required: true}}}, func() (contracts.Normalizer[stateTestArtifact], error) { return stateTestNormalizer{}, nil }))
+	}
+	register(registries.Outputs.RegisterWithSpec(pipeline.ModuleSpec{Key: "reference/output", Stage: pipeline.StageOutput, ExecutionClass: contracts.ExecutionClassDeterministic, Requires: []string{"normalized"}, Provides: []string{"output"}}, func() (contracts.OutputEncoder, error) { return stateTestOutput{}, nil }))
+	return catalogFromRegistries(registries)
+}
+
+type referenceContractCodecB struct{}
+
+type referenceContractCodecA struct{}
+
+func (referenceContractCodecA) Kind() contracts.ArtifactKind { return referenceContractKindAlpha }
+func (referenceContractCodecA) Schema() contracts.ArtifactSchema {
+	return contracts.ArtifactSchema{ID: "reference.alpha", Name: "reference_alpha", Version: "v1", JSONSchema: []byte(`{"type":"object"}`)}
+}
+func (referenceContractCodecA) MediaType() string { return "application/json" }
+func (referenceContractCodecA) EncodeCandidate(stateTestArtifact) ([]byte, error) {
+	return []byte(`{"value":"ok"}`), nil
+}
+func (referenceContractCodecA) Encode(stateTestArtifact) ([]byte, error) {
+	return []byte(`{"value":"ok"}`), nil
+}
+func (referenceContractCodecA) Decode([]byte) (stateTestArtifact, error) {
+	return stateTestArtifact{Value: "ok"}, nil
+}
+
+func (referenceContractCodecB) Kind() contracts.ArtifactKind { return referenceContractKindBeta }
+func (referenceContractCodecB) Schema() contracts.ArtifactSchema {
+	return contracts.ArtifactSchema{ID: "reference.beta", Name: "reference_beta", Version: "v1", JSONSchema: []byte(`{"type":"object"}`)}
+}
+func (referenceContractCodecB) MediaType() string { return "application/json" }
+func (referenceContractCodecB) EncodeCandidate(stateTestArtifact) ([]byte, error) {
+	return []byte(`{"value":"ok"}`), nil
+}
+func (referenceContractCodecB) Encode(stateTestArtifact) ([]byte, error) {
+	return []byte(`{"value":"ok"}`), nil
+}
+func (referenceContractCodecB) Decode([]byte) (stateTestArtifact, error) {
+	return stateTestArtifact{Value: "ok"}, nil
+}
+
+func referenceContractLane(t *testing.T, resolved pipeline.ResolvedPipeline, id string) pipeline.ResolvedArtifactLane {
+	t.Helper()
+	for _, step := range resolved.Steps {
+		for _, lane := range step.ArtifactLanes {
+			if lane.ID == id {
+				return lane
+			}
+		}
+	}
+	t.Fatalf("lane %q not found", id)
+	return pipeline.ResolvedArtifactLane{}
+}
+
+func referenceContractBindingSource(bindings []pipeline.ReferenceBinding, slot string) string {
+	for _, binding := range bindings {
+		if binding.SlotName == slot {
+			return binding.Source
+		}
+	}
+	return ""
+}
+
+func referenceContractFindBinding(bindings []pipeline.ReferenceBinding, slot string) *pipeline.ReferenceBinding {
+	for i := range bindings {
+		if bindings[i].SlotName == slot {
+			return &bindings[i]
+		}
+	}
+	return nil
+}
+
+func referenceContractFileURI(path string) string {
+	absolute, err := filepath.Abs(path)
+	if err != nil {
+		absolute = path
+	}
+	return "file://" + filepath.ToSlash(absolute)
+}

internal/cli/run_contract_test.go

@@ -0,0 +1,654 @@
+package cli
+
+import (
+	"bytes"
+	"context"
+	"errors"
+	"fmt"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+
+	"gitea.maximumdirect.net/eric/notarius/internal/core/artifacts"
+	"gitea.maximumdirect.net/eric/notarius/internal/core/config"
+	"gitea.maximumdirect.net/eric/notarius/internal/framework/checkpoint"
+	"gitea.maximumdirect.net/eric/notarius/internal/framework/contracts"
+	"gitea.maximumdirect.net/eric/notarius/internal/framework/pipeline"
+)
+
+func TestRunControlsRejectSyntaxWithoutAllocatingState(t *testing.T) {
+	tests := []struct {
+		name string
+		args func(stateTestRoots) []string
+	}{
+		{name: "missing pipeline", args: func(roots stateTestRoots) []string {
+			return []string{"run", "--config", roots.config, "--input", roots.input}
+		}},
+		{name: "missing input", args: func(roots stateTestRoots) []string {
+			return []string{"run", "sample", "--config", roots.config}
+		}},
+		{name: "unknown flag", args: func(roots stateTestRoots) []string {
+			return []string{"run", "sample", "--config", roots.config, "--input", roots.input, "--unknown"}
+		}},
+		{name: "blank output directory", args: func(roots stateTestRoots) []string {
+			return []string{"run", "sample", "--config", roots.config, "--input", roots.input, "--output-dir", ""}
+		}},
+		{name: "blank debug directory", args: func(roots stateTestRoots) []string {
+			return []string{"run", "sample", "--config", roots.config, "--input", roots.input, "--debug-dir", ""}
+		}},
+		{name: "debug directory without debug", args: func(roots stateTestRoots) []string {
+			return []string{"run", "sample", "--config", roots.config, "--input", roots.input, "--debug-dir", filepath.Join(filepath.Dir(roots.debug), "requested-debug")}
+		}},
+		{name: "blank session ID", args: func(roots stateTestRoots) []string {
+			return []string{"run", "sample", "--config", roots.config, "--input", roots.input, "--session-id", ""}
+		}},
+		{name: "multiple pipeline IDs", args: func(roots stateTestRoots) []string {
+			return []string{"run", "sample", "extra", "--config", roots.config, "--input", roots.input}
+		}},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			roots := newStateTestRoots(t)
+			var stdout, stderr bytes.Buffer
+			code := RunWithOptions(tt.args(roots), &stdout, &stderr, newStateTestHarness().options())
+			if code != 2 || stdout.Len() != 0 || stderr.Len() == 0 {
+				t.Fatalf("code=%d stdout=%q stderr=%q", code, stdout.String(), stderr.String())
+			}
+			assertAbsent(t, roots.output)
+			assertAbsent(t, roots.debug)
+		})
+	}
+}
+
+func TestRecomputeStepCLIContract(t *testing.T) {
+	tests := []struct {
+		name       string
+		configure  func(*testing.T, stateTestRoots)
+		flags      []string
+		wantCode   int
+		wantOutput string
+		wantError  string
+	}{
+		{
+			name: "explicit step",
+			configure: func(t *testing.T, roots stateTestRoots) {
+				replaceStateTestConfigLine(t, roots.config, "    artifacts:\n      items:\n        extract: test/extract\n        merge: test/merge\n        normalize: test/normalize\n", "    steps:\n      - id: chosen\n        artifacts:\n          items:\n            extract: test/extract\n            merge: test/merge\n            normalize: test/normalize\n")
+			},
+			flags:      []string{"--resume", "--recompute-step", "chosen"},
+			wantCode:   0,
+			wantOutput: "outputs=1",
+		},
+		{name: "implicit default step", flags: []string{"--resume", "--recompute-step", "default"}, wantCode: 0, wantOutput: "outputs=1"},
+		{name: "repeated flag", flags: []string{"--resume", "--recompute-step", "default", "--recompute-step", "default"}, wantCode: 2, wantError: "specified only once"},
+		{name: "empty step", flags: []string{"--resume", "--recompute-step", ""}, wantCode: 2, wantError: "must not be empty"},
+		{name: "unknown step", flags: []string{"--resume", "--recompute-step", "missing"}, wantCode: 1, wantError: "unknown pipeline step"},
+		{name: "without resume", flags: []string{"--recompute-step", "default"}, wantCode: 2, wantError: "requires --resume"},
+		{
+			name: "checkpoint recording disabled",
+			configure: func(t *testing.T, roots stateTestRoots) {
+				replaceStateTestConfigLine(t, roots.config, "    enabled: true\n", "    enabled: false\n")
+			},
+			flags: []string{"--resume", "--recompute-step", "default"}, wantCode: 1, wantError: "cache.checkpoints.enabled",
+		},
+		{name: "with only", flags: []string{"--resume", "--recompute-step", "default", "--only", "items"}, wantCode: 2, wantError: "cannot be combined with --only"},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			roots := newStateTestRoots(t)
+			if tt.configure != nil {
+				tt.configure(t, roots)
+			}
+			args := []string{"run", "sample", "--config", roots.config, "--input", roots.input, "--chunk_cache", "bypass"}
+			args = append(args, tt.flags...)
+			var stdout, stderr bytes.Buffer
+			code := RunWithOptions(args, &stdout, &stderr, newStateTestHarness().options())
+			if code != tt.wantCode || (tt.wantOutput != "" && !strings.Contains(stdout.String(), tt.wantOutput)) || (tt.wantError != "" && !strings.Contains(stderr.String(), tt.wantError)) {
+				t.Fatalf("code=%d stdout=%q stderr=%q", code, stdout.String(), stderr.String())
+			}
+		})
+	}
+}
+
+func TestRunValidFailuresClassifyAndReportDebug(t *testing.T) {
+	tests := []struct {
+		name      string
+		args      func(stateTestRoots) []string
+		wantError string
+		wantDebug bool
+	}{
+		{name: "unknown pipeline", args: func(roots stateTestRoots) []string {
+			return []string{"run", "missing", "--config", roots.config, "--input", roots.input, "--chunk_cache", "bypass"}
+		}, wantError: `pipeline "missing"`},
+		{name: "unknown lane", args: func(roots stateTestRoots) []string {
+			return []string{"run", "sample", "--config", roots.config, "--input", roots.input, "--only", "missing", "--chunk_cache", "bypass", "--debug"}
+		}, wantError: `lane "missing"`, wantDebug: true},
+		{name: "unreadable input", args: func(roots stateTestRoots) []string {
+			return []string{"run", "sample", "--config", roots.config, "--input", filepath.Join(filepath.Dir(roots.input), "unreadable.txt"), "--chunk_cache", "bypass", "--debug"}
+		}, wantError: "read input", wantDebug: true},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			roots := newStateTestRoots(t)
+			var stdout, stderr bytes.Buffer
+			code := RunWithOptions(tt.args(roots), &stdout, &stderr, newStateTestHarness().options())
+			if code != 1 || stdout.Len() != 0 || !strings.Contains(stderr.String(), tt.wantError) {
+				t.Fatalf("code=%d stdout=%q stderr=%q", code, stdout.String(), stderr.String())
+			}
+			if tt.wantDebug {
+				if !strings.Contains(stderr.String(), "debug=") {
+					t.Fatalf("stderr=%q, want debug path", stderr.String())
+				}
+				onlyChildDir(t, roots.debug)
+			} else {
+				assertAbsent(t, roots.debug)
+			}
+			assertAbsent(t, roots.output)
+		})
+	}
+}
+
+func TestRunOnlyExecutesSelectedLanes(t *testing.T) {
+	roots := newStateTestRoots(t)
+	data, err := os.ReadFile(roots.config)
+	if err != nil {
+		t.Fatal(err)
+	}
+	data = []byte(replaceRequiredOnce(t, string(data), "    output: test/output\n", "      other:\n        extract: test/extract\n    output: test/output\n"))
+	if err := os.WriteFile(roots.config, data, 0o600); err != nil {
+		t.Fatal(err)
+	}
+	harness := newStateTestHarness()
+	var stdout, stderr bytes.Buffer
+	code := RunWithOptions([]string{"run", "sample", "--config", roots.config, "--input", roots.input, "--only", "items", "--chunk_cache", "bypass"}, &stdout, &stderr, harness.options())
+	if code != 0 || !strings.Contains(stdout.String(), "outputs=1") || stderr.Len() != 0 {
+		t.Fatalf("code=%d stdout=%q stderr=%q", code, stdout.String(), stderr.String())
+	}
+	harness.mu.Lock()
+	extractCalls := harness.extractCalls
+	harness.mu.Unlock()
+	if extractCalls != 1 {
+		t.Fatalf("extract calls = %d, want only the selected lane", extractCalls)
+	}
+}
+
+func TestRunStateRootsHonorEnvironmentFlagsAndDefaults(t *testing.T) {
+	t.Run("environment roots", func(t *testing.T) {
+		roots := newStateTestRoots(t)
+		environmentOutput := filepath.Join(t.TempDir(), "environment-output")
+		environmentDebug := filepath.Join(t.TempDir(), "environment-debug")
+		opts := newStateTestHarness().options()
+		opts.LookupEnv = lookupRunContractEnv(map[string]string{
+			"NOTARIUS_OUTPUT_DIR": environmentOutput,
+			"NOTARIUS_DEBUG_DIR":  environmentDebug,
+		})
+		result := runWithStateRoots(t, roots, opts, nil)
+		if result.code != 0 {
+			t.Fatalf("code=%d stdout=%q stderr=%q", result.code, result.stdout, result.stderr)
+		}
+		assertFile(t, filepath.Join(environmentOutput, filepath.Base(onlyChildDir(t, environmentOutput)), "result.json"))
+		onlyChildDir(t, environmentDebug)
+		assertAbsent(t, roots.output)
+		assertAbsent(t, roots.debug)
+	})
+
+	t.Run("command flags override environment", func(t *testing.T) {
+		roots := newStateTestRoots(t)
+		environmentOutput := filepath.Join(t.TempDir(), "environment-output")
+		environmentDebug := filepath.Join(t.TempDir(), "environment-debug")
+		flagOutput := filepath.Join(t.TempDir(), "flag-output")
+		flagDebug := filepath.Join(t.TempDir(), "flag-debug")
+		opts := newStateTestHarness().options()
+		opts.LookupEnv = lookupRunContractEnv(map[string]string{
+			"NOTARIUS_OUTPUT_DIR": environmentOutput,
+			"NOTARIUS_DEBUG_DIR":  environmentDebug,
+		})
+		result := runWithStateRoots(t, roots, opts, []string{"--output-dir", flagOutput, "--debug-dir", flagDebug})
+		if result.code != 0 {
+			t.Fatalf("code=%d stdout=%q stderr=%q", result.code, result.stdout, result.stderr)
+		}
+		assertFile(t, filepath.Join(flagOutput, filepath.Base(onlyChildDir(t, flagOutput)), "result.json"))
+		onlyChildDir(t, flagDebug)
+		assertAbsent(t, environmentOutput)
+		assertAbsent(t, environmentDebug)
+	})
+
+	t.Run("built-in roots", func(t *testing.T) {
+		roots := newStateTestRoots(t)
+		data, err := os.ReadFile(roots.config)
+		if err != nil {
+			t.Fatal(err)
+		}
+		text := string(data)
+		text = replaceRequiredOnce(t, text, fmt.Sprintf("  directory: %q\n", roots.output), "")
+		text = replaceRequiredOnce(t, text, fmt.Sprintf("  directory: %q\n", roots.debug), "")
+		if err := os.WriteFile(roots.config, []byte(text), 0o600); err != nil {
+			t.Fatal(err)
+		}
+		workDir := t.TempDir()
+		t.Chdir(workDir)
+		opts := newStateTestHarness().options()
+		result := runWithStateRoots(t, roots, opts, nil)
+		if result.code != 0 {
+			t.Fatalf("code=%d stdout=%q stderr=%q", result.code, result.stdout, result.stderr)
+		}
+		assertFile(t, filepath.Join(workDir, "notarius-output", filepath.Base(onlyChildDir(t, filepath.Join(workDir, "notarius-output"))), "result.json"))
+		onlyChildDir(t, filepath.Join(workDir, "notarius-debug"))
+	})
+}
+
+func TestRunLLMProfileOverrideAndValidationUseInjectedBoundaries(t *testing.T) {
+	t.Run("one effective profile reaches the factory and modules", func(t *testing.T) {
+		roots := newStateTestRoots(t)
+		profileDir := writeRunContractProfiles(t, "override-profile")
+		prependRunContractConfig(t, roots, fmt.Sprintf("promptkit:\n  profile_dir: %q\n", profileDir))
+		harness := newStateTestHarness()
+		var factoryProfiles []string
+		opts := harness.options()
+		var factoryOverrides []LLMRuntimeOverrides
+		opts.LLMClientFactory = func(_ context.Context, _ config.Config, profileID string, overrides LLMRuntimeOverrides) (contracts.StructuredLLMClient, []artifacts.LLMProfileManifest, error) {
+			factoryProfiles = append(factoryProfiles, profileID)
+			factoryOverrides = append(factoryOverrides, overrides)
+			return nil, nil, nil
+		}
+		var stdout, stderr bytes.Buffer
+		code := RunWithOptions([]string{"run", "sample", "--config", roots.config, "--input", roots.input, "--chunk_cache", "bypass", "--llm-profile", "override-profile"}, &stdout, &stderr, opts)
+		if code != 0 || stderr.Len() != 0 {
+			t.Fatalf("code=%d stdout=%q stderr=%q", code, stdout.String(), stderr.String())
+		}
+		if len(factoryProfiles) != 1 || factoryProfiles[0] != "override-profile" {
+			t.Fatalf("factory profiles = %#v, want one override profile", factoryProfiles)
+		}
+		if len(factoryOverrides) != 1 || factoryOverrides[0].ReasoningEffort != nil {
+			t.Fatalf("factory overrides = %#v, want inherited reasoning", factoryOverrides)
+		}
+		harness.mu.Lock()
+		profiles := append([]string(nil), harness.moduleProfiles...)
+		harness.mu.Unlock()
+		if len(profiles) < 4 {
+			t.Fatalf("module profiles = %#v, want chunk and lane stage requests", profiles)
+		}
+		for _, profile := range profiles {
+			if profile != "override-profile" {
+				t.Fatalf("module profiles = %#v, want override on every request", profiles)
+			}
+		}
+	})
+
+	t.Run("runtime override applies to validators", func(t *testing.T) {
+		roots := newStateTestRoots(t)
+		profileDir := writeRunContractProfiles(t, "override-profile", "validator-profile")
+		prependRunContractConfig(t, roots, fmt.Sprintf("promptkit:\n  profile_dir: %q\n", profileDir))
+		harness := newStateTestHarness()
+		var validatorProfiles []string
+		opts := harness.options()
+		registerRunContractValidator(t, &opts, &validatorProfiles)
+		factoryProfiles := []string{}
+		opts.LLMClientFactory = func(_ context.Context, _ config.Config, profileID string, _ LLMRuntimeOverrides) (contracts.StructuredLLMClient, []artifacts.LLMProfileManifest, error) {
+			factoryProfiles = append(factoryProfiles, profileID)
+			return nil, nil, nil
+		}
+		var stdout, stderr bytes.Buffer
+		code := RunWithOptions([]string{"run", "sample", "--config", roots.config, "--input", roots.input, "--chunk_cache", "bypass", "--llm-profile", "override-profile"}, &stdout, &stderr, opts)
+		if code != 0 || stderr.Len() != 0 {
+			t.Fatalf("code=%d stdout=%q stderr=%q", code, stdout.String(), stderr.String())
+		}
+		if len(factoryProfiles) != 1 || factoryProfiles[0] != "override-profile" {
+			t.Fatalf("factory profiles = %#v, want one override profile", factoryProfiles)
+		}
+		if len(validatorProfiles) != 1 || validatorProfiles[0] != "override-profile" {
+			t.Fatalf("validator profiles = %#v, want runtime override", validatorProfiles)
+		}
+	})
+
+	t.Run("unknown profile is rejected without factory access", func(t *testing.T) {
+		roots := newStateTestRoots(t)
+		profileDir := writeRunContractProfiles(t, "override-profile")
+		prependRunContractConfig(t, roots, fmt.Sprintf("promptkit:\n  profile_dir: %q\n", profileDir))
+		factoryCalls := 0
+		opts := newStateTestHarness().options()
+		opts.LLMClientFactory = func(context.Context, config.Config, string, LLMRuntimeOverrides) (contracts.StructuredLLMClient, []artifacts.LLMProfileManifest, error) {
+			factoryCalls++
+			return nil, nil, nil
+		}
+		var stdout, stderr bytes.Buffer
+		code := RunWithOptions([]string{"run", "sample", "--config", roots.config, "--input", roots.input, "--chunk_cache", "bypass", "--llm-profile", "missing-profile"}, &stdout, &stderr, opts)
+		if code != 1 || !strings.Contains(stderr.String(), "not configured") || factoryCalls != 0 || stdout.Len() != 0 {
+			t.Fatalf("code=%d stdout=%q stderr=%q factoryCalls=%d", code, stdout.String(), stderr.String(), factoryCalls)
+		}
+	})
+
+	t.Run("pipeline default is rejected before factory access", func(t *testing.T) {
+		roots := newStateTestRoots(t)
+		profileDir := writeRunContractProfiles(t, "configured-profile")
+		prependRunContractConfig(t, roots, fmt.Sprintf("promptkit:\n  profile_dir: %q\n", profileDir))
+		replaceStateTestConfigLine(t, roots.config, "  sample:\n", "  sample:\n    llm_profile: missing-profile\n")
+		factoryCalls := 0
+		opts := newStateTestHarness().options()
+		opts.LLMClientFactory = func(context.Context, config.Config, string, LLMRuntimeOverrides) (contracts.StructuredLLMClient, []artifacts.LLMProfileManifest, error) {
+			factoryCalls++
+			return nil, nil, nil
+		}
+		var stdout, stderr bytes.Buffer
+		code := RunWithOptions([]string{"run", "sample", "--config", roots.config, "--input", roots.input, "--chunk_cache", "bypass"}, &stdout, &stderr, opts)
+		if code != 1 || !strings.Contains(stderr.String(), "not configured") || factoryCalls != 0 || stdout.Len() != 0 {
+			t.Fatalf("code=%d stdout=%q stderr=%q factoryCalls=%d", code, stdout.String(), stderr.String(), factoryCalls)
+		}
+	})
+}
+
+func TestRunReasoningEffortOverrideReachesFactory(t *testing.T) {
+	tests := []struct {
+		name      string
+		flags     []string
+		wantValue string
+		wantSet   bool
+	}{
+		{name: "inherit"},
+		{name: "replace", flags: []string{"--reasoning-effort", " focused "}, wantValue: "focused", wantSet: true},
+		{name: "clear", flags: []string{"--clear-reasoning-effort"}, wantSet: true},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			roots := newStateTestRoots(t)
+			opts := newStateTestHarness().options()
+			var got []LLMRuntimeOverrides
+			opts.LLMClientFactory = func(_ context.Context, _ config.Config, _ string, overrides LLMRuntimeOverrides) (contracts.StructuredLLMClient, []artifacts.LLMProfileManifest, error) {
+				got = append(got, overrides)
+				return nil, nil, nil
+			}
+			args := append([]string{"run", "sample", "--config", roots.config, "--input", roots.input, "--chunk_cache", "bypass"}, tt.flags...)
+			var stdout, stderr bytes.Buffer
+			if code := RunWithOptions(args, &stdout, &stderr, opts); code != 0 || stderr.Len() != 0 {
+				t.Fatalf("code=%d stdout=%q stderr=%q", code, stdout.String(), stderr.String())
+			}
+			if len(got) != 1 {
+				t.Fatalf("factory overrides = %#v, want one call", got)
+			}
+			if !tt.wantSet {
+				if got[0].ReasoningEffort != nil {
+					t.Fatalf("reasoning effort = %q, want inherit", *got[0].ReasoningEffort)
+				}
+				return
+			}
+			if got[0].ReasoningEffort == nil || *got[0].ReasoningEffort != tt.wantValue {
+				t.Fatalf("reasoning effort = %#v, want %q", got[0].ReasoningEffort, tt.wantValue)
+			}
+		})
+	}
+}
+
+func TestRunReasoningEffortOverrideRejectsInvalidSyntax(t *testing.T) {
+	tests := []struct {
+		name      string
+		flags     []string
+		wantError string
+	}{
+		{
+			name:      "mutually exclusive controls",
+			flags:     []string{"--reasoning-effort", "focused", "--clear-reasoning-effort"},
+			wantError: "cannot be combined",
+		},
+		{
+			name:      "empty replacement",
+			flags:     []string{"--reasoning-effort", "  "},
+			wantError: "must not be empty",
+		},
+		{
+			name:      "duplicate replacement",
+			flags:     []string{"--reasoning-effort", "low", "--reasoning-effort", "high"},
+			wantError: "may be specified only once",
+		},
+		{
+			name:      "missing replacement",
+			flags:     []string{"--reasoning-effort"},
+			wantError: "flag needs an argument",
+		},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			roots := newStateTestRoots(t)
+			args := append([]string{"run", "sample", "--config", roots.config, "--input", roots.input}, tt.flags...)
+			var stdout, stderr bytes.Buffer
+			code := RunWithOptions(args, &stdout, &stderr, newStateTestHarness().options())
+			if code != 2 || stdout.Len() != 0 || !strings.Contains(stderr.String(), tt.wantError) {
+				t.Fatalf("code=%d stdout=%q stderr=%q", code, stdout.String(), stderr.String())
+			}
+			assertNoRunState(t, roots)
+		})
+	}
+}
+
+func TestReasoningEffortOverrideSeparatesCheckpointIdentities(t *testing.T) {
+	replacement := " focused "
+	cleared := ""
+	states := []struct {
+		name      string
+		overrides LLMRuntimeOverrides
+		wantValue string
+		wantSet   bool
+	}{
+		{name: "inherit"},
+		{name: "replace", overrides: LLMRuntimeOverrides{ReasoningEffort: &replacement}, wantValue: "focused", wantSet: true},
+		{name: "clear", overrides: LLMRuntimeOverrides{ReasoningEffort: &cleared}, wantValue: "<cleared>", wantSet: true},
+	}
+	digests := make(map[string]string, len(states))
+	for _, state := range states {
+		fingerprints := runtimeOverrideFingerprints("", "", state.overrides)
+		var value string
+		var found bool
+		for _, fingerprint := range fingerprints {
+			if fingerprint.Name == "reasoning_effort_override" {
+				value, found = fingerprint.Value, true
+			}
+		}
+		if found != state.wantSet || (found && value != state.wantValue) {
+			t.Fatalf("%s fingerprint found=%t value=%q, want found=%t value=%q", state.name, found, value, state.wantSet, state.wantValue)
+		}
+		identity, err := checkpoint.NewIdentity(checkpoint.IdentityInput{
+			Pipeline:         pipeline.ResolvedPipeline{ID: "sample", Digest: "sha256:pipeline", Input: pipeline.Binding("test/input")},
+			RawInputDigest:   "sha256:input",
+			RuntimeOverrides: fingerprints,
+		})
+		if err != nil {
+			t.Fatal(err)
+		}
+		digests[state.name] = identity.Digest
+	}
+	if digests["inherit"] == digests["replace"] || digests["inherit"] == digests["clear"] || digests["replace"] == digests["clear"] {
+		t.Fatalf("checkpoint identity digests are not distinct: %#v", digests)
+	}
+}
+
+func TestEffectiveLLMProfileIDsAreSortedDeduplicatedAndLLMOnly(t *testing.T) {
+	resolved := pipeline.ResolvedPipeline{
+		Input:               pipeline.ModuleBinding{LLMProfile: "input-profile"},
+		InputExecutionClass: contracts.ExecutionClassLLMBacked,
+		Chunk:               pipeline.ModuleBinding{LLMProfile: " zeta "},
+		ChunkExecutionClass: contracts.ExecutionClassLLMBacked,
+		Steps: []pipeline.ResolvedPipelineStep{{
+			ID: "default",
+			ArtifactLanes: []pipeline.ResolvedArtifactLane{{
+				Extract:                 pipeline.ModuleBinding{LLMProfile: "alpha"},
+				ExtractExecutionClass:   contracts.ExecutionClassLLMBacked,
+				Merge:                   pipeline.ModuleBinding{LLMProfile: "deterministic-merge"},
+				MergeExecutionClass:     contracts.ExecutionClassDeterministic,
+				Normalize:               pipeline.ModuleBinding{LLMProfile: " gamma "},
+				NormalizeExecutionClass: contracts.ExecutionClassLLMBacked,
+			}},
+		}},
+		ValidatorChains: []pipeline.ResolvedValidatorChain{{Validators: []pipeline.ResolvedValidator{
+			{Binding: pipeline.ModuleBinding{LLMProfile: "deterministic-profile"}, ExecutionClass: contracts.ExecutionClassDeterministic},
+			{Binding: pipeline.ModuleBinding{LLMProfile: "beta"}, ExecutionClass: contracts.ExecutionClassLLMBacked},
+		}}},
+		Output:               pipeline.ModuleBinding{LLMProfile: "output-profile"},
+		OutputExecutionClass: contracts.ExecutionClassLLMBacked,
+	}
+	got := effectiveLLMProfileIDs(resolved)
+	want := []string{"alpha", "beta", "gamma", "input-profile", "output-profile", "zeta"}
+	if strings.Join(got, ",") != strings.Join(want, ",") {
+		t.Fatalf("effective profiles = %#v, want %#v", got, want)
+	}
+}
+
+func TestRunSessionIDUsesExplicitValueOrSourceDocumentID(t *testing.T) {
+	for _, tt := range []struct {
+		name string
+		args []string
+		want string
+	}{
+		{name: "source document", want: "source"},
+		{name: "explicit trimmed value", args: []string{"--session-id", " explicit-session "}, want: "explicit-session"},
+	} {
+		t.Run(tt.name, func(t *testing.T) {
+			roots := newStateTestRoots(t)
+			harness := newStateTestHarness()
+			args := append([]string{"run", "sample", "--config", roots.config, "--input", roots.input, "--chunk_cache", "bypass"}, tt.args...)
+			var stdout, stderr bytes.Buffer
+			code := RunWithOptions(args, &stdout, &stderr, harness.options())
+			if code != 0 || stderr.Len() != 0 {
+				t.Fatalf("code=%d stdout=%q stderr=%q", code, stdout.String(), stderr.String())
+			}
+			harness.mu.Lock()
+			sessions := append([]string(nil), harness.sessionIDs...)
+			harness.mu.Unlock()
+			if len(sessions) < 4 {
+				t.Fatalf("session IDs = %#v, want all prompt-facing module requests", sessions)
+			}
+			for _, session := range sessions {
+				if session != tt.want {
+					t.Fatalf("session IDs = %#v, want %q", sessions, tt.want)
+				}
+			}
+		})
+	}
+}
+
+func TestRunFactoryAndPreparationFailuresAreProcessFailures(t *testing.T) {
+	t.Run("LLM factory", func(t *testing.T) {
+		roots := newStateTestRoots(t)
+		opts := newStateTestHarness().options()
+		opts.LLMClientFactory = func(context.Context, config.Config, string, LLMRuntimeOverrides) (contracts.StructuredLLMClient, []artifacts.LLMProfileManifest, error) {
+			return nil, nil, errors.New("injected LLM factory failure")
+		}
+		var stdout, stderr bytes.Buffer
+		code := RunWithOptions([]string{"run", "sample", "--config", roots.config, "--input", roots.input, "--chunk_cache", "bypass"}, &stdout, &stderr, opts)
+		if code != 1 || !strings.Contains(stderr.String(), "injected LLM factory failure") || stdout.Len() != 0 {
+			t.Fatalf("code=%d stdout=%q stderr=%q", code, stdout.String(), stderr.String())
+		}
+	})
+
+	t.Run("pipeline preparation", func(t *testing.T) {
+		roots := newStateTestRoots(t)
+		data, err := os.ReadFile(roots.config)
+		if err != nil {
+			t.Fatal(err)
+		}
+		data = []byte(replaceRequiredOnce(t, string(data), "extract: test/extract", "extract: test/failing-extract"))
+		if err := os.WriteFile(roots.config, data, 0o600); err != nil {
+			t.Fatal(err)
+		}
+		opts := newStateTestHarness().options()
+		if err := pipeline.RegisterExtractorBuilder(opts.Registries.Extractors, pipeline.ModuleSpec{Key: "test/failing-extract", Stage: pipeline.StageExtract, ExecutionClass: contracts.ExecutionClassDeterministic, Requires: []string{"chunks"}, Provides: []string{"artifact"}, ArtifactKind: stateTestArtifactKind}, func(map[string]any) error { return nil }, func(pipeline.BuildRequest) (contracts.Extractor[stateTestArtifact], error) {
+			return nil, errors.New("injected extractor construction failure")
+		}); err != nil {
+			t.Fatal(err)
+		}
+		opts.Catalog = catalogFromRegistries(opts.Registries)
+		var stdout, stderr bytes.Buffer
+		code := RunWithOptions([]string{"run", "sample", "--config", roots.config, "--input", roots.input, "--chunk_cache", "bypass"}, &stdout, &stderr, opts)
+		if code != 1 || !strings.Contains(stderr.String(), "injected extractor construction failure") || stdout.Len() != 0 {
+			t.Fatalf("code=%d stdout=%q stderr=%q", code, stdout.String(), stderr.String())
+		}
+	})
+}
+
+func TestRunWarningsRemainSuccessfulAndReachDurableSurfaces(t *testing.T) {
+	roots := newStateTestRoots(t)
+	harness := newStateTestHarness()
+	harness.includeWarnings = true
+	harness.chunkWarnings = []contracts.Warning{{Scope: "chunk", ReasonCode: "contract-warning", Message: "warning retained"}}
+	var stdout, stderr bytes.Buffer
+	code := RunWithOptions([]string{"run", "sample", "--config", roots.config, "--input", roots.input, "--chunk_cache", "bypass", "--debug"}, &stdout, &stderr, harness.options())
+	if code != 0 || !strings.Contains(stdout.String(), "outputs=1") || !strings.Contains(stderr.String(), "1 warning(s)") {
+		t.Fatalf("code=%d stdout=%q stderr=%q", code, stdout.String(), stderr.String())
+	}
+	outputPath := filepath.Join(onlyChildDir(t, roots.output), "result.json")
+	output, err := os.ReadFile(outputPath)
+	if err != nil || !strings.Contains(string(output), "contract-warning") {
+		t.Fatalf("durable output = %q, %v", output, err)
+	}
+	bundle := onlyChildDir(t, roots.debug)
+	var warnings []contracts.Warning
+	readStateTestSummaryJSON(t, bundle, "warnings.json", &warnings)
+	if len(warnings) != 1 || warnings[0].ReasonCode != "contract-warning" {
+		t.Fatalf("debug warnings = %#v", warnings)
+	}
+}
+
+func runWithStateRoots(t *testing.T, roots stateTestRoots, opts Options, extra []string) stateTestResult {
+	t.Helper()
+	args := []string{"run", "sample", "--config", roots.config, "--input", roots.input, "--chunk_cache", "bypass", "--debug"}
+	args = append(args, extra...)
+	var stdout, stderr bytes.Buffer
+	return stateTestResult{code: RunWithOptions(args, &stdout, &stderr, opts), stdout: stdout.String(), stderr: stderr.String()}
+}
+
+func lookupRunContractEnv(values map[string]string) func(string) (string, bool) {
+	return func(name string) (string, bool) {
+		value, ok := values[name]
+		return value, ok
+	}
+}
+
+func prependRunContractConfig(t *testing.T, roots stateTestRoots, prefix string) {
+	t.Helper()
+	data, err := os.ReadFile(roots.config)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if err := os.WriteFile(roots.config, append([]byte(prefix), data...), 0o600); err != nil {
+		t.Fatal(err)
+	}
+}
+
+func writeRunContractProfiles(t *testing.T, ids ...string) string {
+	t.Helper()
+	dir := t.TempDir()
+	for _, id := range ids {
+		profile := fmt.Sprintf("id: %s\nendpoint: http://127.0.0.1:1/v1\nmodel: %s-model\n", id, id)
+		if err := os.WriteFile(filepath.Join(dir, id+".yaml"), []byte(profile), 0o600); err != nil {
+			t.Fatal(err)
+		}
+	}
+	return dir
+}
+
+func registerRunContractValidator(t *testing.T, opts *Options, profiles *[]string) {
+	t.Helper()
+	if err := pipeline.RegisterTypedValidatorBuilder(opts.Registries.Validators, stateTestArtifactKind, pipeline.ValidatorSpec{Key: "run-contract-validator", ExecutionClass: contracts.ExecutionClassLLMBacked}, func(map[string]any) error { return nil }, func(pipeline.BuildRequest) (contracts.TypedValidator[stateTestArtifact], error) {
+		return runContractValidator{profiles: profiles}, nil
+	}); err != nil {
+		t.Fatal(err)
+	}
+	if err := opts.Registries.ValidatorChains.Register(pipeline.ValidatorChainMapping{Stage: pipeline.StageExtract, Module: "test/extract", Validators: []pipeline.ModuleBinding{{Module: "run-contract-validator", LLMProfile: "validator-profile"}}}); err != nil {
+		t.Fatal(err)
+	}
+	opts.Catalog = catalogFromRegistries(opts.Registries)
+}
+
+type runContractValidator struct {
+	profiles *[]string
+}
+
+func (v runContractValidator) Name() string { return "run-contract-validator" }
+
+func (v runContractValidator) ExecutionClass() contracts.ExecutionClass {
+	return contracts.ExecutionClassLLMBacked
+}
+
+func (v runContractValidator) Validate(_ context.Context, req contracts.TypedValidationRequest[stateTestArtifact]) (contracts.ValidationResult, error) {
+	*v.profiles = append(*v.profiles, req.LLMProfile)
+	return contracts.ValidationResult{Approved: true}, nil
+}

internal/cli/run_id.go

@@ -0,0 +1,34 @@
+package cli
+
+import (
+	"crypto/rand"
+	"encoding/hex"
+	"fmt"
+	"io"
+	"path/filepath"
+	"strings"
+	"time"
+)
+
+type RunIDGenerator func(time.Time) (string, error)
+
+func defaultRunIDGenerator(startedAt time.Time) (string, error) {
+	var suffix [16]byte
+	if _, err := io.ReadFull(rand.Reader, suffix[:]); err != nil {
+		return "", fmt.Errorf("read random run ID suffix: %w", err)
+	}
+	return fmt.Sprintf("run-%d-%s", startedAt.UnixNano(), hex.EncodeToString(suffix[:])), nil
+}
+
+func validateRunID(runID string) error {
+	if runID == "" {
+		return fmt.Errorf("run ID must not be empty")
+	}
+	if runID != strings.TrimSpace(runID) {
+		return fmt.Errorf("run ID %q must not have surrounding whitespace", runID)
+	}
+	if strings.ContainsAny(runID, `/\\`) || filepath.IsAbs(runID) || filepath.Clean(runID) != runID || runID == "." || runID == ".." {
+		return fmt.Errorf("run ID %q must be one safe path component", runID)
+	}
+	return nil
+}

internal/cli/run_id_test.go

@@ -0,0 +1,87 @@
+package cli
+
+import (
+	"os"
+	"path/filepath"
+	"regexp"
+	"strings"
+	"testing"
+	"time"
+
+	"gitea.maximumdirect.net/eric/notarius/internal/framework/contracts"
+)
+
+func TestDefaultRunIDGeneratorProducesUniqueSafeIDs(t *testing.T) {
+	startedAt := time.Unix(0, 123456789).UTC()
+	pattern := regexp.MustCompile(`^run-123456789-[0-9a-f]{32}$`)
+	seen := make(map[string]struct{}, 256)
+	for i := 0; i < 256; i++ {
+		runID, err := defaultRunIDGenerator(startedAt)
+		if err != nil {
+			t.Fatal(err)
+		}
+		if !pattern.MatchString(runID) {
+			t.Fatalf("run ID %q does not match production format", runID)
+		}
+		if err := validateRunID(runID); err != nil {
+			t.Fatalf("run ID %q is not path-safe: %v", runID, err)
+		}
+		if _, exists := seen[runID]; exists {
+			t.Fatalf("duplicate run ID %q", runID)
+		}
+		seen[runID] = struct{}{}
+	}
+}
+
+func TestWriteOutputFilesSupportsNestedLogicalPaths(t *testing.T) {
+	runPath := filepath.Join(t.TempDir(), "output", "run-safe")
+	if err := writeOutputFiles(runPath, []contracts.OutputFile{{Name: "nested/result.json", Bytes: []byte("result")}}); err != nil {
+		t.Fatal(err)
+	}
+	data, err := os.ReadFile(filepath.Join(runPath, "nested", "result.json"))
+	if err != nil || string(data) != "result" {
+		t.Fatalf("nested output = %q, %v", data, err)
+	}
+}
+
+func TestWriteOutputFilesRejectsUnsafeNamesBeforeAllocatingRunDirectory(t *testing.T) {
+	outputRoot := filepath.Join(t.TempDir(), "output")
+	runPath := filepath.Join(outputRoot, "run-safe")
+	for _, name := range []string{"", "../outside", "/absolute", `nested\\outside`, "nested/../outside"} {
+		t.Run(name, func(t *testing.T) {
+			if err := writeOutputFiles(runPath, []contracts.OutputFile{{Name: "safe.json"}, {Name: name}}); err == nil {
+				t.Fatalf("writeOutputFiles accepted %q", name)
+			}
+			if _, err := os.Stat(outputRoot); !os.IsNotExist(err) {
+				t.Fatalf("output root exists or stat failed after %q: %v", name, err)
+			}
+		})
+	}
+}
+
+func TestWriteOutputFilesRetainsNewPartialDirectoryAndPreservesSibling(t *testing.T) {
+	outputRoot := filepath.Join(t.TempDir(), "output")
+	siblingPath := filepath.Join(outputRoot, "sibling")
+	if err := os.MkdirAll(siblingPath, 0o755); err != nil {
+		t.Fatal(err)
+	}
+	sentinelPath := filepath.Join(siblingPath, "sentinel")
+	if err := os.WriteFile(sentinelPath, []byte("preserve sibling"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+
+	runPath := filepath.Join(outputRoot, "run-safe")
+	err := writeOutputFiles(runPath, []contracts.OutputFile{
+		{Name: "blocked", Bytes: []byte("partial output")},
+		{Name: "blocked/nested.json", Bytes: []byte("unreachable")},
+	})
+	if err == nil || !strings.Contains(err.Error(), "create output directory") {
+		t.Fatalf("writeOutputFiles() error = %v, want later directory failure", err)
+	}
+	if got, err := os.ReadFile(filepath.Join(runPath, "blocked")); err != nil || string(got) != "partial output" {
+		t.Fatalf("partial output = %q, %v", got, err)
+	}
+	if got, err := os.ReadFile(sentinelPath); err != nil || string(got) != "preserve sibling" {
+		t.Fatalf("sibling sentinel = %q, %v", got, err)
+	}
+}

internal/cli/run_result.go

@@ -0,0 +1,111 @@
+package cli
+
+import (
+	"encoding/json"
+	"fmt"
+	"io"
+	"path/filepath"
+	"strings"
+
+	"gitea.maximumdirect.net/eric/notarius/internal/framework/pipeline"
+)
+
+const runResultSchemaVersion = "notarius.run-result.v1"
+
+type runResult struct {
+	SchemaVersion         string `json:"schema_version"`
+	RunID                 string `json:"run_id"`
+	PipelineID            string `json:"pipeline_id"`
+	OutputDirectory       string `json:"output_directory"`
+	IndexFile             string `json:"index_file,omitempty"`
+	NormalizedOutputCount int    `json:"normalized_output_count"`
+	RejectedOutputCount   int    `json:"rejected_output_count"`
+	WarningCount          int    `json:"warning_count"`
+	ValidationStatus      string `json:"validation_status"`
+	DebugDirectory        string `json:"debug_directory,omitempty"`
+}
+
+func newRunResult(resolved pipeline.ResolvedPipeline, output pipeline.RunOutput, outputDirectory, debugDirectory string) (runResult, error) {
+	if strings.TrimSpace(output.Manifest.RunID) == "" {
+		return runResult{}, fmt.Errorf("run result requires a run ID")
+	}
+	if strings.TrimSpace(resolved.ID) == "" {
+		return runResult{}, fmt.Errorf("run result requires a resolved pipeline ID")
+	}
+	if strings.TrimSpace(output.Manifest.PipelineID) == "" {
+		return runResult{}, fmt.Errorf("run result requires a manifest pipeline ID")
+	}
+	if output.Manifest.PipelineID != resolved.ID {
+		return runResult{}, fmt.Errorf("run result pipeline ID does not match resolved pipeline")
+	}
+	if strings.TrimSpace(output.Manifest.ValidationStatus) == "" {
+		return runResult{}, fmt.Errorf("run result requires a validation status")
+	}
+	if strings.TrimSpace(outputDirectory) == "" {
+		return runResult{}, fmt.Errorf("run result requires an output directory")
+	}
+
+	absOutputDirectory, err := filepath.Abs(outputDirectory)
+	if err != nil {
+		return runResult{}, fmt.Errorf("make output directory absolute: %w", err)
+	}
+
+	result := runResult{
+		SchemaVersion:         runResultSchemaVersion,
+		RunID:                 output.Manifest.RunID,
+		PipelineID:            output.Manifest.PipelineID,
+		OutputDirectory:       absOutputDirectory,
+		NormalizedOutputCount: len(output.NormalizeOutputs),
+		RejectedOutputCount:   len(output.Rejected),
+		WarningCount:          len(output.Warnings),
+		ValidationStatus:      output.Manifest.ValidationStatus,
+	}
+
+	if strings.TrimSpace(debugDirectory) != "" {
+		absDebugDirectory, err := filepath.Abs(debugDirectory)
+		if err != nil {
+			return runResult{}, fmt.Errorf("make debug directory absolute: %w", err)
+		}
+		result.DebugDirectory = absDebugDirectory
+	}
+
+	if resolved.Output.Module == pipeline.DefaultOutputModule {
+		indexCount := 0
+		for _, file := range output.OutputFiles {
+			if file.Name == "index.json" {
+				indexCount++
+			}
+		}
+		if indexCount != 1 {
+			return runResult{}, fmt.Errorf("production JSON output must contain exactly one index.json file")
+		}
+		result.IndexFile = "index.json"
+	}
+
+	return result, nil
+}
+
+func encodeRunResult(result runResult) ([]byte, error) {
+	encoded, err := json.Marshal(result)
+	if err != nil {
+		return nil, fmt.Errorf("encode run result: %w", err)
+	}
+	return append(encoded, '\n'), nil
+}
+
+func writeRunResult(writer io.Writer, content []byte) error {
+	for len(content) > 0 {
+		written, err := writer.Write(content)
+		if written < 0 || written > len(content) {
+			return io.ErrShortWrite
+		}
+		content = content[written:]
+		if err != nil {
+			return err
+		}
+		if written == 0 {
+			return io.ErrShortWrite
+		}
+	}
+	return nil
+}

internal/cli/run_result_command_test.go

@@ -0,0 +1,199 @@
+package cli
+
+import (
+	"bytes"
+	"encoding/json"
+	"errors"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+
+	"gitea.maximumdirect.net/eric/notarius/internal/framework/contracts"
+	alwaysreject "gitea.maximumdirect.net/eric/notarius/internal/modules/generic/validate/always_reject"
+)
+
+func TestMaintainedMinimalInvocationEmitsRunResult(t *testing.T) {
+	outputRoot := filepath.Join(t.TempDir(), "output")
+	var stdout, stderr strings.Builder
+	code := RunWithOptions([]string{
+		"run", "dnd-session",
+		"--config", repositoryPath("examples", "dnd-minimal.config.yml"),
+		"--input", repositoryPath("examples", "seriatim-minimal-transcript.json"),
+		"--only", "spells", "--chunk_cache", "bypass", "--output-dir", outputRoot, "--json",
+	}, &stdout, &stderr, productionRunOptions(t, &productionFakeLLMClient{}))
+	if code != 0 || stderr.Len() != 0 {
+		t.Fatalf("code=%d stdout=%q stderr=%q", code, stdout.String(), stderr.String())
+	}
+
+	receipt := decodeRunResultDocument(t, stdout.String())
+	if got := receipt["schema_version"]; got != "notarius.run-result.v1" {
+		t.Fatalf("schema_version = %q", got)
+	}
+	if got := receipt["run_id"]; got != productionRunID {
+		t.Fatalf("run_id = %q", got)
+	}
+	if got := receipt["pipeline_id"]; got != "dnd-session" {
+		t.Fatalf("pipeline_id = %q", got)
+	}
+	if got := receipt["index_file"]; got != "index.json" {
+		t.Fatalf("index_file = %q", got)
+	}
+	if got := receipt["normalized_output_count"]; got != float64(1) {
+		t.Fatalf("normalized_output_count = %v", got)
+	}
+	if got := receipt["rejected_output_count"]; got != float64(0) {
+		t.Fatalf("rejected_output_count = %v", got)
+	}
+	if got := receipt["warning_count"]; got != float64(0) {
+		t.Fatalf("warning_count = %v", got)
+	}
+	if got := receipt["validation_status"]; got != "approved" {
+		t.Fatalf("validation_status = %q", got)
+	}
+
+	outputDirectory, ok := receipt["output_directory"].(string)
+	if !ok || !filepath.IsAbs(outputDirectory) || outputDirectory != filepath.Join(outputRoot, productionRunID) {
+		t.Fatalf("output_directory = %q", receipt["output_directory"])
+	}
+	indexFile := receipt["index_file"].(string)
+	assertFile(t, filepath.Join(outputDirectory, indexFile))
+}
+
+func TestRunResultReportsWarningsAndDebugBundle(t *testing.T) {
+	roots := newStateTestRoots(t)
+	harness := newStateTestHarness()
+	harness.chunkWarnings = []contracts.Warning{{Scope: "chunk", ReasonCode: "contract-warning", Message: "warning retained"}}
+	var stdout, stderr bytes.Buffer
+	code := RunWithOptions([]string{
+		"run", "sample", "--config", roots.config, "--input", roots.input,
+		"--chunk_cache", "bypass", "--debug", "--json",
+	}, &stdout, &stderr, harness.options())
+	if code != 0 || !strings.Contains(stderr.String(), "1 warning(s)") {
+		t.Fatalf("code=%d stdout=%q stderr=%q", code, stdout.String(), stderr.String())
+	}
+
+	receipt := decodeRunResultDocument(t, stdout.String())
+	if got := receipt["warning_count"]; got != float64(1) {
+		t.Fatalf("warning_count = %v", got)
+	}
+	debugDirectory, ok := receipt["debug_directory"].(string)
+	if !ok || !filepath.IsAbs(debugDirectory) || debugDirectory != onlyChildDir(t, roots.debug) {
+		t.Fatalf("debug_directory = %q", receipt["debug_directory"])
+	}
+	if strings.Contains(stdout.String(), "complete:") || strings.Contains(stdout.String(), "debug=") {
+		t.Fatalf("machine stdout contains human reporting: %q", stdout.String())
+	}
+}
+
+func TestRunResultReportsSuccessfulRejection(t *testing.T) {
+	roots := newStateTestRoots(t)
+	configBytes, err := os.ReadFile(roots.config)
+	if err != nil {
+		t.Fatal(err)
+	}
+	configBytes = []byte(replaceRequiredOnce(t, string(configBytes), "        normalize: test/normalize\n", "        normalize:\n          module: test/normalize\n          validators:\n            - generic/always_reject\n"))
+	if err := os.WriteFile(roots.config, configBytes, 0o600); err != nil {
+		t.Fatal(err)
+	}
+
+	harness := newStateTestHarness()
+	opts := harness.options()
+	if err := alwaysreject.RegisterTyped[stateTestArtifact](opts.Registries.Validators, stateTestArtifactKind); err != nil {
+		t.Fatal(err)
+	}
+	opts.Catalog = catalogFromRegistries(opts.Registries)
+
+	var stdout, stderr bytes.Buffer
+	code := RunWithOptions([]string{
+		"run", "sample", "--config", roots.config, "--input", roots.input,
+		"--chunk_cache", "bypass", "--json",
+	}, &stdout, &stderr, opts)
+	if code != 0 || stderr.Len() != 0 {
+		t.Fatalf("code=%d stdout=%q stderr=%q", code, stdout.String(), stderr.String())
+	}
+	receipt := decodeRunResultDocument(t, stdout.String())
+	if got := receipt["normalized_output_count"]; got != float64(0) {
+		t.Fatalf("normalized_output_count = %v", got)
+	}
+	if got := receipt["rejected_output_count"]; got != float64(1) {
+		t.Fatalf("rejected_output_count = %v", got)
+	}
+	if got := receipt["validation_status"]; got != "rejected" {
+		t.Fatalf("validation_status = %q", got)
+	}
+}
+
+func TestRunResultIsAbsentForSyntaxAndRuntimeFailures(t *testing.T) {
+	t.Run("syntax", func(t *testing.T) {
+		var stdout, stderr bytes.Buffer
+		code := RunWithOptions([]string{"run", "sample", "--json"}, &stdout, &stderr, newStateTestHarness().options())
+		if code != 2 || stdout.Len() != 0 || stderr.Len() == 0 {
+			t.Fatalf("code=%d stdout=%q stderr=%q", code, stdout.String(), stderr.String())
+		}
+	})
+
+	t.Run("runtime", func(t *testing.T) {
+		roots := newStateTestRoots(t)
+		harness := newStateTestHarness()
+		harness.extractErr = errors.New("injected extraction failure")
+		var stdout, stderr bytes.Buffer
+		code := RunWithOptions([]string{
+			"run", "sample", "--config", roots.config, "--input", roots.input,
+			"--chunk_cache", "bypass", "--json",
+		}, &stdout, &stderr, harness.options())
+		if code != 1 || stdout.Len() != 0 || stderr.Len() == 0 {
+			t.Fatalf("code=%d stdout=%q stderr=%q", code, stdout.String(), stderr.String())
+		}
+	})
+}
+
+func TestRunResultDeliveryFailureRetainsPublishedBundles(t *testing.T) {
+	roots := newStateTestRoots(t)
+	writerErr := errors.New("result writer sentinel")
+	stdout := &resultDeliveryWriter{err: writerErr}
+	var stderr bytes.Buffer
+	code := RunWithOptions([]string{
+		"run", "sample", "--config", roots.config, "--input", roots.input,
+		"--chunk_cache", "bypass", "--debug", "--json",
+	}, stdout, &stderr, newStateTestHarness().options())
+	if code != 1 || !strings.Contains(stderr.String(), "write run result") || strings.Contains(stderr.String(), writerErr.Error()) {
+		t.Fatalf("code=%d stderr=%q", code, stderr.String())
+	}
+	if stdout.accepted.Len() != 0 {
+		t.Fatalf("accepted stdout = %q", stdout.accepted.String())
+	}
+	assertStateTestOutput(t, roots.output)
+	debugBundle := onlyChildDir(t, roots.debug)
+	report := readStateTestRunReport(t, debugBundle)
+	if !report.Succeeded {
+		t.Fatalf("debug report = %#v, want successful persisted run", report)
+	}
+	if strings.Contains(readAllFiles(t, debugBundle), writerErr.Error()) {
+		t.Fatalf("debug bundle contains result writer error")
+	}
+}
+
+func decodeRunResultDocument(t *testing.T, stdout string) map[string]any {
+	t.Helper()
+	if strings.Count(stdout, "\n") != 1 {
+		t.Fatalf("stdout = %q, want one JSON document", stdout)
+	}
+	var receipt map[string]any
+	if err := json.Unmarshal([]byte(stdout), &receipt); err != nil {
+		t.Fatalf("decode run result: %v; stdout=%q", err, stdout)
+	}
+	return receipt
+}
+
+type resultDeliveryWriter struct {
+	err      error
+	accepted bytes.Buffer
+}
+
+func (w *resultDeliveryWriter) Write(content []byte) (int, error) {
+	if w.err != nil {
+		return 0, w.err
+	}
+	return w.accepted.Write(content)
+}

internal/cli/run_result_test.go

@@ -0,0 +1,200 @@
+package cli
+
+import (
+	"bytes"
+	"encoding/json"
+	"errors"
+	"io"
+	"path/filepath"
+	"testing"
+
+	"gitea.maximumdirect.net/eric/notarius/internal/core/artifacts"
+	"gitea.maximumdirect.net/eric/notarius/internal/framework/contracts"
+	"gitea.maximumdirect.net/eric/notarius/internal/framework/pipeline"
+)
+
+func TestRunResultEncodesRequiredFieldsAndCounts(t *testing.T) {
+	result, err := newRunResult(testResolvedPipeline(pipeline.DefaultOutputModule), testRunOutput(), "relative-output", "relative-debug")
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	encoded, err := encodeRunResult(result)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if encoded[len(encoded)-1] != '\n' || bytes.Count(encoded, []byte{'\n'}) != 1 {
+		t.Fatalf("encoded result is not one newline-terminated object: %q", encoded)
+	}
+
+	var decoded map[string]any
+	if err := json.Unmarshal(encoded, &decoded); err != nil {
+		t.Fatal(err)
+	}
+	if got := decoded["schema_version"]; got != runResultSchemaVersion {
+		t.Fatalf("schema_version = %q", got)
+	}
+	if got := decoded["run_id"]; got != "run-123" {
+		t.Fatalf("run_id = %q", got)
+	}
+	if got := decoded["pipeline_id"]; got != "sample" {
+		t.Fatalf("pipeline_id = %q", got)
+	}
+	if got := decoded["validation_status"]; got != "rejected" {
+		t.Fatalf("validation_status = %q", got)
+	}
+	if got := decoded["index_file"]; got != "index.json" {
+		t.Fatalf("index_file = %q", got)
+	}
+	if got := decoded["normalized_output_count"]; got != float64(2) {
+		t.Fatalf("normalized_output_count = %v", got)
+	}
+	if got := decoded["rejected_output_count"]; got != float64(1) {
+		t.Fatalf("rejected_output_count = %v", got)
+	}
+	if got := decoded["warning_count"]; got != float64(1) {
+		t.Fatalf("warning_count = %v", got)
+	}
+	if got := decoded["output_directory"]; got != filepath.Join(mustWorkingDirectory(t), "relative-output") {
+		t.Fatalf("output_directory = %q", got)
+	}
+	if got := decoded["debug_directory"]; got != filepath.Join(mustWorkingDirectory(t), "relative-debug") {
+		t.Fatalf("debug_directory = %q", got)
+	}
+}
+
+func TestRunResultRejectsInvalidRequiredValues(t *testing.T) {
+	tests := []struct {
+		name      string
+		resolved  pipeline.ResolvedPipeline
+		output    pipeline.RunOutput
+		directory string
+	}{
+		{name: "blank run ID", resolved: testResolvedPipeline(pipeline.DefaultOutputModule), output: testRunOutputWithout(func(output *pipeline.RunOutput) { output.Manifest.RunID = " " }), directory: "output"},
+		{name: "blank resolved pipeline ID", resolved: pipeline.ResolvedPipeline{Output: pipeline.ModuleBinding{Module: pipeline.DefaultOutputModule}}, output: testRunOutput(), directory: "output"},
+		{name: "blank manifest pipeline ID", resolved: testResolvedPipeline(pipeline.DefaultOutputModule), output: testRunOutputWithout(func(output *pipeline.RunOutput) { output.Manifest.PipelineID = "" }), directory: "output"},
+		{name: "mismatched pipeline IDs", resolved: testResolvedPipeline(pipeline.DefaultOutputModule), output: testRunOutputWithout(func(output *pipeline.RunOutput) { output.Manifest.PipelineID = "other" }), directory: "output"},
+		{name: "blank validation status", resolved: testResolvedPipeline(pipeline.DefaultOutputModule), output: testRunOutputWithout(func(output *pipeline.RunOutput) { output.Manifest.ValidationStatus = " " }), directory: "output"},
+		{name: "blank output directory", resolved: testResolvedPipeline(pipeline.DefaultOutputModule), output: testRunOutput(), directory: " "},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			if _, err := newRunResult(tt.resolved, tt.output, tt.directory, ""); err == nil {
+				t.Fatal("newRunResult() succeeded")
+			}
+		})
+	}
+}
+
+func TestRunResultOmitsIndexFileForOtherOutputModules(t *testing.T) {
+	result, err := newRunResult(testResolvedPipeline("test/output"), testRunOutputWithout(func(output *pipeline.RunOutput) {
+		output.OutputFiles = nil
+	}), "output", "")
+	if err != nil {
+		t.Fatal(err)
+	}
+	if result.IndexFile != "" {
+		t.Fatalf("index_file = %q", result.IndexFile)
+	}
+	encoded, err := encodeRunResult(result)
+	if err != nil {
+		t.Fatal(err)
+	}
+	var decoded map[string]any
+	if err := json.Unmarshal(encoded, &decoded); err != nil {
+		t.Fatal(err)
+	}
+	if _, ok := decoded["index_file"]; ok {
+		t.Fatalf("encoded non-JSON result contains index_file: %s", encoded)
+	}
+	if _, ok := decoded["debug_directory"]; ok {
+		t.Fatalf("encoded result without debug capture contains debug_directory: %s", encoded)
+	}
+}
+
+func TestRunResultRequiresOneProductionIndexFile(t *testing.T) {
+	tests := []struct {
+		name  string
+		files []contracts.OutputFile
+	}{
+		{name: "missing", files: nil},
+		{name: "duplicate", files: []contracts.OutputFile{{Name: "index.json"}, {Name: "index.json"}}},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			output := testRunOutput()
+			output.OutputFiles = tt.files
+			if _, err := newRunResult(testResolvedPipeline(pipeline.DefaultOutputModule), output, "output", ""); err == nil {
+				t.Fatal("newRunResult() succeeded")
+			}
+		})
+	}
+}
+
+func TestWriteRunResultCompletesAndReportsWriterFailure(t *testing.T) {
+	content := []byte("result\n")
+	var target bytes.Buffer
+	if err := writeRunResult(partialResultWriter{writer: &target, limit: 2}, content); err != nil {
+		t.Fatal(err)
+	}
+	if got := target.String(); got != string(content) {
+		t.Fatalf("written result = %q", got)
+	}
+
+	writerErr := errors.New("result writer failed")
+	if err := writeRunResult(failingResultWriter{err: writerErr}, content); !errors.Is(err, writerErr) {
+		t.Fatalf("writeRunResult() error = %v", err)
+	}
+	if err := writeRunResult(zeroResultWriter{}, content); !errors.Is(err, io.ErrShortWrite) {
+		t.Fatalf("zero-progress error = %v", err)
+	}
+}
+
+func testResolvedPipeline(outputModule string) pipeline.ResolvedPipeline {
+	return pipeline.ResolvedPipeline{ID: "sample", Output: pipeline.ModuleBinding{Module: outputModule}}
+}
+
+func testRunOutput() pipeline.RunOutput {
+	return pipeline.RunOutput{
+		Manifest:         artifacts.RunManifest{RunID: "run-123", PipelineID: "sample", ValidationStatus: "rejected"},
+		NormalizeOutputs: []contracts.SerializedOutput{{}, {}},
+		Rejected:         []contracts.RejectedOutput{{}},
+		Warnings:         []contracts.Warning{{}},
+		OutputFiles:      []contracts.OutputFile{{Name: "index.json"}},
+	}
+}
+
+func testRunOutputWithout(change func(*pipeline.RunOutput)) pipeline.RunOutput {
+	output := testRunOutput()
+	change(&output)
+	return output
+}
+
+func mustWorkingDirectory(t *testing.T) string {
+	t.Helper()
+	workingDirectory, err := filepath.Abs(".")
+	if err != nil {
+		t.Fatal(err)
+	}
+	return workingDirectory
+}
+
+type partialResultWriter struct {
+	writer io.Writer
+	limit  int
+}
+
+func (w partialResultWriter) Write(content []byte) (int, error) {
+	if len(content) > w.limit {
+		content = content[:w.limit]
+	}
+	return w.writer.Write(content)
+}
+
+type failingResultWriter struct{ err error }
+
+func (w failingResultWriter) Write([]byte) (int, error) { return 0, w.err }
+
+type zeroResultWriter struct{}
+
+func (zeroResultWriter) Write([]byte) (int, error) { return 0, nil }

internal/cli/run_terminal.go

@@ -0,0 +1,90 @@
+package cli
+
+import (
+	"errors"
+	"fmt"
+	"io"
+
+	"gitea.maximumdirect.net/eric/notarius/internal/core/debugbundle"
+	"gitea.maximumdirect.net/eric/notarius/internal/framework/pipeline"
+)
+
+type DebugTerminalWriter interface {
+	WriteRunReport(debugbundle.RunReport) error
+	WriteError(string) error
+}
+
+type pipelineCommandState struct {
+	report       debugbundle.RunReport
+	terminalized bool
+}
+
+func newPipelineCommandState(runID, pipelineID, outputPath string) *pipelineCommandState {
+	return &pipelineCommandState{report: debugbundle.RunReport{
+		RunID:      runID,
+		PipelineID: pipelineID,
+		OutputPath: outputPath,
+	}}
+}
+
+func (s *pipelineCommandState) setDebugPath(debugPath string) {
+	if s != nil {
+		s.report.DebugPath = debugPath
+	}
+}
+
+func (s *pipelineCommandState) observeOutput(output pipeline.RunOutput) {
+	if s == nil {
+		return
+	}
+	s.report.OutputCount = len(output.NormalizeOutputs)
+	s.report.RejectedCount = len(output.Rejected)
+	s.report.WarningCount = len(output.Warnings)
+	s.report.ValidationStatus = output.Manifest.ValidationStatus
+}
+
+func (s *pipelineCommandState) terminalize(writer DebugTerminalWriter, primaryErr error) (error, error) {
+	if s == nil || s.terminalized {
+		return primaryErr, nil
+	}
+	s.terminalized = true
+	if writer == nil {
+		return primaryErr, nil
+	}
+
+	report := s.report
+	report.Succeeded = primaryErr == nil
+	reportErr := writer.WriteRunReport(report)
+	if reportErr != nil {
+		reportErr = fmt.Errorf("write debug run report: %w", reportErr)
+		if primaryErr == nil {
+			primaryErr = reportErr
+			reportErr = nil
+		}
+	}
+
+	var errorLogErr error
+	if primaryErr != nil {
+		if err := writer.WriteError(primaryErr.Error()); err != nil {
+			errorLogErr = fmt.Errorf("write debug error log: %w", err)
+		}
+	}
+	return primaryErr, errors.Join(reportErr, errorLogErr)
+}
+
+func failPipelineCommand(stderr io.Writer, state *pipelineCommandState, writer DebugTerminalWriter, primaryErr error, persistenceErrs ...error) int {
+	primaryErr, terminalErr := state.terminalize(writer, primaryErr)
+	persistenceErrs = append(persistenceErrs, terminalErr)
+	return writePipelineCommandFailure(stderr, state, primaryErr, errors.Join(persistenceErrs...))
+}
+
+func writePipelineCommandFailure(stderr io.Writer, state *pipelineCommandState, primaryErr, persistenceErr error) int {
+	fmt.Fprintf(stderr, "notarius: %v\n", primaryErr)
+	if persistenceErr != nil {
+		fmt.Fprintf(stderr, "notarius: %v\n", persistenceErr)
+	}
+	if state != nil && state.report.DebugPath != "" {
+		fmt.Fprintf(stderr, "notarius: debug=%s\n", state.report.DebugPath)
+	}
+	return 1
+}

internal/cli/spell_catalog_identity_contract_test.go

@@ -0,0 +1,503 @@
+package cli
+
+import (
+	"crypto/sha256"
+	"encoding/hex"
+	"encoding/json"
+	"os"
+	"path/filepath"
+	"reflect"
+	"strings"
+	"testing"
+
+	"gitea.maximumdirect.net/eric/notarius/internal/core/artifacts"
+	"gitea.maximumdirect.net/eric/notarius/internal/core/config"
+	"gitea.maximumdirect.net/eric/notarius/internal/core/source"
+	"gitea.maximumdirect.net/eric/notarius/internal/framework/checkpoint"
+	"gitea.maximumdirect.net/eric/notarius/internal/framework/contracts"
+	"gitea.maximumdirect.net/eric/notarius/internal/framework/pipeline"
+	"gitea.maximumdirect.net/eric/notarius/internal/modules/dnd"
+	"gitea.maximumdirect.net/eric/notarius/internal/modules/dnd/extract/spells"
+	spellnormalize "gitea.maximumdirect.net/eric/notarius/internal/modules/dnd/normalize/spells"
+	spellcatalog "gitea.maximumdirect.net/eric/notarius/internal/modules/dnd/spells/catalog"
+)
+
+func TestSpellCatalogBytesAffectCheckpointIdentityButNotSemanticDigest(t *testing.T) {
+	components := productionTestComponents(t)
+	configPath := writeProductionSpellCatalogContractConfig(t)
+	effective, err := loadMaintainedExample(t, configPath).Resolve(resolveInputForMaintainedExample(components, "dnd-session"))
+	if err != nil {
+		t.Fatalf("resolve production configuration: %v", err)
+	}
+	overlayPath := filepath.Join(t.TempDir(), "catalog.json")
+	resolved := effective.ResolvedPipeline
+	bindings := resolved.Steps[0].ArtifactLanes[0].ExtractReferences.Bindings
+	catalogBindingIndex := -1
+	for index, binding := range bindings {
+		if binding.SlotName == spellcatalog.SpellCatalogReferenceSlot {
+			catalogBindingIndex = index
+			break
+		}
+	}
+	if catalogBindingIndex < 0 {
+		t.Fatalf("spell catalog bindings = %#v, want catalog binding", bindings)
+	}
+	resolved.Steps[0].ArtifactLanes[0].ExtractReferences.Bindings[catalogBindingIndex].Source = overlayPath
+	normalizeBindings := resolved.Steps[0].ArtifactLanes[0].NormalizeReferences.Bindings
+	normalizeCatalogBindingIndex := -1
+	for index, binding := range normalizeBindings {
+		if binding.SlotName == spellcatalog.SpellCatalogReferenceSlot {
+			normalizeCatalogBindingIndex = index
+			break
+		}
+	}
+	if normalizeCatalogBindingIndex < 0 {
+		t.Fatalf("normalize spell catalog bindings = %#v, want catalog binding", normalizeBindings)
+	}
+	resolved.Steps[0].ArtifactLanes[0].NormalizeReferences.Bindings[normalizeCatalogBindingIndex].Source = overlayPath
+
+	if err := os.WriteFile(overlayPath, []byte(reorderedOverlayA), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	materializedA, _, err := pipeline.MaterializeReferences(resolved, catalogFromRegistries(components.registries), pipeline.ReferenceMaterializationOptions{ConfigPath: configPath, WorkingDir: filepath.Dir(configPath)})
+	if err != nil {
+		t.Fatalf("materialize first catalog: %v", err)
+	}
+	identityA := catalogCheckpointIdentity(t, materializedA)
+	metadataA := catalogExtractorMetadata(t, materializedA)
+	normalizerMetadataA := catalogNormalizerMetadata(t, materializedA)
+	referenceA := catalogReference(t, materializedA)
+
+	if err := os.WriteFile(overlayPath, []byte(reorderedOverlayB), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	materializedB, _, err := pipeline.MaterializeReferences(resolved, catalogFromRegistries(components.registries), pipeline.ReferenceMaterializationOptions{ConfigPath: configPath, WorkingDir: filepath.Dir(configPath)})
+	if err != nil {
+		t.Fatalf("materialize reordered catalog: %v", err)
+	}
+	identityB := catalogCheckpointIdentity(t, materializedB)
+	metadataB := catalogExtractorMetadata(t, materializedB)
+	normalizerMetadataB := catalogNormalizerMetadata(t, materializedB)
+	referenceB := catalogReference(t, materializedB)
+
+	if identityA.Digest == identityB.Digest {
+		t.Fatalf("checkpoint identity digest = %q for both raw catalog files, want invalidation", identityA.Digest)
+	}
+	if referenceA.Digest == referenceB.Digest || referenceA.OriginURI != referenceB.OriginURI {
+		t.Fatalf("catalog reference provenance changed from %#v to %#v, want same origin and different raw digest", referenceA, referenceB)
+	}
+	digestA, ok := metadataA["catalog_digest"].(string)
+	if !ok {
+		t.Fatalf("first extractor catalog metadata = %#v, want digest", metadataA)
+	}
+	digestB, ok := metadataB["catalog_digest"].(string)
+	if !ok || digestA != digestB {
+		t.Fatalf("extractor catalog digests = %q and %q, want same semantic digest", digestA, digestB)
+	}
+	if got, want := metadataA["catalog_overlay_ids"], []string{"campaign.a", "campaign.b"}; !reflect.DeepEqual(got, want) || !reflect.DeepEqual(metadataB["catalog_overlay_ids"], want) {
+		t.Fatalf("extractor overlay IDs = %#v and %#v, want %#v", got, metadataB["catalog_overlay_ids"], want)
+	}
+	normalizerDigestA, ok := normalizerMetadataA["catalog_digest"].(string)
+	normalizerDigestB, okB := normalizerMetadataB["catalog_digest"].(string)
+	if !ok || !okB || normalizerDigestA != digestA || normalizerDigestB != digestB {
+		t.Fatalf("normalizer catalog digests = %#v and %#v, want extractor semantic digests %q and %q", normalizerMetadataA["catalog_digest"], normalizerMetadataB["catalog_digest"], digestA, digestB)
+	}
+	if got, want := normalizerMetadataA["catalog_overlay_ids"], []string{"campaign.a", "campaign.b"}; !reflect.DeepEqual(got, want) || !reflect.DeepEqual(normalizerMetadataB["catalog_overlay_ids"], want) {
+		t.Fatalf("normalizer overlay IDs = %#v and %#v, want %#v", got, normalizerMetadataB["catalog_overlay_ids"], want)
+	}
+}
+
+func TestConfiguredSpellCatalogBindingChangesResolvedPipelineIdentity(t *testing.T) {
+	base := productionSpellCatalogContractConfig(t)
+	changed := strings.Replace(base, repositoryPath("examples", "dnd-spell-catalog.json"), filepath.Join(t.TempDir(), "alternate-spell-catalog.json"), 1)
+	if changed == base {
+		t.Fatal("production configuration did not contain the maintained catalog binding")
+	}
+	root := t.TempDir()
+	firstPath := filepath.Join(root, "first.yml")
+	secondPath := filepath.Join(root, "second.yml")
+	if err := os.WriteFile(firstPath, []byte(base), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.WriteFile(secondPath, []byte(changed), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	components := productionTestComponents(t)
+	first, err := loadMaintainedExample(t, firstPath).Resolve(resolveInputForMaintainedExample(components, "dnd-session"))
+	if err != nil {
+		t.Fatalf("resolve first configuration: %v", err)
+	}
+	second, err := loadMaintainedExample(t, secondPath).Resolve(resolveInputForMaintainedExample(components, "dnd-session"))
+	if err != nil {
+		t.Fatalf("resolve changed configuration: %v", err)
+	}
+	if first.ResolvedPipeline.Digest == second.ResolvedPipeline.Digest {
+		t.Fatalf("resolved pipeline digest = %q for different catalog bindings, want change", first.ResolvedPipeline.Digest)
+	}
+}
+
+func TestSemanticSpellCatalogFingerprintChangesCheckpointIdentityWithoutReferenceChange(t *testing.T) {
+	components := productionTestComponents(t)
+	configPath := writeProductionSpellCatalogContractConfig(t)
+	effective, err := loadMaintainedExample(t, configPath).Resolve(resolveInputForMaintainedExample(components, "dnd-session"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	materialized, _, err := pipeline.MaterializeReferences(effective.ResolvedPipeline, catalogFromRegistries(components.registries), pipeline.ReferenceMaterializationOptions{ConfigPath: configPath, WorkingDir: filepath.Dir(configPath)})
+	if err != nil {
+		t.Fatal(err)
+	}
+	prepared, err := pipeline.Prepare(materialized, components.registries, pipeline.ModuleDependencies{LLM: &productionFakeLLMClient{}})
+	if err != nil {
+		t.Fatal(err)
+	}
+	fingerprints := prepared.CheckpointFingerprints()
+	wantNames := map[string]struct{}{
+		"extract:spells:" + spells.Key + ":effective_catalog":                                                  {},
+		"extract:spells:" + spells.Key + ":validator:2:extract/dnd/spells/catalog:effective_catalog":           {},
+		"normalize:spells:" + spellnormalize.Key + ":effective_catalog":                                        {},
+		"normalize:spells:" + spellnormalize.Key + ":validator:2:extract/dnd/spells/catalog:effective_catalog": {},
+	}
+	seen := make(map[string]string, len(fingerprints))
+	for _, fingerprint := range fingerprints {
+		if _, ok := wantNames[fingerprint.Name]; ok {
+			seen[fingerprint.Name] = fingerprint.Value
+		}
+	}
+	if len(seen) != len(wantNames) {
+		t.Fatalf("prepared fingerprints = %#v, want scoped extractor and normalize catalog identities", fingerprints)
+	}
+	var catalogDigest string
+	for name, value := range seen {
+		if catalogDigest == "" {
+			catalogDigest = value
+		} else if value != catalogDigest {
+			t.Fatalf("prepared fingerprint %q = %q, want shared semantic catalog digest %q", name, value, catalogDigest)
+		}
+	}
+
+	identityFor := func(values []pipeline.CheckpointFingerprint) checkpoint.Identity {
+		identity, identityErr := checkpoint.NewIdentity(checkpoint.IdentityInput{
+			Pipeline:               materialized,
+			InputKey:               materialized.Input.Module,
+			RawInputDigest:         "sha256:unchanged-input",
+			References:             pipeline.ReferenceProvenance(materialized),
+			ProvenanceFingerprints: checkpointIdentityFingerprints(values),
+		})
+		if identityErr != nil {
+			t.Fatal(identityErr)
+		}
+		return identity
+	}
+	first := identityFor(fingerprints)
+	changed := replaceCheckpointFingerprintValue(t, fingerprints, normalizeSpellCatalogFingerprintName(), "sha256:changed-effective-catalog")
+	assertOnlyCheckpointFingerprintChanged(t, fingerprints, changed, normalizeSpellCatalogFingerprintName())
+	second := identityFor(changed)
+	if first.Digest == second.Digest || reflect.DeepEqual(first.ReferenceDigests, nil) || !reflect.DeepEqual(first.ReferenceDigests, second.ReferenceDigests) {
+		t.Fatalf("identities = %#v / %#v, want semantic invalidation with unchanged reference provenance", first, second)
+	}
+}
+
+func TestChangedSemanticSpellCatalogFingerprintCannotResumeRecordedCheckpoint(t *testing.T) {
+	components := productionTestComponents(t)
+	configPath := writeProductionSpellCatalogContractConfig(t)
+	effective, err := loadMaintainedExample(t, configPath).Resolve(resolveInputForMaintainedExample(components, "dnd-session"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	materialized, _, err := pipeline.MaterializeReferences(effective.ResolvedPipeline, catalogFromRegistries(components.registries), pipeline.ReferenceMaterializationOptions{ConfigPath: configPath, WorkingDir: filepath.Dir(configPath)})
+	if err != nil {
+		t.Fatal(err)
+	}
+	prepared, err := pipeline.Prepare(materialized, components.registries, pipeline.ModuleDependencies{LLM: &productionFakeLLMClient{}})
+	if err != nil {
+		t.Fatal(err)
+	}
+	fingerprints := prepared.CheckpointFingerprints()
+	llmFingerprints := []checkpoint.Fingerprint{{Name: "promptkit_profile_source", Value: "sha256:profile-source-one"}}
+	settings := config.CheckpointCacheConfig{Enabled: true, Directory: t.TempDir()}
+	recorder, _, err := checkpointHandlersForRun(settings, Options{}, materialized, fingerprints, llmFingerprints, []byte("same input"), nil, nil, "", "", LLMRuntimeOverrides{}, false)
+	if err != nil {
+		t.Fatal(err)
+	}
+	doc := source.SourceDocument{ID: "source", Kind: "transcript", Format: "application/json"}
+	doc.Units = []source.SourceUnit{{ID: 1, Kind: "turn", Text: "Aria casts Cure Wounds.", Ref: source.SourceRef{SourceID: doc.ID, StartUnitID: 1, EndUnitID: 1}}}
+	doc.Digest, err = source.DigestDocument(&doc)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if err := recorder.SourceSucceeded(materialized.Input.Module, &doc); err != nil {
+		t.Fatal(err)
+	}
+	normalizeDependencies := []pipeline.CheckpointFingerprint{{Name: "artifact[0]", Value: "sha256:merged-artifact"}}
+	normalizeSchema := contracts.ArtifactSchema{ID: "notarius.dnd.spells", Name: "notarius_dnd_spells", Version: "v1"}
+	normalizeArtifact := pipeline.CheckpointArtifact{
+		LaneID: "spells", ModuleKey: spellnormalize.Key, SourceID: doc.ID,
+		SchemaDigest: contracts.DigestArtifactSchema(normalizeSchema),
+		Artifact: contracts.SerializedArtifact{
+			Kind: dnd.SpellListKind, Schema: normalizeSchema, MediaType: "application/json", Content: []byte(`{"spell_casts":[]}`),
+		},
+	}
+	if err := recorder.NormalizeSucceeded("spells", spellnormalize.Key, normalizeDependencies, normalizeArtifact, nil); err != nil {
+		t.Fatal(err)
+	}
+
+	_, sameLoader, err := checkpointHandlersForRun(settings, Options{}, materialized, fingerprints, llmFingerprints, []byte("same input"), nil, nil, "", "", LLMRuntimeOverrides{}, true)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if _, decision := sameLoader.Source(materialized.Input.Module); !decision.Reused {
+		t.Fatalf("same fingerprint decision = %#v, want reuse", decision)
+	}
+	if restored, decision := sameLoader.Normalize("spells", spellnormalize.Key, normalizeDependencies); !decision.Reused || string(restored.Output.Artifact.Content) != `{"spell_casts":[]}` {
+		t.Fatalf("same normalize checkpoint = %#v, decision=%#v, want reuse", restored, decision)
+	}
+	changed := replaceCheckpointFingerprintValue(t, fingerprints, normalizeSpellCatalogFingerprintName(), "sha256:changed-effective-catalog")
+	assertOnlyCheckpointFingerprintChanged(t, fingerprints, changed, normalizeSpellCatalogFingerprintName())
+	_, changedLoader, err := checkpointHandlersForRun(settings, Options{}, materialized, changed, llmFingerprints, []byte("same input"), nil, nil, "", "", LLMRuntimeOverrides{}, true)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if _, decision := changedLoader.Source(materialized.Input.Module); decision.Reused {
+		t.Fatalf("changed fingerprint decision = %#v, want cold miss", decision)
+	}
+	if _, decision := changedLoader.Normalize("spells", spellnormalize.Key, normalizeDependencies); decision.Reused {
+		t.Fatalf("changed normalize fingerprint decision = %#v, want normalize checkpoint cold miss", decision)
+	}
+	changedMapping := replaceCheckpointFingerprintValue(t, fingerprints, extractSpellMappingFingerprintName(), "dnd.spells.extract_mapping.v3")
+	assertOnlyCheckpointFingerprintChanged(t, fingerprints, changedMapping, extractSpellMappingFingerprintName())
+	_, mappingLoader, err := checkpointHandlersForRun(settings, Options{}, materialized, changedMapping, llmFingerprints, []byte("same input"), nil, nil, "", "", LLMRuntimeOverrides{}, true)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if _, decision := mappingLoader.Source(materialized.Input.Module); decision.Reused {
+		t.Fatalf("changed mapping policy decision = %#v, want cold miss", decision)
+	}
+
+	changedLLMFingerprints := []checkpoint.Fingerprint{{Name: "promptkit_profile_source", Value: "sha256:profile-source-two"}}
+	_, profileLoader, err := checkpointHandlersForRun(settings, Options{}, materialized, fingerprints, changedLLMFingerprints, []byte("same input"), nil, nil, "", "", LLMRuntimeOverrides{}, true)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if _, decision := profileLoader.Source(materialized.Input.Module); decision.Reused {
+		t.Fatalf("changed PromptKit profile source decision = %#v, want cold miss", decision)
+	}
+	if _, decision := profileLoader.Normalize("spells", spellnormalize.Key, normalizeDependencies); decision.Reused {
+		t.Fatalf("changed PromptKit profile normalize decision = %#v, want cold miss", decision)
+	}
+}
+
+func normalizeSpellCatalogFingerprintName() string {
+	return "normalize:spells:" + spellnormalize.Key + ":effective_catalog"
+}
+
+func extractSpellMappingFingerprintName() string {
+	return "extract:spells:" + spells.Key + ":mapping_policy"
+}
+
+func replaceCheckpointFingerprintValue(t *testing.T, fingerprints []pipeline.CheckpointFingerprint, name, value string) []pipeline.CheckpointFingerprint {
+	t.Helper()
+	changed := append([]pipeline.CheckpointFingerprint(nil), fingerprints...)
+	matches := 0
+	for index := range changed {
+		if changed[index].Name == name {
+			changed[index].Value = value
+			matches++
+		}
+	}
+	if matches != 1 {
+		t.Fatalf("checkpoint fingerprints = %#v, want exactly one fingerprint named %q", fingerprints, name)
+	}
+	return changed
+}
+
+func assertOnlyCheckpointFingerprintChanged(t *testing.T, before, after []pipeline.CheckpointFingerprint, changedName string) {
+	t.Helper()
+	if len(before) != len(after) {
+		t.Fatalf("fingerprint lengths = %d and %d, want equal", len(before), len(after))
+	}
+	changes := 0
+	for index := range before {
+		if before[index].Name != after[index].Name {
+			t.Fatalf("fingerprint[%d] name changed from %q to %q", index, before[index].Name, after[index].Name)
+		}
+		if before[index].Value == after[index].Value {
+			continue
+		}
+		changes++
+		if before[index].Name != changedName {
+			t.Fatalf("fingerprint %q changed unexpectedly", before[index].Name)
+		}
+	}
+	if changes != 1 {
+		t.Fatalf("fingerprints changed %d values, want exactly %q", changes, changedName)
+	}
+}
+
+func TestMaintainedProductionOverlayRunAlignsGroundingValidationAndProvenance(t *testing.T) {
+	outputRoot := filepath.Join(t.TempDir(), "output")
+	fake := &productionFakeLLMClient{spellResponse: productionSpellResponse("Aegis of Emberfall")}
+	options := productionRunOptions(t, fake)
+	var stdout, stderr strings.Builder
+	code := RunWithOptions([]string{
+		"run", "dnd-session",
+		"--config", writeProductionSpellCatalogContractConfig(t),
+		"--input", repositoryPath("examples", "seriatim-minimal-transcript.json"),
+		"--only", "spells", "--chunk_cache", "bypass", "--output-dir", outputRoot,
+	}, &stdout, &stderr, options)
+	if code != 0 {
+		t.Fatalf("code=%d stdout=%q stderr=%q", code, stdout.String(), stderr.String())
+	}
+
+	runRoot := filepath.Join(outputRoot, productionRunID)
+	manifest := readProductionJSON[artifacts.RunManifest](t, filepath.Join(runRoot, "manifest.json"))
+	if manifest.ValidationStatus != "approved" || len(manifest.References) == 0 || len(manifest.ArtifactLanes) != 1 {
+		t.Fatalf("manifest = %#v, want approved overlay run with one lane and references", manifest)
+	}
+	lane := manifest.ArtifactLanes[0]
+	extractorMetadata, ok := lane.Metadata["extractor"].(map[string]any)
+	if !ok {
+		t.Fatalf("lane metadata = %#v, want extractor metadata", lane.Metadata)
+	}
+	if extractorMetadata["catalog_base_id"] != spellcatalog.SRD5E2014ID || !strings.HasPrefix(stringValue(extractorMetadata["catalog_digest"]), "sha256:") {
+		t.Fatalf("extractor catalog metadata = %#v, want base ID and semantic digest", extractorMetadata)
+	}
+	if got := stringValues(extractorMetadata["catalog_overlay_ids"]); !reflect.DeepEqual(got, []string{"notarius.example-campaign"}) {
+		t.Fatalf("catalog overlay IDs = %#v, want maintained overlay", got)
+	}
+	normalizerMetadata, ok := lane.Metadata["normalizer"].(map[string]any)
+	if !ok {
+		t.Fatalf("lane metadata = %#v, want normalizer metadata", lane.Metadata)
+	}
+	if normalizerMetadata["catalog_base_id"] != spellcatalog.SRD5E2014ID || !strings.HasPrefix(stringValue(normalizerMetadata["catalog_digest"]), "sha256:") || !reflect.DeepEqual(stringValues(normalizerMetadata["catalog_overlay_ids"]), []string{"notarius.example-campaign"}) {
+		t.Fatalf("normalizer catalog metadata = %#v, want base ID, semantic digest, and overlay IDs", normalizerMetadata)
+	}
+	if normalizerMetadata["catalog_digest"] != extractorMetadata["catalog_digest"] || !reflect.DeepEqual(stringValues(normalizerMetadata["catalog_overlay_ids"]), stringValues(extractorMetadata["catalog_overlay_ids"])) {
+		t.Fatalf("extractor metadata = %#v, normalizer metadata = %#v, want shared catalog identity", extractorMetadata, normalizerMetadata)
+	}
+
+	var catalogProvenances []artifacts.ReferenceProvenance
+	for index := range manifest.References {
+		reference := &manifest.References[index]
+		if reference.SlotName == spellcatalog.SpellCatalogReferenceSlot {
+			catalogProvenances = append(catalogProvenances, *reference)
+		}
+	}
+	if len(catalogProvenances) != 2 {
+		t.Fatalf("manifest references = %#v, want independently materialized extract and normalize catalog provenance", manifest.References)
+	}
+	overlayBytes := readRepositoryFile(t, "examples", "dnd-spell-catalog.json")
+	for _, catalogProvenance := range catalogProvenances {
+		if catalogProvenance.Stage != "extract" && catalogProvenance.Stage != "normalize" {
+			t.Fatalf("catalog provenance = %#v, want extract or normalize scope", catalogProvenance)
+		}
+		if catalogProvenance.LaneID != "spells" || catalogProvenance.OriginType != "file" || catalogProvenance.MediaType != "application/json" || catalogProvenance.SizeBytes != int64(len(overlayBytes)) || catalogProvenance.Digest != digestBytes(overlayBytes) || !strings.Contains(catalogProvenance.OriginURI, "dnd-spell-catalog.json") {
+			t.Fatalf("catalog provenance = %#v, want raw overlay provenance in both scopes", catalogProvenance)
+		}
+	}
+	manifestBytes, err := json.Marshal(manifest)
+	if err != nil {
+		t.Fatal(err)
+	}
+	for _, leaked := range []string{"Aegis of Emberfall", "Emberfall Aegis", "Notarius example campaign spell names"} {
+		if strings.Contains(string(manifestBytes), leaked) {
+			t.Fatalf("manifest leaked overlay content %q", leaked)
+		}
+	}
+
+	requests := fake.requestsFor(spells.PromptID)
+	if len(requests) != 1 {
+		t.Fatalf("spell requests = %d, want one", len(requests))
+	}
+	catalogInput, ok := requests[0].Inputs[spellcatalog.SpellCatalogReferenceSlot]
+	if !ok || !strings.Contains(string(catalogInput.Content), "Aegis of Emberfall") || strings.Contains(string(catalogInput.Content), "Emberfall Aegis") {
+		t.Fatalf("spell catalog prompt input = %#v, want canonical overlay name without alias", catalogInput)
+	}
+	artifact := readProductionJSON[dnd.SpellList](t, filepath.Join(runRoot, "lanes", "spells.json"))
+	if len(artifact.SpellCasts) != 1 || artifact.SpellCasts[0].Spell != "Aegis of Emberfall" {
+		t.Fatalf("artifact = %#v, want accepted overlay-only canonical spell", artifact)
+	}
+	rejected := readProductionJSON[struct {
+		Rejected []json.RawMessage `json:"rejected"`
+	}](t, filepath.Join(runRoot, "rejected.json"))
+	if len(rejected.Rejected) != 0 {
+		t.Fatalf("rejected = %#v, want no rejected output", rejected.Rejected)
+	}
+}
+
+func catalogCheckpointIdentity(t *testing.T, resolved pipeline.ResolvedPipeline) checkpoint.Identity {
+	t.Helper()
+	identity, err := checkpoint.NewIdentity(checkpoint.IdentityInput{
+		Pipeline:       resolved,
+		InputKey:       resolved.Input.Module,
+		RawInputDigest: "sha256:catalog-test-input",
+		References:     pipeline.ReferenceProvenance(resolved),
+	})
+	if err != nil {
+		t.Fatalf("create checkpoint identity: %v", err)
+	}
+	return identity
+}
+
+func catalogExtractorMetadata(t *testing.T, resolved pipeline.ResolvedPipeline) map[string]any {
+	t.Helper()
+	lane := resolved.Steps[0].ArtifactLanes[0]
+	extractor, err := spells.New(&productionFakeLLMClient{}, spells.Options{}, lane.ExtractReferences.ReferenceSet)
+	if err != nil {
+		t.Fatalf("construct extractor: %v", err)
+	}
+	return extractor.ManifestMetadata()
+}
+
+func catalogNormalizerMetadata(t *testing.T, resolved pipeline.ResolvedPipeline) map[string]any {
+	t.Helper()
+	lane := resolved.Steps[0].ArtifactLanes[0]
+	normalizer, err := spellnormalize.New(spellnormalize.Options{}, lane.NormalizeReferences.ReferenceSet)
+	if err != nil {
+		t.Fatalf("construct normalizer: %v", err)
+	}
+	return normalizer.ManifestMetadata()
+}
+
+func catalogReference(t *testing.T, resolved pipeline.ResolvedPipeline) artifacts.ReferenceProvenance {
+	t.Helper()
+	for _, reference := range pipeline.ReferenceProvenance(resolved) {
+		if reference.SlotName == spellcatalog.SpellCatalogReferenceSlot && reference.Stage == "extract" && reference.LaneID == "spells" {
+			return reference
+		}
+	}
+	t.Fatalf("resolved references = %#v, want spell catalog provenance", pipeline.ReferenceProvenance(resolved))
+	return artifacts.ReferenceProvenance{}
+}
+
+func stringValue(value any) string {
+	result, _ := value.(string)
+	return result
+}
+
+func stringValues(value any) []string {
+	raw, err := json.Marshal(value)
+	if err != nil {
+		return nil
+	}
+	var values []string
+	if err := json.Unmarshal(raw, &values); err != nil {
+		return nil
+	}
+	return values
+}
+
+func digestBytes(value []byte) string {
+	sum := sha256.Sum256(value)
+	return "sha256:" + hex.EncodeToString(sum[:])
+}
+
+const reorderedOverlayA = `{
+  "schema_version": "notarius.dnd.spell-catalog-overlay.v1",
+  "catalogs": [
+    {"id":"campaign.a","ruleset":"dnd-5e-2014","source":{"title":"Campaign A"},"spells":[{"name":"Aegis of Emberfall","aliases":["Emberfall Aegis"]}]},
+    {"id":"campaign.b","ruleset":"dnd-5e-2014","source":{"title":"Campaign B"},"spells":[{"name":"Cinder Veil","aliases":["Veil of Cinder","Cinder Shroud"]}]}
+  ]
+}`
+
+const reorderedOverlayB = `{"catalogs":[{"spells":[{"aliases":["Cinder Shroud","Veil of Cinder"],"name":"Cinder Veil"}],"source":{"title":"Campaign B"},"ruleset":"dnd-5e-2014","id":"campaign.b"},{"spells":[{"aliases":["Emberfall Aegis"],"name":"Aegis of Emberfall"}],"source":{"title":"Campaign A"},"ruleset":"dnd-5e-2014","id":"campaign.a"}],"schema_version":"notarius.dnd.spell-catalog-overlay.v1"}`

internal/cli/spell_catalog_retry_contract_test.go

@@ -0,0 +1,160 @@
+package cli
+
+import (
+	"context"
+	"encoding/json"
+	"fmt"
+	"path/filepath"
+	"sync"
+	"testing"
+
+	"gitea.maximumdirect.net/eric/notarius/internal/core/config"
+	"gitea.maximumdirect.net/eric/notarius/internal/core/source"
+	"gitea.maximumdirect.net/eric/notarius/internal/framework/contracts"
+	"gitea.maximumdirect.net/eric/notarius/internal/framework/pipeline"
+	"gitea.maximumdirect.net/eric/notarius/internal/modules/dnd"
+	"gitea.maximumdirect.net/eric/notarius/internal/modules/dnd/extract/spells"
+)
+
+func TestProductionSpellCatalogValidationRetries(t *testing.T) {
+	const retries = 2
+	tests := []struct {
+		name            string
+		responses       []string
+		wantCalls       int
+		wantRejected    bool
+		wantSpell       string
+		wantWarningCode string
+	}{
+		{
+			name: "unknown spell remains rejected after exhaustion",
+			responses: []string{
+				productionSpellResponse("Unknown Spell"),
+				productionSpellResponse("Unknown Spell"),
+				productionSpellResponse("Unknown Spell"),
+			},
+			wantCalls:    retries + 1,
+			wantRejected: true,
+		},
+		{
+			name: "overlay spell becomes valid on retry",
+			responses: []string{
+				productionSpellResponse("Unknown Spell"),
+				productionSpellResponse("Aegis of Emberfall"),
+			},
+			wantCalls:       2,
+			wantSpell:       "Aegis of Emberfall",
+			wantWarningCode: "spell_not_near_source",
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			components := productionTestComponents(t)
+			configPath := writeProductionSpellCatalogContractConfig(t)
+			cfg := loadMaintainedExample(t, configPath)
+			effective, err := cfg.Resolve(config.ResolveInput{PipelineID: "dnd-session", Catalog: catalogFromRegistries(components.registries)})
+			if err != nil {
+				t.Fatalf("resolve production configuration: %v", err)
+			}
+			materialized, _, err := pipeline.MaterializeReferences(effective.ResolvedPipeline, catalogFromRegistries(components.registries), pipeline.ReferenceMaterializationOptions{
+				ConfigPath: configPath,
+				WorkingDir: filepath.Dir(configPath),
+			})
+			if err != nil {
+				t.Fatalf("materialize production references: %v", err)
+			}
+			materialized.Steps[0].ArtifactLanes[0].Extract.Retries = retries
+
+			llmClient := &catalogRetryLLMClient{responses: tt.responses}
+			prepared, err := pipeline.Prepare(materialized, components.registries, pipeline.ModuleDependencies{LLM: llmClient})
+			if err != nil {
+				t.Fatalf("prepare production pipeline: %v", err)
+			}
+			output, err := pipeline.New().Run(context.Background(), pipeline.RunInput{
+				Prepared:       prepared,
+				RawInput:       readRepositoryFile(t, "examples", "seriatim-minimal-transcript.json"),
+				ChunkCacheMode: pipeline.ChunkCacheBypass,
+			})
+			if err != nil {
+				t.Fatalf("Run() error = %v, want nil", err)
+			}
+			if calls := llmClient.CallCount(); calls > retries+1 || calls != tt.wantCalls {
+				t.Fatalf("LLM calls = %d, want %d and no more than %d", calls, tt.wantCalls, retries+1)
+			}
+
+			if tt.wantRejected {
+				if len(output.Rejected) != 1 || len(output.NormalizeOutputs) != 0 {
+					t.Fatalf("rejected = %#v normalized = %#v, want one nonfatal rejection and no merge output", output.Rejected, output.NormalizeOutputs)
+				}
+				rejection := output.Rejected[0]
+				if rejection.ReasonCode != "unknown_spell" || rejection.AttemptCount != retries+1 {
+					t.Fatalf("rejection = %#v, want exhausted unknown-spell rejection", rejection)
+				}
+				if len(output.Warnings) != 0 {
+					t.Fatalf("warnings = %#v, want no warnings from rejected attempts", output.Warnings)
+				}
+				return
+			}
+
+			if len(output.Rejected) != 0 || len(output.NormalizeOutputs) != 1 {
+				t.Fatalf("rejected = %#v normalized = %#v, want only accepted output", output.Rejected, output.NormalizeOutputs)
+			}
+			var value dnd.SpellList
+			if err := json.Unmarshal(output.NormalizeOutputs[0].Artifact.Content, &value); err != nil {
+				t.Fatalf("decode normalized spell list: %v", err)
+			}
+			if len(value.SpellCasts) != 1 || value.SpellCasts[0].Spell != tt.wantSpell {
+				t.Fatalf("normalized spell list = %#v, want accepted overlay spell", value)
+			}
+			if len(output.Warnings) != 2 || output.Warnings[0].ReasonCode != tt.wantWarningCode || output.Warnings[1].ReasonCode != tt.wantWarningCode {
+				t.Fatalf("warnings = %#v, want accepted-attempt warnings from extract and normalize validation", output.Warnings)
+			}
+		})
+	}
+}
+
+type catalogRetryLLMClient struct {
+	mu        sync.Mutex
+	responses []string
+	calls     int
+}
+
+func (client *catalogRetryLLMClient) CompleteStructured(ctx context.Context, req contracts.StructuredCompletionRequest, out any) (contracts.StructuredCompletionResponse, error) {
+	if err := ctx.Err(); err != nil {
+		return contracts.StructuredCompletionResponse{}, err
+	}
+	if req.PromptID != spells.PromptID {
+		return contracts.StructuredCompletionResponse{}, fmt.Errorf("unexpected prompt %q", req.PromptID)
+	}
+	client.mu.Lock()
+	index := client.calls
+	client.calls++
+	client.mu.Unlock()
+	if index >= len(client.responses) {
+		return contracts.StructuredCompletionResponse{}, fmt.Errorf("missing fake response %d", index)
+	}
+	content := []byte(client.responses[index])
+	if err := json.Unmarshal(content, out); err != nil {
+		return contracts.StructuredCompletionResponse{}, fmt.Errorf("populate fake structured target: %w", err)
+	}
+	return contracts.StructuredCompletionResponse{Content: content, Provider: "test", Model: "deterministic", ProfileID: req.ProfileID}, nil
+}
+
+func (client *catalogRetryLLMClient) CallCount() int {
+	client.mu.Lock()
+	defer client.mu.Unlock()
+	return client.calls
+}
+
+func productionSpellResponse(name string) string {
+	content, err := json.Marshal(dnd.SpellList{SpellCasts: []dnd.SpellCast{{
+		Caster:     "Aria",
+		Spell:      name,
+		SourceRefs: []source.SourceRef{{SourceID: "session-alpha", StartUnitID: 1, EndUnitID: 1}},
+	}}})
+	if err != nil {
+		panic(err)
+	}
+	return string(content)
+}

internal/cli/state_hardening_test.go

@@ -0,0 +1,988 @@
+package cli
+
+import (
+	"bytes"
+	"context"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"os"
+	"path/filepath"
+	"runtime"
+	"strings"
+	"sync"
+	"testing"
+	"time"
+
+	"gitea.maximumdirect.net/eric/notarius/internal/core/artifacts"
+	"gitea.maximumdirect.net/eric/notarius/internal/core/config"
+	"gitea.maximumdirect.net/eric/notarius/internal/core/debugbundle"
+	"gitea.maximumdirect.net/eric/notarius/internal/core/source"
+	"gitea.maximumdirect.net/eric/notarius/internal/framework/chunkplan"
+	"gitea.maximumdirect.net/eric/notarius/internal/framework/contracts"
+	frameworkdebug "gitea.maximumdirect.net/eric/notarius/internal/framework/debug"
+	"gitea.maximumdirect.net/eric/notarius/internal/framework/pipeline"
+)
+
+const stateTestDigest = "sha256:e511d8906649b78eb639b11215fa57a9652a1a64f4aefa3ed68320dbda46f439"
+
+func TestRunStateSurfaceMatrix(t *testing.T) {
+	for _, debug := range []bool{false, true} {
+		for _, resume := range []bool{false, true} {
+			for _, mode := range []string{"auto", "bypass", "refresh"} {
+				name := fmt.Sprintf("debug=%t/resume=%t/cache=%s", debug, resume, mode)
+				t.Run(name, func(t *testing.T) {
+					roots := newStateTestRoots(t)
+					harness := newStateTestHarness()
+					opts := harness.options()
+					var storeRoots []string
+					opts.ChunkPlanStoreFactory = func(root string) (pipeline.ChunkPlanStore, error) {
+						storeRoots = append(storeRoots, root)
+						return chunkplan.NewFilesystemStore(root)
+					}
+					result := runStateTest(t, roots, opts, debug, resume, mode)
+					if result.code != 0 {
+						t.Fatalf("code=%d stderr=%q", result.code, result.stderr)
+					}
+					assertStateTestOutput(t, roots.output)
+					if mode == "bypass" {
+						assertAbsent(t, roots.plans)
+						if len(storeRoots) != 0 {
+							t.Fatalf("chunk plan store roots = %v, want none", storeRoots)
+						}
+					} else {
+						assertFile(t, filepath.Join(roots.plans, strings.TrimPrefix(stateTestDigest, "sha256:"), "plan.json"))
+						if len(storeRoots) != 1 || storeRoots[0] != roots.plans {
+							t.Fatalf("chunk plan store roots = %v, want [%q]", storeRoots, roots.plans)
+						}
+					}
+					assertAnyFile(t, roots.checkpoints)
+					assertRestrictedTree(t, roots.checkpoints)
+					if debug {
+						bundle := onlyChildDir(t, roots.debug)
+						assertFile(t, filepath.Join(bundle, "summary", "invocation.json"))
+						assertAnyFile(t, filepath.Join(bundle, "trace"))
+						assertRestrictedTree(t, roots.debug)
+					} else {
+						assertAbsent(t, roots.debug)
+					}
+				})
+			}
+		}
+	}
+}
+
+func TestRunKeepsStateRootsIndependentAndReusesSelectedCheckpointRoot(t *testing.T) {
+	roots := newStateTestRoots(t)
+	harness := newStateTestHarness()
+	first := runStateTest(t, roots, harness.options(), true, false, "auto")
+	if first.code != 0 {
+		t.Fatalf("first run code=%d stderr=%q", first.code, first.stderr)
+	}
+	planPath := filepath.Join(roots.plans, strings.TrimPrefix(stateTestDigest, "sha256:"), "plan.json")
+	initialPlan, err := os.ReadFile(planPath)
+	if err != nil {
+		t.Fatal(err)
+	}
+	firstBundle := onlyChildDir(t, roots.debug)
+
+	second := runStateTest(t, roots, harness.options(), false, false, "auto")
+	if second.code != 0 {
+		t.Fatalf("second run code=%d stderr=%q", second.code, second.stderr)
+	}
+	if harness.chunkCalls != 1 {
+		t.Fatalf("chunk calls after debug toggle = %d, want 1", harness.chunkCalls)
+	}
+	if harness.extractCalls != 2 {
+		t.Fatalf("extract calls after two recording-only runs = %d, want 2", harness.extractCalls)
+	}
+	if got, err := os.ReadFile(planPath); err != nil || !bytes.Equal(got, initialPlan) {
+		t.Fatalf("chunk plan changed after debug toggle: %v", err)
+	}
+	if _, err := os.Stat(firstBundle); err != nil {
+		t.Fatalf("initial debug bundle was removed: %v", err)
+	}
+
+	checkpointRoot := roots.checkpoints
+	extractCallsBeforeResume := harness.extractCalls
+	seed := runStateTest(t, roots, harness.options(), false, true, "auto")
+	if seed.code != 0 {
+		t.Fatalf("checkpoint seed code=%d stderr=%q", seed.code, seed.stderr)
+	}
+	if harness.extractCalls != extractCallsBeforeResume {
+		t.Fatalf("extract calls after reusing recording-only checkpoint = %d, want %d", harness.extractCalls, extractCallsBeforeResume)
+	}
+	extractCalls := harness.extractCalls
+	checkpointFiles := readTree(t, checkpointRoot)
+	reused := runStateTest(t, roots, harness.options(), false, true, "auto")
+	if reused.code != 0 {
+		t.Fatalf("checkpoint reuse code=%d stderr=%q", reused.code, reused.stderr)
+	}
+	if harness.extractCalls != extractCalls {
+		t.Fatalf("extract calls after checkpoint reuse = %d, want %d", harness.extractCalls, extractCalls)
+	}
+	if got := readTree(t, checkpointRoot); !sameFiles(got, checkpointFiles) {
+		t.Fatal("reused checkpoint was rewritten")
+	}
+}
+
+func TestRunRecomputesOnlyAfterExplicitChunkPlanRemoval(t *testing.T) {
+	roots := newStateTestRoots(t)
+	harness := newStateTestHarness()
+	first := runStateTest(t, roots, harness.options(), true, false, "auto")
+	if first.code != 0 {
+		t.Fatalf("first run code=%d stderr=%q", first.code, first.stderr)
+	}
+	firstOutput := onlyChildDir(t, roots.output)
+	firstBundle := onlyChildDir(t, roots.debug)
+	entry := filepath.Join(roots.plans, strings.TrimPrefix(stateTestDigest, "sha256:"))
+	if err := os.RemoveAll(entry); err != nil {
+		t.Fatal(err)
+	}
+	second := runStateTest(t, roots, harness.options(), false, false, "auto")
+	if second.code != 0 {
+		t.Fatalf("second run code=%d stderr=%q", second.code, second.stderr)
+	}
+	if harness.chunkCalls != 2 {
+		t.Fatalf("chunk calls = %d, want 2 after removing exact cache entry", harness.chunkCalls)
+	}
+	assertFile(t, filepath.Join(firstOutput, "result.json"))
+	assertFile(t, filepath.Join(firstBundle, "summary", "run-report.json"))
+}
+
+func TestRunRetainsDebugBundlesAcrossFailures(t *testing.T) {
+	t.Run("configuration failure precedes allocation", func(t *testing.T) {
+		root := filepath.Join(t.TempDir(), "debug")
+		var stdout, stderr bytes.Buffer
+		code := RunWithOptions([]string{"run", "sample", "--config", filepath.Join(t.TempDir(), "missing.yml"), "--input", "missing", "--debug", "--debug-dir", root}, &stdout, &stderr, newStateTestHarness().options())
+		if code != 1 || !strings.Contains(stderr.String(), "config file") {
+			t.Fatalf("code=%d stderr=%q", code, stderr.String())
+		}
+		assertAbsent(t, root)
+	})
+
+	for _, failure := range []struct {
+		name     string
+		expected string
+		setup    func(*testing.T, stateTestRoots, *stateTestHarness) Options
+	}{
+		{"resolution", "pipeline \"missing\"", func(t *testing.T, roots stateTestRoots, h *stateTestHarness) Options { return h.options() }},
+		{"pipeline", "synthetic extraction failure", func(t *testing.T, roots stateTestRoots, h *stateTestHarness) Options {
+			h.extractErr = errors.New("synthetic extraction failure")
+			return h.options()
+		}},
+		{"output", "create output parent", func(t *testing.T, roots stateTestRoots, h *stateTestHarness) Options {
+			if err := os.WriteFile(roots.output, []byte("not a directory"), 0o600); err != nil {
+				t.Fatal(err)
+			}
+			return h.options()
+		}},
+		{"summary", "write debug invocation metadata", func(t *testing.T, roots stateTestRoots, h *stateTestHarness) Options {
+			opts := h.options()
+			opts.DebugRecorderFactory = func(traceRoot string) (pipeline.DebugRecorder, error) {
+				if err := os.RemoveAll(filepath.Join(filepath.Dir(traceRoot), "summary")); err != nil {
+					return nil, err
+				}
+				if err := os.WriteFile(filepath.Join(filepath.Dir(traceRoot), "summary"), []byte("blocked"), 0o600); err != nil {
+					return nil, err
+				}
+				return frameworkdebug.NewFilesystemRecorder(traceRoot)
+			}
+			return opts
+		}},
+		{"trace", "trace unavailable", func(t *testing.T, roots stateTestRoots, h *stateTestHarness) Options {
+			opts := h.options()
+			opts.DebugRecorderFactory = func(string) (pipeline.DebugRecorder, error) { return failingDebugRecorder{}, nil }
+			return opts
+		}},
+	} {
+		t.Run(failure.name, func(t *testing.T) {
+			roots := newStateTestRoots(t)
+			harness := newStateTestHarness()
+			opts := failure.setup(t, roots, harness)
+			failureStderr := ""
+			if failure.name == "resolution" {
+				var stdout, stderr bytes.Buffer
+				code := RunWithOptions([]string{"run", "missing", "--config", roots.config, "--input", roots.input, "--debug"}, &stdout, &stderr, opts)
+				if code != 1 {
+					t.Fatalf("code=%d stderr=%q", code, stderr.String())
+				}
+				failureStderr = stderr.String()
+			} else {
+				result := runStateTest(t, roots, opts, true, false, "bypass")
+				if result.code != 1 {
+					t.Fatalf("code=%d stderr=%q", result.code, result.stderr)
+				}
+				failureStderr = result.stderr
+			}
+			if !strings.Contains(failureStderr, failure.expected) || !strings.Contains(failureStderr, "debug=") {
+				t.Fatalf("stderr=%q, want %q and debug path", failureStderr, failure.expected)
+			}
+			bundle := onlyChildDir(t, roots.debug)
+			if !strings.Contains(readAllFiles(t, bundle), "synthetic") && failure.name == "pipeline" {
+				t.Fatal("pipeline failure was not retained in debug bundle")
+			}
+		})
+	}
+}
+
+func TestRunDebugArtifactsRedactSecretsButRetainApplicationData(t *testing.T) {
+	roots := newStateTestRoots(t)
+	t.Setenv("STATE_TEST_UNRELATED_ENV", "HOST_ONLY_SENTINEL")
+	if err := os.WriteFile(filepath.Join(filepath.Dir(roots.input), "unrelated.txt"), []byte("HOST_ONLY_FILE_SENTINEL"), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	harness := newStateTestHarness()
+	result := runStateTest(t, roots, harness.options(), true, false, "bypass")
+	if result.code != 0 {
+		t.Fatalf("code=%d stderr=%q", result.code, result.stderr)
+	}
+	bundle := onlyChildDir(t, roots.debug)
+	summary := readAllFiles(t, filepath.Join(bundle, "summary"))
+	trace := readAllFiles(t, filepath.Join(bundle, "trace"))
+	for _, forbidden := range []string{"sk-secretvalue", "Bearer secretvalue", "HOST_ONLY_SENTINEL", "HOST_ONLY_FILE_SENTINEL"} {
+		if strings.Contains(summary, forbidden) || strings.Contains(trace, forbidden) {
+			t.Fatalf("debug bundle contains %q", forbidden)
+		}
+	}
+	if strings.Contains(summary, "application content") {
+		t.Fatal("summary contains raw application input")
+	}
+	if !strings.Contains(trace, "application content") {
+		t.Fatal("trace does not retain expected application input")
+	}
+}
+
+func TestRunRedactsSensitiveModuleOptionsFromConfigAndPipelineSummaries(t *testing.T) {
+	roots := newStateTestRoots(t)
+	data, err := os.ReadFile(roots.config)
+	if err != nil {
+		t.Fatal(err)
+	}
+	configText := replaceRequiredOnce(t, string(data), "    input: test/input\n", `    input:
+      module: test/input
+      options:
+        api_key: CONFIG_SUMMARY_SECRET_SENTINEL
+        safe: SAFE_OPTION_SENTINEL
+        nested:
+          - - password: PIPELINE_SUMMARY_SECRET_SENTINEL
+              neighbor: SAFE_NESTED_OPTION_SENTINEL
+`)
+	if err := os.WriteFile(roots.config, []byte(configText), 0o600); err != nil {
+		t.Fatal(err)
+	}
+
+	result := runStateTest(t, roots, newStateTestHarness().options(), true, false, "bypass")
+	if result.code != 0 {
+		t.Fatalf("code=%d stderr=%q", result.code, result.stderr)
+	}
+	summaryRoot := filepath.Join(onlyChildDir(t, roots.debug), "summary")
+	for _, name := range []string{"effective-config.json", "resolved-pipeline.json"} {
+		contents, err := os.ReadFile(filepath.Join(summaryRoot, name))
+		if err != nil {
+			t.Fatal(err)
+		}
+		text := string(contents)
+		for _, secret := range []string{"CONFIG_SUMMARY_SECRET_SENTINEL", "PIPELINE_SUMMARY_SECRET_SENTINEL"} {
+			if strings.Contains(text, secret) {
+				t.Fatalf("%s contains %q: %s", name, secret, text)
+			}
+		}
+		for _, retained := range []string{"[REDACTED]", "SAFE_OPTION_SENTINEL", "SAFE_NESTED_OPTION_SENTINEL"} {
+			if !strings.Contains(text, retained) {
+				t.Fatalf("%s does not contain %q: %s", name, retained, text)
+			}
+		}
+	}
+}
+
+func TestRunUsesOneInjectedIdentityForDebugOutputAndManifest(t *testing.T) {
+	roots := newStateTestRoots(t)
+	harness := newStateTestHarness()
+	opts := harness.options()
+	const runID = "run-1000000000-11111111111111111111111111111111"
+	opts.RunIDGenerator = func(time.Time) (string, error) { return runID, nil }
+
+	result := runStateTest(t, roots, opts, true, false, "bypass")
+	if result.code != 0 {
+		t.Fatalf("code=%d stderr=%q", result.code, result.stderr)
+	}
+	outputPath := filepath.Join(roots.output, runID)
+	debugPath := filepath.Join(roots.debug, runID)
+	assertFile(t, filepath.Join(outputPath, "result.json"))
+	assertFile(t, filepath.Join(debugPath, "summary", "run-manifest.json"))
+	if !strings.Contains(result.stdout, "output="+outputPath) || !strings.Contains(result.stdout, "debug="+debugPath) {
+		t.Fatalf("stdout=%q, want shared run identity", result.stdout)
+	}
+	data, err := os.ReadFile(filepath.Join(debugPath, "summary", "run-manifest.json"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	var manifest artifacts.RunManifest
+	if err := json.Unmarshal(data, &manifest); err != nil {
+		t.Fatal(err)
+	}
+	if manifest.RunID != runID {
+		t.Fatalf("manifest run ID = %q, want %q", manifest.RunID, runID)
+	}
+	wantStartedAt := time.Unix(1, 0).UTC()
+	if manifest.StartedAt == nil || !manifest.StartedAt.Equal(wantStartedAt) {
+		t.Fatalf("manifest started at = %v, want %v", manifest.StartedAt, wantStartedAt)
+	}
+	var invocation debugbundle.Invocation
+	readStateTestSummaryJSON(t, debugPath, "invocation.json", &invocation)
+	if invocation.RunID != runID || !invocation.StartedAt.Equal(wantStartedAt) {
+		t.Fatalf("debug invocation identity = %#v, want run %q at %v", invocation, runID, wantStartedAt)
+	}
+	report := readStateTestRunReport(t, debugPath)
+	if !report.Succeeded || report.RunID != runID || report.PipelineID != "sample" || report.OutputPath != outputPath || report.DebugPath != debugPath || report.OutputCount != 1 || report.RejectedCount != 0 || report.WarningCount != 0 || report.ValidationStatus != "approved" {
+		t.Fatalf("success report = %#v", report)
+	}
+	if !strings.Contains(result.stdout, "outputs=1 rejected=0") {
+		t.Fatalf("stdout=%q, want report counts", result.stdout)
+	}
+}
+
+func TestRunWritesTerminalArtifactsForResolutionPipelineAndOutputFailures(t *testing.T) {
+	for _, tc := range []struct {
+		name             string
+		pipelineID       string
+		wantError        string
+		wantOutputs      int
+		wantValidation   string
+		configureFailure func(*testing.T, stateTestRoots, *stateTestHarness)
+	}{
+		{name: "resolution", pipelineID: "missing", wantError: `pipeline "missing"`},
+		{name: "pipeline", pipelineID: "sample", wantError: "synthetic extraction failure", wantValidation: "failed", configureFailure: func(_ *testing.T, _ stateTestRoots, h *stateTestHarness) {
+			h.extractErr = errors.New("synthetic extraction failure")
+		}},
+		{name: "output", pipelineID: "sample", wantError: "create output parent", wantOutputs: 1, wantValidation: "approved", configureFailure: func(t *testing.T, roots stateTestRoots, _ *stateTestHarness) {
+			if err := os.WriteFile(roots.output, []byte("not a directory"), 0o600); err != nil {
+				t.Fatal(err)
+			}
+		}},
+	} {
+		t.Run(tc.name, func(t *testing.T) {
+			roots := newStateTestRoots(t)
+			harness := newStateTestHarness()
+			if tc.configureFailure != nil {
+				tc.configureFailure(t, roots, harness)
+			}
+			opts := harness.options()
+			var stdout, stderr bytes.Buffer
+			args := []string{"run", tc.pipelineID, "--config", roots.config, "--input", roots.input, "--chunk_cache", "bypass", "--debug"}
+			code := RunWithOptions(args, &stdout, &stderr, opts)
+			if code != 1 || !strings.Contains(stderr.String(), tc.wantError) {
+				t.Fatalf("code=%d stderr=%q", code, stderr.String())
+			}
+			bundlePath := onlyChildDir(t, roots.debug)
+			runID := filepath.Base(bundlePath)
+			report := readStateTestRunReport(t, bundlePath)
+			if report.Succeeded || report.RunID != runID || report.PipelineID != tc.pipelineID || report.OutputPath != filepath.Join(roots.output, runID) || report.DebugPath != bundlePath || report.OutputCount != tc.wantOutputs || report.RejectedCount != 0 || report.WarningCount != 0 || report.ValidationStatus != tc.wantValidation {
+				t.Fatalf("failure report = %#v", report)
+			}
+			errorLog, err := os.ReadFile(filepath.Join(bundlePath, "summary", "error.log"))
+			if err != nil || !strings.Contains(string(errorLog), tc.wantError) {
+				t.Fatalf("error log = %q, %v", errorLog, err)
+			}
+		})
+	}
+}
+
+func TestRunRetainsPartialPipelineOutcomeInFailureSummary(t *testing.T) {
+	roots := newStateTestRoots(t)
+	harness := newStateTestHarness()
+	harness.chunkWarnings = []contracts.Warning{{Scope: "chunk", ReasonCode: "partial-warning", Message: "warning retained before failure"}}
+	harness.extractErr = errors.New("synthetic partial pipeline failure")
+
+	result := runStateTest(t, roots, harness.options(), true, true, "bypass")
+	if result.code != 1 {
+		t.Fatalf("code=%d stderr=%q", result.code, result.stderr)
+	}
+	bundlePath := onlyChildDir(t, roots.debug)
+	report := readStateTestRunReport(t, bundlePath)
+	if report.Succeeded || report.OutputCount != 0 || report.RejectedCount != 0 || report.WarningCount != 1 || report.ValidationStatus != "failed" {
+		t.Fatalf("partial failure report = %#v", report)
+	}
+
+	var manifest artifacts.RunManifest
+	readStateTestSummaryJSON(t, bundlePath, "run-manifest.json", &manifest)
+	if manifest.RunID != report.RunID || manifest.PipelineID != "sample" || manifest.ValidationStatus != "failed" {
+		t.Fatalf("partial manifest = %#v", manifest)
+	}
+	var warnings []contracts.Warning
+	readStateTestSummaryJSON(t, bundlePath, "warnings.json", &warnings)
+	if len(warnings) != 1 || warnings[0].ReasonCode != "partial-warning" {
+		t.Fatalf("partial warnings = %#v", warnings)
+	}
+	var events []pipeline.CheckpointEvent
+	readStateTestSummaryJSON(t, bundlePath, "checkpoint-events.json", &events)
+	if len(events) == 0 || events[0].Stage != "source" {
+		t.Fatalf("partial checkpoint events = %#v, want retained source decision", events)
+	}
+	var chunkPlan artifacts.ChunkPlanSummary
+	readStateTestSummaryJSON(t, bundlePath, "chunk-plan.json", &chunkPlan)
+	if chunkPlan.Mode != "bypass" || chunkPlan.ValidationStatus == "not_run" {
+		t.Fatalf("partial chunk plan = %#v", chunkPlan)
+	}
+}
+
+func TestRunTerminalPersistenceFailuresDoNotRecurseOrHidePrimaryError(t *testing.T) {
+	for _, tc := range []struct {
+		name          string
+		reportErr     error
+		errorLogErr   error
+		wantSecondary string
+	}{
+		{name: "run report", reportErr: errors.New("injected run report failure"), wantSecondary: "injected run report failure"},
+		{name: "error log", errorLogErr: errors.New("injected error log failure"), wantSecondary: "injected error log failure"},
+	} {
+		t.Run(tc.name, func(t *testing.T) {
+			roots := newStateTestRoots(t)
+			harness := newStateTestHarness()
+			harness.extractErr = errors.New("primary pipeline failure")
+			opts := harness.options()
+			var terminal *recordingTerminalWriter
+			opts.DebugTerminalFactory = func(delegate *debugbundle.SummaryWriter) DebugTerminalWriter {
+				terminal = &recordingTerminalWriter{delegate: delegate, reportErr: tc.reportErr, errorLogErr: tc.errorLogErr}
+				return terminal
+			}
+
+			result := runStateTest(t, roots, opts, true, false, "bypass")
+			if result.code != 1 {
+				t.Fatalf("code=%d stderr=%q", result.code, result.stderr)
+			}
+			if terminal == nil {
+				t.Fatal("terminal writer was not constructed")
+			}
+			if terminal.reportCalls != 1 || terminal.errorLogCalls != 1 {
+				t.Fatalf("terminal calls = report:%d error:%d", terminal.reportCalls, terminal.errorLogCalls)
+			}
+			primaryIndex := strings.Index(result.stderr, "primary pipeline failure")
+			secondaryIndex := strings.Index(result.stderr, tc.wantSecondary)
+			debugIndex := strings.Index(result.stderr, "debug=")
+			if primaryIndex < 0 || secondaryIndex <= primaryIndex || debugIndex <= secondaryIndex {
+				t.Fatalf("stderr order = %q", result.stderr)
+			}
+		})
+	}
+}
+
+func TestRunReportFailureOnSuccessIsTerminalizedWithoutRetry(t *testing.T) {
+	roots := newStateTestRoots(t)
+	opts := newStateTestHarness().options()
+	var terminal *recordingTerminalWriter
+	opts.DebugTerminalFactory = func(delegate *debugbundle.SummaryWriter) DebugTerminalWriter {
+		terminal = &recordingTerminalWriter{delegate: delegate, reportErr: errors.New("injected success report failure")}
+		return terminal
+	}
+
+	result := runStateTest(t, roots, opts, true, false, "bypass")
+	if result.code != 1 || !strings.Contains(result.stderr, "write debug run report") || !strings.Contains(result.stderr, "injected success report failure") {
+		t.Fatalf("code=%d stdout=%q stderr=%q", result.code, result.stdout, result.stderr)
+	}
+	if terminal == nil {
+		t.Fatal("terminal writer was not constructed")
+	}
+	if terminal.reportCalls != 1 || terminal.errorLogCalls != 1 {
+		t.Fatalf("terminal calls = report:%d error:%d", terminal.reportCalls, terminal.errorLogCalls)
+	}
+	if result.stdout != "" {
+		t.Fatalf("stdout=%q, want no success message", result.stdout)
+	}
+	bundlePath := onlyChildDir(t, roots.debug)
+	errorLog, err := os.ReadFile(filepath.Join(bundlePath, "summary", "error.log"))
+	if err != nil || !strings.Contains(string(errorLog), "injected success report failure") {
+		t.Fatalf("error log = %q, %v", errorLog, err)
+	}
+}
+
+func TestRunWithoutDebugDoesNotUseTerminalSummaryWriter(t *testing.T) {
+	roots := newStateTestRoots(t)
+	harness := newStateTestHarness()
+	harness.extractErr = errors.New("non-debug pipeline failure")
+	opts := harness.options()
+	factoryCalls := 0
+	opts.DebugTerminalFactory = func(delegate *debugbundle.SummaryWriter) DebugTerminalWriter {
+		factoryCalls++
+		return delegate
+	}
+
+	result := runStateTest(t, roots, opts, false, false, "bypass")
+	if result.code != 1 || !strings.Contains(result.stderr, "non-debug pipeline failure") {
+		t.Fatalf("code=%d stderr=%q", result.code, result.stderr)
+	}
+	if factoryCalls != 0 {
+		t.Fatalf("terminal summary factory calls = %d, want 0", factoryCalls)
+	}
+	assertAbsent(t, roots.debug)
+}
+
+func TestRunRefusesExistingOutputDirectoryWithoutChangingIt(t *testing.T) {
+	roots := newStateTestRoots(t)
+	const runID = "run-1000000000-22222222222222222222222222222222"
+	runPath := filepath.Join(roots.output, runID)
+	if err := os.MkdirAll(filepath.Join(runPath, "nested"), 0o755); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.WriteFile(filepath.Join(runPath, "sentinel"), []byte("existing output"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.WriteFile(filepath.Join(runPath, "nested", "data"), []byte("preserve me"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	before := readTree(t, runPath)
+	opts := newStateTestHarness().options()
+	opts.RunIDGenerator = func(time.Time) (string, error) { return runID, nil }
+
+	result := runStateTest(t, roots, opts, true, false, "bypass")
+	if result.code != 1 || !strings.Contains(result.stderr, "output run directory") || !strings.Contains(result.stderr, "already exists") {
+		t.Fatalf("code=%d stderr=%q", result.code, result.stderr)
+	}
+	if after := readTree(t, runPath); !sameFiles(after, before) {
+		t.Fatalf("existing output changed: before=%v after=%v", before, after)
+	}
+	bundlePath := filepath.Join(roots.debug, runID)
+	report := readStateTestRunReport(t, bundlePath)
+	if report.Succeeded || report.RunID != runID || report.OutputPath != runPath || report.DebugPath != bundlePath || report.OutputCount != 1 || report.ValidationStatus != "approved" {
+		t.Fatalf("output collision report = %#v", report)
+	}
+	errorLog, err := os.ReadFile(filepath.Join(bundlePath, "summary", "error.log"))
+	if err != nil || !strings.Contains(string(errorLog), "already exists") {
+		t.Fatalf("output collision error log = %q, %v", errorLog, err)
+	}
+}
+
+func TestRepeatedRunIdentityCannotOverwriteFirstOutput(t *testing.T) {
+	roots := newStateTestRoots(t)
+	const runID = "run-1000000000-33333333333333333333333333333333"
+	harness := newStateTestHarness()
+	opts := harness.options()
+	opts.RunIDGenerator = func(time.Time) (string, error) { return runID, nil }
+
+	first := runStateTest(t, roots, opts, false, false, "bypass")
+	if first.code != 0 {
+		t.Fatalf("first code=%d stderr=%q", first.code, first.stderr)
+	}
+	runPath := filepath.Join(roots.output, runID)
+	before := readTree(t, runPath)
+	second := runStateTest(t, roots, opts, false, false, "bypass")
+	if second.code != 1 || !strings.Contains(second.stderr, "already exists") {
+		t.Fatalf("second code=%d stderr=%q", second.code, second.stderr)
+	}
+	if after := readTree(t, runPath); !sameFiles(after, before) {
+		t.Fatalf("first output changed: before=%v after=%v", before, after)
+	}
+}
+
+func TestRunRefusesExistingDebugBundleWithoutChangingIt(t *testing.T) {
+	roots := newStateTestRoots(t)
+	const runID = "run-1000000000-44444444444444444444444444444444"
+	bundlePath := filepath.Join(roots.debug, runID)
+	if err := os.MkdirAll(bundlePath, 0o700); err != nil {
+		t.Fatal(err)
+	}
+	sentinelPath := filepath.Join(bundlePath, "sentinel")
+	if err := os.WriteFile(sentinelPath, []byte("existing debug"), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	opts := newStateTestHarness().options()
+	opts.RunIDGenerator = func(time.Time) (string, error) { return runID, nil }
+
+	result := runStateTest(t, roots, opts, true, false, "bypass")
+	if result.code != 1 || !strings.Contains(result.stderr, "debug bundle") || !strings.Contains(result.stderr, "already exists") {
+		t.Fatalf("code=%d stderr=%q", result.code, result.stderr)
+	}
+	if got, err := os.ReadFile(sentinelPath); err != nil || string(got) != "existing debug" {
+		t.Fatalf("sentinel = %q, %v", got, err)
+	}
+	assertAbsent(t, roots.output)
+}
+
+func TestRunIDGenerationFailurePrecedesDebugAllocation(t *testing.T) {
+	roots := newStateTestRoots(t)
+	opts := newStateTestHarness().options()
+	opts.RunIDGenerator = func(time.Time) (string, error) { return "", errors.New("random source unavailable") }
+
+	result := runStateTest(t, roots, opts, true, false, "bypass")
+	if result.code != 1 || !strings.Contains(result.stderr, "generate run ID: random source unavailable") {
+		t.Fatalf("code=%d stderr=%q", result.code, result.stderr)
+	}
+	assertAbsent(t, roots.debug)
+	assertAbsent(t, roots.output)
+}
+
+func TestRunRejectsUnsafeGeneratedIdentityBeforePathUse(t *testing.T) {
+	roots := newStateTestRoots(t)
+	opts := newStateTestHarness().options()
+	opts.RunIDGenerator = func(time.Time) (string, error) { return "../outside", nil }
+
+	result := runStateTest(t, roots, opts, true, false, "bypass")
+	if result.code != 1 || !strings.Contains(result.stderr, "invalid generated run ID") || !strings.Contains(result.stderr, "one safe path component") {
+		t.Fatalf("code=%d stderr=%q", result.code, result.stderr)
+	}
+	assertAbsent(t, roots.debug)
+	assertAbsent(t, roots.output)
+}
+
+type stateTestRoots struct{ config, input, output, plans, checkpoints, debug string }
+
+func newStateTestRoots(t *testing.T) stateTestRoots {
+	t.Helper()
+	base := t.TempDir()
+	roots := stateTestRoots{input: filepath.Join(base, "input.txt"), output: filepath.Join(base, "output"), plans: filepath.Join(base, "plans"), checkpoints: filepath.Join(base, "checkpoints"), debug: filepath.Join(base, "debug")}
+	if err := os.WriteFile(roots.input, []byte("application content Bearer secretvalue sk-secretvalue"), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	roots.config = filepath.Join(base, "config.yml")
+	config := fmt.Sprintf("version: 4\noutput:\n  directory: %q\ncache:\n  chunk_plans:\n    directory: %q\n    mode: auto\n  checkpoints:\n    enabled: true\n    directory: %q\ndebug:\n  directory: %q\npipelines:\n  sample:\n    input: test/input\n    chunk: test/chunk\n    artifacts:\n      items:\n        extract: test/extract\n        merge: test/merge\n        normalize: test/normalize\n    output: test/output\n", roots.output, roots.plans, roots.checkpoints, roots.debug)
+	if err := os.WriteFile(roots.config, []byte(config), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	return roots
+}
+
+type stateTestResult struct {
+	code           int
+	stdout, stderr string
+}
+
+func runStateTest(t *testing.T, roots stateTestRoots, opts Options, debug, resume bool, mode string) stateTestResult {
+	t.Helper()
+	args := []string{"run", "sample", "--config", roots.config, "--input", roots.input, "--chunk_cache", mode}
+	if debug {
+		args = append(args, "--debug")
+	}
+	if resume {
+		args = append(args, "--resume")
+	}
+	var stdout, stderr bytes.Buffer
+	return stateTestResult{RunWithOptions(args, &stdout, &stderr, opts), stdout.String(), stderr.String()}
+}
+
+func assertStateTestOutput(t *testing.T, root string) {
+	t.Helper()
+	output := onlyChildDir(t, root)
+	data, err := os.ReadFile(filepath.Join(output, "result.json"))
+	if err != nil || string(data) != "{\"ok\":true}\n" {
+		t.Fatalf("output = %q, %v", data, err)
+	}
+}
+
+func onlyChildDir(t *testing.T, root string) string {
+	t.Helper()
+	entries, err := os.ReadDir(root)
+	if err != nil {
+		t.Fatal(err)
+	}
+	var dirs []string
+	for _, entry := range entries {
+		if entry.IsDir() {
+			dirs = append(dirs, filepath.Join(root, entry.Name()))
+		}
+	}
+	if len(dirs) != 1 {
+		t.Fatalf("directories in %q = %v, want one", root, dirs)
+	}
+	return dirs[0]
+}
+
+func assertFile(t *testing.T, path string) {
+	t.Helper()
+	if info, err := os.Stat(path); err != nil || info.IsDir() {
+		t.Fatalf("file %q: %v", path, err)
+	}
+}
+func assertAbsent(t *testing.T, path string) {
+	t.Helper()
+	if _, err := os.Stat(path); !os.IsNotExist(err) {
+		t.Fatalf("%q exists or stat failed: %v", path, err)
+	}
+}
+func assertAnyFile(t *testing.T, root string) {
+	t.Helper()
+	if text := readAllFiles(t, root); text == "" {
+		t.Fatalf("no files under %q", root)
+	}
+}
+
+func readAllFiles(t *testing.T, root string) string {
+	t.Helper()
+	var content strings.Builder
+	if err := filepath.WalkDir(root, func(path string, entry os.DirEntry, err error) error {
+		if err != nil {
+			return err
+		}
+		if entry.IsDir() {
+			return nil
+		}
+		data, err := os.ReadFile(path)
+		if err != nil {
+			return err
+		}
+		content.Write(data)
+		return nil
+	}); err != nil {
+		t.Fatal(err)
+	}
+	return content.String()
+}
+
+func readStateTestRunReport(t *testing.T, bundlePath string) debugbundle.RunReport {
+	t.Helper()
+	var report debugbundle.RunReport
+	readStateTestSummaryJSON(t, bundlePath, "run-report.json", &report)
+	return report
+}
+
+func readStateTestSummaryJSON(t *testing.T, bundlePath, name string, target any) {
+	t.Helper()
+	data, err := os.ReadFile(filepath.Join(bundlePath, "summary", name))
+	if err != nil {
+		t.Fatal(err)
+	}
+	if err := json.Unmarshal(data, target); err != nil {
+		t.Fatal(err)
+	}
+}
+
+func assertRestrictedTree(t *testing.T, root string) {
+	t.Helper()
+	if runtime.GOOS == "windows" {
+		return
+	}
+	if err := filepath.Walk(root, func(path string, info os.FileInfo, err error) error {
+		if err != nil {
+			return err
+		}
+		want := os.FileMode(0o600)
+		if info.IsDir() {
+			want = 0o700
+		}
+		if info.Mode().Perm() != want {
+			return fmt.Errorf("%s has mode %o, want %o", path, info.Mode().Perm(), want)
+		}
+		return nil
+	}); err != nil {
+		t.Fatal(err)
+	}
+}
+
+func readTree(t *testing.T, root string) map[string][]byte {
+	t.Helper()
+	files := map[string][]byte{}
+	if err := filepath.WalkDir(root, func(path string, entry os.DirEntry, err error) error {
+		if err != nil {
+			return err
+		}
+		if entry.IsDir() {
+			return nil
+		}
+		data, err := os.ReadFile(path)
+		if err != nil {
+			return err
+		}
+		relative, err := filepath.Rel(root, path)
+		if err != nil {
+			return err
+		}
+		files[relative] = data
+		return nil
+	}); err != nil {
+		t.Fatal(err)
+	}
+	return files
+}
+func sameFiles(left, right map[string][]byte) bool {
+	if len(left) != len(right) {
+		return false
+	}
+	for path, data := range left {
+		if !bytes.Equal(data, right[path]) {
+			return false
+		}
+	}
+	return true
+}
+
+type stateTestHarness struct {
+	mu                       sync.Mutex
+	chunkCalls, extractCalls int
+	runIDCalls               uint64
+	extractErr               error
+	chunkWarnings            []contracts.Warning
+	moduleProfiles           []string
+	sessionIDs               []string
+	outputWarnings           []contracts.Warning
+	includeWarnings          bool
+}
+
+func newStateTestHarness() *stateTestHarness { return &stateTestHarness{} }
+func (h *stateTestHarness) options() Options {
+	registries := pipeline.Registries{Inputs: pipeline.NewInputAdapterRegistry(), Chunkers: pipeline.NewChunkerRegistry(), ArtifactCodecs: pipeline.NewArtifactCodecRegistry(), Extractors: pipeline.NewExtractorRegistry(), Mergers: pipeline.NewMergerRegistry(), Normalizers: pipeline.NewNormalizerRegistry(), Validators: pipeline.NewValidatorRegistry(), ValidatorChains: pipeline.NewValidatorChainRegistry(), Outputs: pipeline.NewOutputEncoderRegistry()}
+	if err := pipeline.RegisterArtifactCodec(registries.ArtifactCodecs, stateTestCodec{}); err != nil {
+		panic(err)
+	}
+	if err := registries.Inputs.RegisterBuilderWithSpec(pipeline.ModuleSpec{Key: "test/input", Stage: pipeline.StageInput, ExecutionClass: contracts.ExecutionClassDeterministic, Provides: []string{"source"}}, func(map[string]any) error { return nil }, func(pipeline.BuildRequest) (contracts.InputAdapter, error) { return stateTestInput{}, nil }); err != nil {
+		panic(err)
+	}
+	if err := registries.Chunkers.RegisterBuilderWithSpec(pipeline.ModuleSpec{Key: "test/chunk", Stage: pipeline.StageChunk, ExecutionClass: contracts.ExecutionClassLLMBacked, Requires: []string{"source"}, Provides: []string{"chunks"}, ReferenceSlots: []contracts.ReferenceSlot{{Name: "cache-reference"}}}, func(map[string]any) error { return nil }, func(pipeline.BuildRequest) (contracts.Chunker, error) { return stateTestChunker{h}, nil }); err != nil {
+		panic(err)
+	}
+	if err := pipeline.RegisterExtractor(registries.Extractors, pipeline.ModuleSpec{Key: "test/extract", Stage: pipeline.StageExtract, ExecutionClass: contracts.ExecutionClassLLMBacked, Requires: []string{"chunks"}, Provides: []string{"artifact"}, ArtifactKind: stateTestArtifactKind}, func() (contracts.Extractor[stateTestArtifact], error) { return stateTestExtractor{h}, nil }); err != nil {
+		panic(err)
+	}
+	if err := pipeline.RegisterMerger(registries.Mergers, pipeline.ModuleSpec{Key: "test/merge", Stage: pipeline.StageMerge, ExecutionClass: contracts.ExecutionClassLLMBacked, Requires: []string{"artifact"}, Provides: []string{"merged"}, ArtifactKind: stateTestArtifactKind}, func() (contracts.Merger[stateTestArtifact], error) { return stateTestMerger{harness: h}, nil }); err != nil {
+		panic(err)
+	}
+	if err := pipeline.RegisterNormalizer(registries.Normalizers, pipeline.ModuleSpec{Key: "test/normalize", Stage: pipeline.StageNormalize, ExecutionClass: contracts.ExecutionClassLLMBacked, Requires: []string{"merged"}, Provides: []string{"normalized"}, ArtifactKind: stateTestArtifactKind}, func() (contracts.Normalizer[stateTestArtifact], error) { return stateTestNormalizer{harness: h}, nil }); err != nil {
+		panic(err)
+	}
+	if err := registries.Outputs.RegisterWithSpec(pipeline.ModuleSpec{Key: "test/output", Stage: pipeline.StageOutput, ExecutionClass: contracts.ExecutionClassDeterministic, Requires: []string{"normalized"}, Provides: []string{"output"}}, func() (contracts.OutputEncoder, error) {
+		return stateTestOutput{harness: h, includeWarnings: h.includeWarnings}, nil
+	}); err != nil {
+		panic(err)
+	}
+	return Options{Catalog: catalogFromRegistries(registries), Registries: registries, LookupEnv: emptyLookup, Now: func() time.Time { return time.Unix(1, 0) }, RunIDGenerator: func(startedAt time.Time) (string, error) {
+		h.mu.Lock()
+		defer h.mu.Unlock()
+		h.runIDCalls++
+		return fmt.Sprintf("run-%d-%032x", startedAt.UnixNano(), h.runIDCalls), nil
+	}, UserCacheDir: func() (string, error) { return "", errors.New("unexpected user cache lookup") }, LLMClientFactory: func(context.Context, config.Config, string, LLMRuntimeOverrides) (contracts.StructuredLLMClient, []artifacts.LLMProfileManifest, error) {
+		return nil, nil, nil
+	}}
+}
+
+type stateTestInput struct{}
+
+func (stateTestInput) Key() string { return "test/input" }
+func (stateTestInput) Parse(_ context.Context, req contracts.ParseRequest) (*source.SourceDocument, error) {
+	doc := &source.SourceDocument{ID: "source", Kind: "text", Format: "text/plain", Units: []source.SourceUnit{{ID: 1, Kind: "text", Text: string(req.Raw), Ref: source.SourceRef{SourceID: "source", StartUnitID: 1, EndUnitID: 1}}}}
+	digest, err := source.DigestDocument(doc)
+	if err != nil {
+		return nil, err
+	}
+	doc.Digest = digest
+	return doc, nil
+}
+
+type stateTestChunker struct{ harness *stateTestHarness }
+
+func (stateTestChunker) Key() string                               { return "test/chunk" }
+func (stateTestChunker) ReferenceSlots() []contracts.ReferenceSlot { return nil }
+func (c stateTestChunker) Plan(_ context.Context, req contracts.ChunkRequest) (contracts.ChunkPlanResult, error) {
+	c.harness.mu.Lock()
+	c.harness.moduleProfiles = append(c.harness.moduleProfiles, req.LLMProfile)
+	c.harness.sessionIDs = append(c.harness.sessionIDs, req.SessionID)
+	c.harness.mu.Unlock()
+	c.harness.mu.Lock()
+	c.harness.chunkCalls++
+	c.harness.mu.Unlock()
+	return contracts.ChunkPlanResult{Plan: source.ChunkPlan{SourceDigest: req.Source.Digest, Ranges: []source.ChunkRange{{StartUnitID: 1, EndUnitID: 1}}}, Warnings: append([]contracts.Warning(nil), c.harness.chunkWarnings...)}, nil
+}
+
+const stateTestArtifactKind contracts.ArtifactKind = "test/artifact"
+
+type stateTestArtifact struct {
+	Value string `json:"value"`
+}
+type stateTestCodec struct{}
+
+func (stateTestCodec) Kind() contracts.ArtifactKind { return stateTestArtifactKind }
+func (stateTestCodec) Schema() contracts.ArtifactSchema {
+	return contracts.ArtifactSchema{ID: "test.artifact", Name: "test_artifact", Version: "v1", JSONSchema: []byte(`{"type":"object"}`)}
+}
+func (stateTestCodec) MediaType() string { return "application/json" }
+func (stateTestCodec) EncodeCandidate(v stateTestArtifact) ([]byte, error) {
+	return []byte(`{"value":"ok"}`), nil
+}
+func (stateTestCodec) Encode(v stateTestArtifact) ([]byte, error) {
+	return []byte(`{"value":"ok"}`), nil
+}
+func (stateTestCodec) Decode([]byte) (stateTestArtifact, error) {
+	return stateTestArtifact{Value: "ok"}, nil
+}
+
+type stateTestExtractor struct{ harness *stateTestHarness }
+
+func (stateTestExtractor) Key() string                               { return "test/extract" }
+func (stateTestExtractor) ReferenceSlots() []contracts.ReferenceSlot { return nil }
+func (e stateTestExtractor) Extract(_ context.Context, req contracts.TypedExtractionRequest) (contracts.TypedExtractionResult[stateTestArtifact], error) {
+	e.harness.mu.Lock()
+	defer e.harness.mu.Unlock()
+	e.harness.extractCalls++
+	e.harness.moduleProfiles = append(e.harness.moduleProfiles, req.LLMProfile)
+	e.harness.sessionIDs = append(e.harness.sessionIDs, req.SessionID)
+	if e.harness.extractErr != nil {
+		return contracts.TypedExtractionResult[stateTestArtifact]{}, e.harness.extractErr
+	}
+	return contracts.TypedExtractionResult[stateTestArtifact]{Value: stateTestArtifact{Value: "ok"}}, nil
+}
+
+type stateTestMerger struct{ harness *stateTestHarness }
+
+func (stateTestMerger) Key() string { return "test/merge" }
+func (m stateTestMerger) Merge(_ context.Context, req contracts.TypedMergeRequest[stateTestArtifact]) (contracts.TypedMergeResult[stateTestArtifact], error) {
+	m.harness.mu.Lock()
+	m.harness.moduleProfiles = append(m.harness.moduleProfiles, req.LLMProfile)
+	m.harness.sessionIDs = append(m.harness.sessionIDs, req.SessionID)
+	m.harness.mu.Unlock()
+	return contracts.TypedMergeResult[stateTestArtifact]{Value: req.ExtractOutputs[0].Value}, nil
+}
+
+type stateTestNormalizer struct{ harness *stateTestHarness }
+
+func (stateTestNormalizer) Key() string                               { return "test/normalize" }
+func (stateTestNormalizer) ReferenceSlots() []contracts.ReferenceSlot { return nil }
+func (n stateTestNormalizer) Normalize(_ context.Context, req contracts.TypedNormalizeRequest[stateTestArtifact]) (contracts.TypedNormalizeResult[stateTestArtifact], error) {
+	n.harness.mu.Lock()
+	n.harness.moduleProfiles = append(n.harness.moduleProfiles, req.LLMProfile)
+	n.harness.sessionIDs = append(n.harness.sessionIDs, req.SessionID)
+	n.harness.mu.Unlock()
+	return contracts.TypedNormalizeResult[stateTestArtifact]{Value: req.MergeOutput.Value}, nil
+}
+
+type stateTestOutput struct {
+	harness         *stateTestHarness
+	includeWarnings bool
+}
+
+func (o stateTestOutput) Key() string { return "test/output" }
+func (o stateTestOutput) Encode(_ context.Context, req contracts.OutputRequest) (contracts.OutputResult, error) {
+	o.harness.mu.Lock()
+	o.harness.outputWarnings = append([]contracts.Warning(nil), req.Warnings...)
+	o.harness.mu.Unlock()
+	data := []byte("{\"ok\":true}\n")
+	if o.includeWarnings && len(req.Warnings) > 0 {
+		data = []byte(fmt.Sprintf("{\"ok\":true,\"warnings\":%q}\n", req.Warnings[0].ReasonCode))
+	}
+	return contracts.OutputResult{Files: []contracts.OutputFile{{Name: "result.json", Bytes: data}}}, nil
+}
+
+type failingDebugRecorder struct{}
+
+func (failingDebugRecorder) Enabled() bool                   { return true }
+func (failingDebugRecorder) WriteJSON(string, any) error     { return errors.New("trace unavailable") }
+func (failingDebugRecorder) WriteBytes(string, []byte) error { return errors.New("trace unavailable") }
+
+type recordingTerminalWriter struct {
+	delegate                   DebugTerminalWriter
+	reportErr, errorLogErr     error
+	reportCalls, errorLogCalls int
+}
+
+func (w *recordingTerminalWriter) WriteRunReport(report debugbundle.RunReport) error {
+	w.reportCalls++
+	if w.reportErr != nil {
+		return w.reportErr
+	}
+	return w.delegate.WriteRunReport(report)
+}
+
+func (w *recordingTerminalWriter) WriteError(message string) error {
+	w.errorLogCalls++
+	if w.errorLogErr != nil {
+		return w.errorLogErr
+	}
+	return w.delegate.WriteError(message)
+}

internal/core/artifacts/artifacts.go

@@ -1,91 +1,173 @@
 package artifacts
 
 import (
-	"encoding/json"
+	"strings"
 	"time"
-
-	"gitea.maximumdirect.net/eric/notarius/internal/core/source"
 )
 
-type ArtifactCandidate struct {
-	Index         int                `json:"index"`
-	ExtractorKey  string             `json:"extractor_key"`
-	ArtifactType  string             `json:"artifact_type"`
-	SchemaVersion string             `json:"schema_version"`
-	Payload       json.RawMessage    `json:"payload"`
-	SourceRefs    []source.SourceRef `json:"source_refs,omitempty"`
-	Metadata      map[string]any     `json:"metadata,omitempty"`
-}
-
-type Artifact struct {
-	ExtractorKey  string             `json:"extractor_key"`
-	ArtifactType  string             `json:"artifact_type"`
-	SchemaVersion string             `json:"schema_version"`
-	Payload       json.RawMessage    `json:"payload"`
-	SourceRefs    []source.SourceRef `json:"source_refs,omitempty"`
-	Metadata      map[string]any     `json:"metadata,omitempty"`
-}
-
-type RejectedArtifact struct {
-	Candidate     ArtifactCandidate `json:"candidate"`
-	ValidatorName string            `json:"validator_name"`
-	ReasonCode    string            `json:"reason_code"`
-	Message       string            `json:"message"`
-}
-
 type ArtifactLaneManifest struct {
+	StepID     string         `json:"step_id,omitempty"`
 	ID         string         `json:"id"`
 	Extractor  string         `json:"extractor"`
 	Merger     string         `json:"merger"`
 	Normalizer string         `json:"normalizer"`
-	Validators []string       `json:"validators,omitempty"`
 	Metadata   map[string]any `json:"metadata,omitempty"`
 }
 
+type ValidatorChainManifest struct {
+	Stage      string              `json:"stage"`
+	LaneID     string              `json:"lane_id,omitempty"`
+	ModuleKey  string              `json:"module_key"`
+	Validators []ValidatorManifest `json:"validators"`
+}
+
+type ValidatorManifest struct {
+	Key            string `json:"key"`
+	ExecutionClass string `json:"execution_class"`
+}
+
 type LLMProfileManifest struct {
-	ID       string `json:"id"`
+	ID              string `json:"id"`
-	Provider string `json:"provider,omitempty"`
+	Provider        string `json:"provider,omitempty"`
-	Model    string `json:"model,omitempty"`
+	Model           string `json:"model,omitempty"`
+	BackendID       string `json:"backend_id,omitempty"`
+	ReasoningEffort string `json:"reasoning_effort,omitempty"`
+}
+
+// Normalized returns the canonical representation used for manifest identity
+// and publication.
+func (profile LLMProfileManifest) Normalized() LLMProfileManifest {
+	profile.ID = strings.TrimSpace(profile.ID)
+	profile.Provider = strings.TrimSpace(profile.Provider)
+	profile.Model = strings.TrimSpace(profile.Model)
+	profile.BackendID = strings.TrimSpace(profile.BackendID)
+	profile.ReasoningEffort = strings.TrimSpace(profile.ReasoningEffort)
+	return profile
+}
+
+// IdentityKey returns an opaque, deterministic key for the effective profile.
+func (profile LLMProfileManifest) IdentityKey() string {
+	profile = profile.Normalized()
+	return profile.ID + "\x00" +
+		profile.Provider + "\x00" +
+		profile.Model + "\x00" +
+		profile.BackendID + "\x00" +
+		profile.ReasoningEffort
+}
+
+type ReferenceProvenance struct {
+	Stage            string `json:"stage,omitempty"`
+	StepID           string `json:"step_id,omitempty"`
+	LaneID           string `json:"lane_id,omitempty"`
+	SlotName         string `json:"slot_name"`
+	OriginType       string `json:"origin_type"`
+	OriginURI        string `json:"origin_uri,omitempty"`
+	Digest           string `json:"digest,omitempty"`
+	MediaType        string `json:"media_type,omitempty"`
+	SizeBytes        int64  `json:"size_bytes,omitempty"`
+	BindingSource    string `json:"binding_source,omitempty"`
+	ArtifactKind     string `json:"artifact_kind,omitempty"`
+	SchemaID         string `json:"schema_id,omitempty"`
+	SchemaName       string `json:"schema_name,omitempty"`
+	SchemaVersion    string `json:"schema_version,omitempty"`
+	SchemaDigest     string `json:"schema_digest,omitempty"`
+	ProducerPipeline string `json:"producer_pipeline_id,omitempty"`
+	ProducerStep     string `json:"producer_step_id,omitempty"`
+	ProducerLane     string `json:"producer_lane_id,omitempty"`
+	ProducerModule   string `json:"producer_module_key,omitempty"`
+}
+
+type OutputSchemaProvenance struct {
+	ID      string `json:"id,omitempty"`
+	Name    string `json:"name,omitempty"`
+	Version string `json:"version,omitempty"`
+}
+
+type NormalizedOutputManifest struct {
+	StepID    string                 `json:"step_id,omitempty"`
+	LaneID    string                 `json:"lane_id"`
+	ModuleKey string                 `json:"module_key,omitempty"`
+	SourceID  string                 `json:"source_id,omitempty"`
+	MediaType string                 `json:"media_type,omitempty"`
+	Schema    OutputSchemaProvenance `json:"schema,omitempty"`
+}
+
+type RejectedOutputManifest struct {
+	Stage                  string `json:"stage"`
+	StepID                 string `json:"step_id,omitempty"`
+	LaneID                 string `json:"lane_id,omitempty"`
+	ModuleKey              string `json:"module_key,omitempty"`
+	ChunkID                string `json:"chunk_id,omitempty"`
+	ChunkIndex             int    `json:"chunk_index,omitempty"`
+	ValidatorName          string `json:"validator_name,omitempty"`
+	ReasonCode             string `json:"reason_code,omitempty"`
+	Message                string `json:"message,omitempty"`
+	AttemptCount           int    `json:"attempt_count,omitempty"`
+	DiagnosticArtifactPath string `json:"diagnostic_artifact_path,omitempty"`
+}
+
+type CheckpointDecisionManifest struct {
+	Stage      string `json:"stage"`
+	StepID     string `json:"step_id,omitempty"`
+	LaneID     string `json:"lane_id,omitempty"`
+	ModuleKey  string `json:"module_key,omitempty"`
+	Category   string `json:"category"`
+	ReasonCode string `json:"reason_code,omitempty"`
+	Detail     string `json:"detail,omitempty"`
+}
+
+type ChunkPlanManifest struct {
+	Mode                string                `json:"mode"`
+	Action              string                `json:"action,omitempty"`
+	SourceDigest        string                `json:"source_digest,omitempty"`
+	PlanDigest          string                `json:"plan_digest,omitempty"`
+	PlanSchemaVersion   string                `json:"plan_schema_version,omitempty"`
+	RequestedModule     string                `json:"requested_module"`
+	ProducerInputModule string                `json:"producer_input_module,omitempty"`
+	ProducerModule      string                `json:"producer_module,omitempty"`
+	ProducerLLMProfile  string                `json:"producer_llm_profile,omitempty"`
+	ProducerReferences  []ReferenceProvenance `json:"producer_references,omitempty"`
+	ProducerMetadata    map[string]any        `json:"producer_metadata,omitempty"`
+	CreatedAt           *time.Time            `json:"created_at,omitempty"`
+}
+
+// ChunkPlanSummary is deliberately limited to cache and validation decisions.
+// It must never contain plan units, source content, annotations, or model I/O.
+type ChunkPlanSummary struct {
+	Mode              string `json:"mode"`
+	SourceDigest      string `json:"source_digest,omitempty"`
+	CandidateDigest   string `json:"candidate_digest,omitempty"`
+	RequestedModule   string `json:"requested_module"`
+	LookupStatus      string `json:"lookup_status"`
+	LookupReason      string `json:"lookup_reason,omitempty"`
+	Action            string `json:"action,omitempty"`
+	ValidationStatus  string `json:"validation_status"`
+	PublicationStatus string `json:"publication_status"`
 }
 
 type RunManifest struct {
-	RunID            string                 `json:"run_id,omitempty"`
+	RunID               string                       `json:"run_id,omitempty"`
-	PipelineID       string                 `json:"pipeline_id,omitempty"`
+	PipelineID          string                       `json:"pipeline_id,omitempty"`
-	PipelineDigest   string                 `json:"pipeline_digest,omitempty"`
+	PipelineDigest      string                       `json:"pipeline_digest,omitempty"`
-	InputModule      string                 `json:"input_module,omitempty"`
+	InputModule         string                       `json:"input_module,omitempty"`
-	Chunker          string                 `json:"chunker,omitempty"`
+	Chunker             string                       `json:"chunker,omitempty"`
-	SourceDigests    []string               `json:"source_digests,omitempty"`
+	ChunkPlan           *ChunkPlanManifest           `json:"chunk_plan,omitempty"`
-	Extractors       []string               `json:"extractors,omitempty"`
+	SourceDigests       []string                     `json:"source_digests,omitempty"`
-	Merger           string                 `json:"merger,omitempty"`
+	Extractors          []string                     `json:"extractors,omitempty"`
-	Normalizer       string                 `json:"normalizer,omitempty"`
+	Merger              string                       `json:"merger,omitempty"`
-	OutputEncoder    string                 `json:"output_encoder,omitempty"`
+	Normalizer          string                       `json:"normalizer,omitempty"`
-	ArtifactLanes    []ArtifactLaneManifest `json:"artifact_lanes,omitempty"`
+	OutputEncoder       string                       `json:"output_encoder,omitempty"`
-	LLMProfiles      []LLMProfileManifest   `json:"llm_profiles,omitempty"`
+	ModuleMetadata      map[string]map[string]any    `json:"module_metadata,omitempty"`
-	SchemaVersion    string                 `json:"schema_version,omitempty"`
+	ArtifactLanes       []ArtifactLaneManifest       `json:"artifact_lanes,omitempty"`
-	ValidationStatus string                 `json:"validation_status,omitempty"`
+	ValidatorChains     []ValidatorChainManifest     `json:"validator_chains,omitempty"`
-	StartedAt        *time.Time             `json:"started_at,omitempty"`
+	References          []ReferenceProvenance        `json:"references,omitempty"`
-	CompletedAt      *time.Time             `json:"completed_at,omitempty"`
+	NormalizedOutputs   []NormalizedOutputManifest   `json:"normalized_outputs,omitempty"`
-}
+	RejectedOutputs     []RejectedOutputManifest     `json:"rejected_outputs,omitempty"`
-
+	CheckpointDecisions []CheckpointDecisionManifest `json:"checkpoint_decisions,omitempty"`
-func ArtifactFromCandidate(candidate ArtifactCandidate) Artifact {
+	LLMProfiles         []LLMProfileManifest         `json:"llm_profiles,omitempty"`
-	return Artifact{
+	Metadata            map[string]any               `json:"metadata,omitempty"`
-		ExtractorKey:  candidate.ExtractorKey,
+	SchemaVersion       string                       `json:"schema_version,omitempty"`
-		ArtifactType:  candidate.ArtifactType,
+	ValidationStatus    string                       `json:"validation_status,omitempty"`
-		SchemaVersion: candidate.SchemaVersion,
+	StartedAt           *time.Time                   `json:"started_at,omitempty"`
-		Payload:       append(json.RawMessage(nil), candidate.Payload...),
+	CompletedAt         *time.Time                   `json:"completed_at,omitempty"`
-		SourceRefs:    append([]source.SourceRef(nil), candidate.SourceRefs...),
-		Metadata:      copyMetadata(candidate.Metadata),
-	}
-}
-
-func copyMetadata(metadata map[string]any) map[string]any {
-	if len(metadata) == 0 {
-		return nil
-	}
-
-	copied := make(map[string]any, len(metadata))
-	for key, value := range metadata {
-		copied[key] = value
-	}
-	return copied
 }

internal/core/artifacts/artifacts_test.go

@@ -2,116 +2,10 @@ package artifacts
 
 import (
 	"encoding/json"
-	"reflect"
+	"strings"
 	"testing"
-
-	"gitea.maximumdirect.net/eric/notarius/internal/core/source"
 )
 
-func TestArtifactFromCandidatePreservesCandidateFields(t *testing.T) {
-	candidate := ArtifactCandidate{
-		Index:         7,
-		ExtractorKey:  "generic-extractor",
-		ArtifactType:  "generic-artifact",
-		SchemaVersion: "v1",
-		Payload:       json.RawMessage(`{"name":"example"}`),
-		SourceRefs: []source.SourceRef{
-			{SourceID: "source-1", StartUnitID: "u1", EndUnitID: "u2"},
-		},
-		Metadata: map[string]any{
-			"confidence": 0.75,
-		},
-	}
-
-	artifact := ArtifactFromCandidate(candidate)
-
-	if artifact.ExtractorKey != candidate.ExtractorKey {
-		t.Fatalf("ExtractorKey = %q, want %q", artifact.ExtractorKey, candidate.ExtractorKey)
-	}
-	if artifact.ArtifactType != candidate.ArtifactType {
-		t.Fatalf("ArtifactType = %q, want %q", artifact.ArtifactType, candidate.ArtifactType)
-	}
-	if artifact.SchemaVersion != candidate.SchemaVersion {
-		t.Fatalf("SchemaVersion = %q, want %q", artifact.SchemaVersion, candidate.SchemaVersion)
-	}
-	if string(artifact.Payload) != string(candidate.Payload) {
-		t.Fatalf("Payload = %s, want %s", artifact.Payload, candidate.Payload)
-	}
-	if !reflect.DeepEqual(artifact.SourceRefs, candidate.SourceRefs) {
-		t.Fatalf("SourceRefs = %#v, want %#v", artifact.SourceRefs, candidate.SourceRefs)
-	}
-	if !reflect.DeepEqual(artifact.Metadata, candidate.Metadata) {
-		t.Fatalf("Metadata = %#v, want %#v", artifact.Metadata, candidate.Metadata)
-	}
-
-	candidate.Payload[0] = '['
-	candidate.SourceRefs[0].StartUnitID = "changed"
-	candidate.Metadata["confidence"] = 0.5
-
-	if string(artifact.Payload) != `{"name":"example"}` {
-		t.Fatalf("Payload changed after candidate mutation: %s", artifact.Payload)
-	}
-	if artifact.SourceRefs[0].StartUnitID != "u1" {
-		t.Fatalf("SourceRefs changed after candidate mutation: %#v", artifact.SourceRefs)
-	}
-	if artifact.Metadata["confidence"] != 0.75 {
-		t.Fatalf("Metadata changed after candidate mutation: %#v", artifact.Metadata)
-	}
-}
-
-func TestJSONMarshalUsesExpectedFieldNames(t *testing.T) {
-	candidate := ArtifactCandidate{
-		Index:         1,
-		ExtractorKey:  "generic-extractor",
-		ArtifactType:  "generic-artifact",
-		SchemaVersion: "v1",
-		Payload:       json.RawMessage(`{"value":true}`),
-		SourceRefs: []source.SourceRef{
-			{SourceID: "source-1", StartUnitID: "u1", EndUnitID: "u1"},
-		},
-		Metadata: map[string]any{
-			"reviewed": true,
-		},
-	}
-	rejected := RejectedArtifact{
-		Candidate:     candidate,
-		ValidatorName: "generic-validator",
-		ReasonCode:    "invalid",
-		Message:       "candidate was not accepted",
-	}
-
-	gotJSON, err := json.Marshal(rejected)
-	if err != nil {
-		t.Fatalf("json.Marshal() error = %v", err)
-	}
-
-	var got map[string]any
-	if err := json.Unmarshal(gotJSON, &got); err != nil {
-		t.Fatalf("json.Unmarshal() error = %v", err)
-	}
-
-	assertHasKeys(t, got, "candidate", "validator_name", "reason_code", "message")
-
-	gotCandidate, ok := got["candidate"].(map[string]any)
-	if !ok {
-		t.Fatalf("candidate = %#v, want object", got["candidate"])
-	}
-	assertHasKeys(t, gotCandidate, "index", "extractor_key", "artifact_type", "schema_version", "payload", "source_refs", "metadata")
-
-	gotRefs, ok := gotCandidate["source_refs"].([]any)
-	if !ok {
-		t.Fatalf("source_refs = %#v, want array", gotCandidate["source_refs"])
-	}
-	if len(gotRefs) != 1 {
-		t.Fatalf("len(source_refs) = %d, want 1", len(gotRefs))
-	}
-	gotRef, ok := gotRefs[0].(map[string]any)
-	if !ok {
-		t.Fatalf("source_refs[0] = %#v, want object", gotRefs[0])
-	}
-	assertHasKeys(t, gotRef, "source_id", "start_unit_id", "end_unit_id")
-}
-
 func TestRunManifestOmitsEmptyOptionalFields(t *testing.T) {
 	gotJSON, err := json.Marshal(RunManifest{})
 	if err != nil {
@@ -123,12 +17,49 @@ func TestRunManifestOmitsEmptyOptionalFields(t *testing.T) {
 	}
 }
 
+func TestRunManifestChunkPlanIsAdditiveAndOmitsPlanContent(t *testing.T) {
+	manifest := RunManifest{ChunkPlan: &ChunkPlanManifest{
+		Mode: "auto", Action: "reused", SourceDigest: "sha256:source", PlanDigest: "sha256:plan",
+		PlanSchemaVersion: "notarius.chunk-plan.v2", RequestedModule: "chunk/current",
+		ProducerInputModule: "input/original", ProducerModule: "chunk/original",
+	}}
+	encoded, err := json.Marshal(manifest)
+	if err != nil {
+		t.Fatal(err)
+	}
+	text := string(encoded)
+	for _, want := range []string{`"chunk_plan"`, `"action":"reused"`, `"requested_module":"chunk/current"`, `"producer_module":"chunk/original"`} {
+		if !strings.Contains(text, want) {
+			t.Fatalf("manifest JSON %s does not contain %s", text, want)
+		}
+	}
+	for _, forbidden := range []string{`"plan"`, `"units"`, `"annotations"`} {
+		if strings.Contains(text, forbidden) {
+			t.Fatalf("manifest JSON contains forbidden field %s: %s", forbidden, text)
+		}
+	}
+
+	var legacy RunManifest
+	if err := json.Unmarshal([]byte(`{"pipeline_id":"legacy"}`), &legacy); err != nil {
+		t.Fatal(err)
+	}
+	if legacy.PipelineID != "legacy" || legacy.ChunkPlan != nil {
+		t.Fatalf("legacy manifest = %#v", legacy)
+	}
+}
+
 func TestRunManifestIncludesPipelineAndArtifactLaneFields(t *testing.T) {
 	manifest := RunManifest{
 		PipelineID:     "pipeline-1",
 		PipelineDigest: "sha256:abc123",
 		LLMProfiles: []LLMProfileManifest{
-			{ID: "default", Provider: "openai-compatible", Model: "model-a"},
+			{
+				ID:              "default",
+				Provider:        "promptkit",
+				Model:           "model-a",
+				BackendID:       "openrouter",
+				ReasoningEffort: "high",
+			},
 		},
 		ArtifactLanes: []ArtifactLaneManifest{
 			{
@@ -136,12 +67,21 @@ func TestRunManifestIncludesPipelineAndArtifactLaneFields(t *testing.T) {
 				Extractor:  "event-extractor",
 				Merger:     "appendorder",
 				Normalizer: "noop",
-				Validators: []string{"grounded"},
 				Metadata: map[string]any{
 					"extractor": map[string]any{"prompt_id": "test.prompt"},
 				},
 			},
 		},
+		ValidatorChains: []ValidatorChainManifest{
+			{
+				Stage:     "extract",
+				LaneID:    "events",
+				ModuleKey: "event-extractor",
+				Validators: []ValidatorManifest{
+					{Key: "grounded", ExecutionClass: "deterministic"},
+				},
+			},
+		},
 	}
 
 	gotJSON, err := json.Marshal(manifest)
@@ -154,7 +94,7 @@ func TestRunManifestIncludesPipelineAndArtifactLaneFields(t *testing.T) {
 		t.Fatalf("json.Unmarshal() error = %v", err)
 	}
 
-	assertHasKeys(t, got, "pipeline_id", "pipeline_digest", "artifact_lanes", "llm_profiles")
+	assertHasKeys(t, got, "pipeline_id", "pipeline_digest", "artifact_lanes", "validator_chains", "llm_profiles")
 
 	profiles, ok := got["llm_profiles"].([]any)
 	if !ok {
@@ -167,7 +107,13 @@ func TestRunManifestIncludesPipelineAndArtifactLaneFields(t *testing.T) {
 	if !ok {
 		t.Fatalf("llm_profiles[0] = %#v, want object", profiles[0])
 	}
-	assertHasKeys(t, profile, "id", "provider", "model")
+	assertHasKeys(t, profile, "id", "provider", "model", "backend_id", "reasoning_effort")
+	if profile["provider"] != "promptkit" {
+		t.Fatalf("llm_profiles[0].provider = %#v, want promptkit", profile["provider"])
+	}
+	if profile["backend_id"] != "openrouter" || profile["reasoning_effort"] != "high" {
+		t.Fatalf("llm_profiles[0] = %#v, want backend and reasoning provenance", profile)
+	}
 
 	lanes, ok := got["artifact_lanes"].([]any)
 	if !ok {
@@ -180,7 +126,94 @@ func TestRunManifestIncludesPipelineAndArtifactLaneFields(t *testing.T) {
 	if !ok {
 		t.Fatalf("artifact_lanes[0] = %#v, want object", lanes[0])
 	}
-	assertHasKeys(t, lane, "id", "extractor", "merger", "normalizer", "validators", "metadata")
+	assertHasKeys(t, lane, "id", "extractor", "merger", "normalizer", "metadata")
+
+	chains, ok := got["validator_chains"].([]any)
+	if !ok {
+		t.Fatalf("validator_chains = %#v, want array", got["validator_chains"])
+	}
+	if len(chains) != 1 {
+		t.Fatalf("len(validator_chains) = %d, want 1", len(chains))
+	}
+	chain, ok := chains[0].(map[string]any)
+	if !ok {
+		t.Fatalf("validator_chains[0] = %#v, want object", chains[0])
+	}
+	assertHasKeys(t, chain, "stage", "lane_id", "module_key", "validators")
+}
+
+func TestRunManifestIncludesReferenceProvenance(t *testing.T) {
+	manifest := RunManifest{
+		References: []ReferenceProvenance{
+			{
+				Stage:         "extract",
+				LaneID:        "events",
+				SlotName:      "roster",
+				OriginType:    "file",
+				OriginURI:     "file:///tmp/roster.txt",
+				Digest:        "sha256:reference",
+				MediaType:     "text/plain; charset=utf-8",
+				SizeBytes:     12,
+				BindingSource: "config",
+			},
+		},
+	}
+
+	gotJSON, err := json.Marshal(manifest)
+	if err != nil {
+		t.Fatalf("json.Marshal() error = %v", err)
+	}
+
+	var got RunManifest
+	if err := json.Unmarshal(gotJSON, &got); err != nil {
+		t.Fatalf("json.Unmarshal() error = %v", err)
+	}
+	if len(got.References) != 1 {
+		t.Fatalf("len(References) = %d, want 1", len(got.References))
+	}
+	reference := got.References[0]
+	if reference.Stage != "extract" || reference.LaneID != "events" || reference.SlotName != "roster" || reference.OriginType != "file" || reference.OriginURI != "file:///tmp/roster.txt" {
+		t.Fatalf("reference provenance = %#v, want lane-scoped origin details", reference)
+	}
+	if reference.Digest != "sha256:reference" || reference.MediaType != "text/plain; charset=utf-8" || reference.SizeBytes != 12 || reference.BindingSource != "config" {
+		t.Fatalf("reference provenance = %#v, want digest/media/size/source details", reference)
+	}
+}
+
+func TestRunManifestIncludesTopLevelModuleMetadata(t *testing.T) {
+	manifest := RunManifest{
+		ModuleMetadata: map[string]map[string]any{
+			"chunker": {
+				"prompt_id":            "dnd.scenes",
+				"prompt_version":       "v1",
+				"prompt_sha256":        "sha256:abc123",
+				"response_schema_key":  "dnd_scenes",
+				"response_schema_name": "dnd_scenes",
+			},
+		},
+	}
+
+	gotJSON, err := json.Marshal(manifest)
+	if err != nil {
+		t.Fatalf("json.Marshal() error = %v", err)
+	}
+
+	var got map[string]any
+	if err := json.Unmarshal(gotJSON, &got); err != nil {
+		t.Fatalf("json.Unmarshal() error = %v", err)
+	}
+
+	moduleMetadata, ok := got["module_metadata"].(map[string]any)
+	if !ok {
+		t.Fatalf("module_metadata = %#v, want object", got["module_metadata"])
+	}
+	assertHasKeys(t, moduleMetadata, "chunker")
+
+	chunkerMetadata, ok := moduleMetadata["chunker"].(map[string]any)
+	if !ok {
+		t.Fatalf("module_metadata.chunker = %#v, want object", moduleMetadata["chunker"])
+	}
+	assertHasKeys(t, chunkerMetadata, "prompt_id", "prompt_version", "prompt_sha256", "response_schema_key", "response_schema_name")
 }
 
 func assertHasKeys(t *testing.T, values map[string]any, keys ...string) {

internal/core/config/cache.go

@@ -0,0 +1,31 @@
+package config
+
+import (
+	"fmt"
+	"path/filepath"
+	"strings"
+)
+
+// DefaultChunkPlanRoot resolves the existing per-user chunk-plan cache root.
+func DefaultChunkPlanRoot(userCacheDir func() (string, error)) (string, error) {
+	return defaultCacheFamilyRoot(userCacheDir, "chunk-plans")
+}
+
+func DefaultCheckpointRoot(userCacheDir func() (string, error)) (string, error) {
+	return defaultCacheFamilyRoot(userCacheDir, "checkpoints")
+}
+
+func defaultCacheFamilyRoot(userCacheDir func() (string, error), family string) (string, error) {
+	if userCacheDir == nil {
+		return "", fmt.Errorf("user cache directory resolver must not be nil")
+	}
+	root, err := userCacheDir()
+	if err != nil {
+		return "", fmt.Errorf("resolve user cache directory: %w", err)
+	}
+	root = strings.TrimSpace(root)
+	if root == "" {
+		return "", fmt.Errorf("user cache directory must not be empty")
+	}
+	return filepath.Join(filepath.Clean(root), "notarius", family), nil
+}

internal/core/config/config.go

@@ -1,66 +1,82 @@
 package config
 
 import (
-	"gitea.maximumdirect.net/eric/notarius/internal/core/diagnostics"
 	"gitea.maximumdirect.net/eric/notarius/internal/framework/pipeline"
 )
 
-const SupportedFileConfigVersion = 1
+const SupportedFileConfigVersion = 4
 
 type Config struct {
-	LLMProfiles map[string]LLMProfile               `json:"llm_profiles"`
+	PromptKit   PromptKitConfig                     `json:"promptkit,omitempty"`
 	Pipelines   map[string]pipeline.PipelineProfile `json:"pipelines"`
 	Concurrency ConcurrencyConfig                   `json:"concurrency"`
-	Diagnostics DiagnosticsConfig                   `json:"diagnostics"`
+	Output      OutputConfig                        `json:"output"`
+	Cache       CacheConfig                         `json:"cache"`
+	Debug       DebugConfig                         `json:"debug"`
 }
 
-type LLMProfile struct {
+type PromptKitConfig struct {
-	Provider       string `json:"provider,omitempty"`
+	ProfileDir   string                       `json:"profile_dir,omitempty"`
-	BaseURL        string `json:"base_url,omitempty"`
+	ProfileFile  string                       `json:"profile_file,omitempty"`
-	Model          string `json:"model,omitempty"`
+	LocalBackend *PromptKitLocalBackendConfig `json:"local_backend,omitempty"`
-	APIKey         string `json:"api_key,omitempty"`
+}
-	APIKeyEnv      string `json:"api_key_env,omitempty"`
+
-	TimeoutSeconds int    `json:"timeout_seconds,omitempty"`
+type PromptKitLocalBackendConfig struct {
-	MaxRetries     int    `json:"max_retries,omitempty"`
+	Endpoint         string `json:"endpoint"`
-	MaxConcurrency int    `json:"max_concurrency,omitempty"`
+	ConcurrencyLimit int    `json:"concurrency_limit"`
 }
 
 type ConcurrencyConfig struct {
-	TotalLLM int `json:"total_llm"`
+	TotalLLM     int            `json:"total_llm"`
+	StageWorkers map[string]int `json:"stage_workers"`
+
+	extractWorkersConfigured bool
+	defaultedExtractWorkers  int
 }
 
-type DiagnosticsConfig struct {
+type OutputConfig struct {
-	WorkDir   string                    `json:"work_dir"`
+	Directory string `json:"directory"`
-	Retention diagnostics.RetentionMode `json:"retention"`
+}
+
+type CacheConfig struct {
+	ChunkPlans  ChunkPlanCacheConfig  `json:"chunk_plans"`
+	Checkpoints CheckpointCacheConfig `json:"checkpoints"`
+}
+
+type ChunkPlanCacheConfig struct {
+	Directory string                  `json:"directory,omitempty"`
+	Mode      pipeline.ChunkCacheMode `json:"mode"`
+}
+
+type CheckpointCacheConfig struct {
+	Enabled   bool   `json:"enabled"`
+	Directory string `json:"directory,omitempty"`
+}
+type DebugConfig struct {
+	Directory string `json:"directory"`
 }
 
 func Default() Config {
 	return Config{
-		LLMProfiles: map[string]LLMProfile{
-			pipeline.DefaultLLMProfile: {
-				Provider:       "openai-compatible",
-				TimeoutSeconds: 600,
-				MaxRetries:     3,
-				MaxConcurrency: 1,
-			},
-		},
 		Pipelines: map[string]pipeline.PipelineProfile{},
 		Concurrency: ConcurrencyConfig{
-			TotalLLM: 1,
+			TotalLLM:                1,
-		},
+			StageWorkers:            map[string]int{"extract": 1},
-		Diagnostics: DiagnosticsConfig{
+			defaultedExtractWorkers: 1,
-			WorkDir:   "/tmp/notarius",
-			Retention: diagnostics.RetentionAuto,
 		},
+		Output: OutputConfig{Directory: "./notarius-output"},
+		Cache:  CacheConfig{ChunkPlans: ChunkPlanCacheConfig{Mode: pipeline.ChunkCacheAuto}},
+		Debug:  DebugConfig{Directory: "./notarius-debug"},
 	}
 }
 
 func cloneConfig(in Config) Config {
 	out := in
-	out.LLMProfiles = make(map[string]LLMProfile, len(in.LLMProfiles))
+	if in.PromptKit.LocalBackend != nil {
-	for key, profile := range in.LLMProfiles {
+		localBackend := *in.PromptKit.LocalBackend
-		out.LLMProfiles[key] = profile
+		out.PromptKit.LocalBackend = &localBackend
 	}
+	out.Concurrency.StageWorkers = cloneIntMap(in.Concurrency.StageWorkers)
 	out.Pipelines = make(map[string]pipeline.PipelineProfile, len(in.Pipelines))
 	for key, profile := range in.Pipelines {
 		out.Pipelines[key] = clonePipelineProfile(profile)
@@ -68,11 +84,59 @@ func cloneConfig(in Config) Config {
 	return out
 }
 
+func cloneIntMap(in map[string]int) map[string]int {
+	if len(in) == 0 {
+		return nil
+	}
+	out := make(map[string]int, len(in))
+	for key, value := range in {
+		out[key] = value
+	}
+	return out
+}
+
+func (c *ConcurrencyConfig) recomputeStageWorkerDefaults() {
+	if c == nil {
+		return
+	}
+	if c.StageWorkers == nil {
+		c.StageWorkers = make(map[string]int)
+	}
+	if !c.extractWorkersConfigured {
+		if value, ok := c.StageWorkers["extract"]; ok && (c.defaultedExtractWorkers == 0 || value != c.defaultedExtractWorkers) {
+			c.extractWorkersConfigured = true
+			return
+		}
+		c.StageWorkers["extract"] = c.TotalLLM
+		c.defaultedExtractWorkers = c.TotalLLM
+	}
+}
+
 func clonePipelineProfile(in pipeline.PipelineProfile) pipeline.PipelineProfile {
 	out := in
 	out.Input = cloneModuleBinding(in.Input)
 	out.Chunk = cloneModuleBinding(in.Chunk)
 	out.Output = cloneModuleBinding(in.Output)
+	out.References = cloneReferenceSourceMap(in.References)
+	if len(in.Artifacts) > 0 {
+		out.Artifacts = make(map[string]pipeline.ArtifactLaneProfile, len(in.Artifacts))
+		for key, lane := range in.Artifacts {
+			out.Artifacts[key] = cloneArtifactLaneProfile(lane)
+		}
+	}
+	if in.Steps != nil {
+		out.Steps = make([]pipeline.PipelineStepProfile, len(in.Steps))
+		for i, step := range in.Steps {
+			out.Steps[i] = clonePipelineStepProfile(step)
+		}
+	}
+	return out
+}
+
+func clonePipelineStepProfile(in pipeline.PipelineStepProfile) pipeline.PipelineStepProfile {
+	out := in
+	out.ID = in.ID
+	out.References = cloneReferenceSourceMap(in.References)
 	if len(in.Artifacts) > 0 {
 		out.Artifacts = make(map[string]pipeline.ArtifactLaneProfile, len(in.Artifacts))
 		for key, lane := range in.Artifacts {
@@ -87,6 +151,7 @@ func cloneArtifactLaneProfile(in pipeline.ArtifactLaneProfile) pipeline.Artifact
 	out.Extract = cloneModuleBinding(in.Extract)
 	out.Merge = cloneModuleBinding(in.Merge)
 	out.Normalize = cloneModuleBinding(in.Normalize)
+	out.References = cloneReferenceSourceMap(in.References)
 	if len(in.Validators) > 0 {
 		out.Validators = make([]pipeline.ModuleBinding, len(in.Validators))
 		for i, binding := range in.Validators {
@@ -96,11 +161,55 @@ func cloneArtifactLaneProfile(in pipeline.ArtifactLaneProfile) pipeline.Artifact
 	return out
 }
 
+func cloneStringMap(in map[string]string) map[string]string {
+	if len(in) == 0 {
+		return nil
+	}
+	out := make(map[string]string, len(in))
+	for key, value := range in {
+		out[key] = value
+	}
+	return out
+}
+
+func cloneReferenceSourceMap(in map[string]pipeline.ReferenceSource) map[string]pipeline.ReferenceSource {
+	if len(in) == 0 {
+		return nil
+	}
+	out := make(map[string]pipeline.ReferenceSource, len(in))
+	for key, source := range in {
+		out[key] = cloneReferenceSource(source)
+	}
+	return out
+}
+
+func cloneReferenceSource(in pipeline.ReferenceSource) pipeline.ReferenceSource {
+	out := in
+	if in.Artifact != nil {
+		artifact := *in.Artifact
+		out.Artifact = &artifact
+	}
+	return out
+}
+
 func cloneModuleBinding(in pipeline.ModuleBinding) pipeline.ModuleBinding {
 	out := in
 	if len(in.Options) > 0 {
 		out.Options = cloneOptions(in.Options)
 	}
+	out.References = cloneReferenceSourceMap(in.References)
+	out.Validators = cloneValidatorOverride(in.Validators)
+	return out
+}
+
+func cloneValidatorOverride(in pipeline.ValidatorOverride) pipeline.ValidatorOverride {
+	out := pipeline.ValidatorOverride{Set: in.Set}
+	if len(in.Validators) > 0 {
+		out.Validators = make([]pipeline.ModuleBinding, len(in.Validators))
+		for i, binding := range in.Validators {
+			out.Validators[i] = cloneModuleBinding(binding)
+		}
+	}
 	return out
 }
 

internal/core/config/config_test.go

@@ -1,78 +0,0 @@
-package config
-
-import (
-	"testing"
-
-	"gitea.maximumdirect.net/eric/notarius/internal/core/diagnostics"
-	"gitea.maximumdirect.net/eric/notarius/internal/framework/pipeline"
-)
-
-func TestDefaultValues(t *testing.T) {
-	cfg := Default()
-
-	defaultProfile, ok := cfg.LLMProfiles[pipeline.DefaultLLMProfile]
-	if !ok {
-		t.Fatalf("expected default LLM profile")
-	}
-	if defaultProfile.Provider != "openai-compatible" {
-		t.Fatalf("unexpected provider: %q", defaultProfile.Provider)
-	}
-	if defaultProfile.BaseURL != "" || defaultProfile.Model != "" {
-		t.Fatalf("default profile should not require base URL/model yet: %+v", defaultProfile)
-	}
-	if defaultProfile.TimeoutSeconds != 600 || defaultProfile.MaxRetries != 3 || defaultProfile.MaxConcurrency != 1 {
-		t.Fatalf("unexpected default LLM operational values: %+v", defaultProfile)
-	}
-	if len(cfg.Pipelines) != 0 {
-		t.Fatalf("expected no built-in pipeline profiles, got %v", cfg.Pipelines)
-	}
-	if cfg.Concurrency.TotalLLM != 1 {
-		t.Fatalf("unexpected total LLM concurrency: %d", cfg.Concurrency.TotalLLM)
-	}
-	if cfg.Diagnostics.WorkDir != "/tmp/notarius" {
-		t.Fatalf("unexpected diagnostics work dir: %q", cfg.Diagnostics.WorkDir)
-	}
-	if cfg.Diagnostics.Retention != diagnostics.RetentionAuto {
-		t.Fatalf("unexpected diagnostics retention: %q", cfg.Diagnostics.Retention)
-	}
-}
-
-func TestApplyFileConfigMergesWithDefaults(t *testing.T) {
-	fileCfg, err := ParseFileConfigYAML([]byte(`
-version: 1
-llm_profiles:
-  default:
-    model: test-model
-pipelines:
-  example:
-    input: fake/input
-    artifacts:
-      events:
-        extract: fake/extract
-`))
-	if err != nil {
-		t.Fatalf("ParseFileConfigYAML: %v", err)
-	}
-
-	cfg := Default()
-	if err := cfg.applyFileConfigWithLookup(fileCfg, emptyLookup); err != nil {
-		t.Fatalf("ApplyFileConfig: %v", err)
-	}
-
-	profile := cfg.LLMProfiles[pipeline.DefaultLLMProfile]
-	if profile.Model != "test-model" {
-		t.Fatalf("expected file model, got %+v", profile)
-	}
-	if profile.Provider != "openai-compatible" || profile.TimeoutSeconds != 600 || profile.MaxRetries != 3 {
-		t.Fatalf("expected default LLM fields to be preserved, got %+v", profile)
-	}
-	if cfg.Concurrency.TotalLLM != 1 {
-		t.Fatalf("expected default concurrency preserved, got %d", cfg.Concurrency.TotalLLM)
-	}
-	if cfg.Diagnostics.Retention != diagnostics.RetentionAuto {
-		t.Fatalf("expected default diagnostics retention preserved, got %q", cfg.Diagnostics.Retention)
-	}
-	if _, ok := cfg.Pipelines["example"]; !ok {
-		t.Fatalf("expected file pipeline to be applied")
-	}
-}

internal/core/config/effective_config.go

@@ -3,9 +3,7 @@ package config
 import (
 	"fmt"
 	"strings"
-	"time"
 
-	"gitea.maximumdirect.net/eric/notarius/internal/framework/llm"
 	"gitea.maximumdirect.net/eric/notarius/internal/framework/pipeline"
 )
 
@@ -14,16 +12,21 @@ type ResolveInput struct {
 	Only               []string
 	Catalog            pipeline.ModuleCatalog
 	LLMProfileOverride string
+	ReferenceOverrides []pipeline.ReferenceBinding
+	ReferenceUnbinds   []pipeline.ReferenceUnbind
 }
 
 type EffectiveConfig struct {
-	Config           Config
+	Config             Config
-	PipelineID       string
+	PipelineID         string
-	Only             []string
+	Only               []string
-	ResolvedPipeline pipeline.ResolvedPipeline
+	ReferenceOverrides []pipeline.ReferenceBinding
+	ReferenceUnbinds   []pipeline.ReferenceUnbind
+	ResolvedPipeline   pipeline.ResolvedPipeline
 }
 
 func (c Config) Resolve(input ResolveInput) (EffectiveConfig, error) {
+	c.Concurrency.recomputeStageWorkerDefaults()
 	if err := c.Validate(); err != nil {
 		return EffectiveConfig{}, err
 	}
@@ -39,41 +42,27 @@ func (c Config) Resolve(input ResolveInput) (EffectiveConfig, error) {
 	}
 	profile = clonePipelineProfile(profile)
 	profile.ID = pipelineID
-	if override := strings.TrimSpace(input.LLMProfileOverride); override != "" {
-		if !hasLLMProfile(c.LLMProfiles, override) {
-			return EffectiveConfig{}, fmt.Errorf("LLM profile override %q is not configured", override)
-		}
-		applyLLMProfileOverride(&profile, override)
-	}
 
-	resolved, err := pipeline.ResolvePipeline(profile, pipeline.ResolveOptions{Only: input.Only}, input.Catalog)
+	resolved, err := pipeline.ResolvePipeline(profile, pipeline.ResolveOptions{
+		Only:               input.Only,
+		LLMProfileOverride: input.LLMProfileOverride,
+		ReferenceOverrides: append([]pipeline.ReferenceBinding(nil), input.ReferenceOverrides...),
+		ReferenceUnbinds:   append([]pipeline.ReferenceUnbind(nil), input.ReferenceUnbinds...),
+	}, input.Catalog)
 	if err != nil {
 		return EffectiveConfig{}, fmt.Errorf("resolve pipeline %q: %w", pipelineID, err)
 	}
 
 	return EffectiveConfig{
-		Config:           cloneConfig(c),
+		Config:             cloneConfig(c),
-		PipelineID:       pipelineID,
+		PipelineID:         pipelineID,
-		Only:             append([]string(nil), input.Only...),
+		Only:               append([]string(nil), input.Only...),
-		ResolvedPipeline: resolved,
+		ReferenceOverrides: append([]pipeline.ReferenceBinding(nil), input.ReferenceOverrides...),
+		ReferenceUnbinds:   append([]pipeline.ReferenceUnbind(nil), input.ReferenceUnbinds...),
+		ResolvedPipeline:   resolved,
 	}, nil
 }
 
-func applyLLMProfileOverride(profile *pipeline.PipelineProfile, profileID string) {
-	profile.Input.LLMProfile = profileID
-	profile.Chunk.LLMProfile = profileID
-	profile.Output.LLMProfile = profileID
-	for laneID, lane := range profile.Artifacts {
-		lane.Extract.LLMProfile = profileID
-		lane.Merge.LLMProfile = profileID
-		lane.Normalize.LLMProfile = profileID
-		for i := range lane.Validators {
-			lane.Validators[i].LLMProfile = profileID
-		}
-		profile.Artifacts[laneID] = lane
-	}
-}
-
 func lookupPipelineProfile(profiles map[string]pipeline.PipelineProfile, pipelineID string) (pipeline.PipelineProfile, bool) {
 	pipelineID = strings.TrimSpace(pipelineID)
 	for rawID, profile := range profiles {
@@ -83,36 +72,3 @@ func lookupPipelineProfile(profiles map[string]pipeline.PipelineProfile, pipelin
 	}
 	return pipeline.PipelineProfile{}, false
 }
-
-func (c Config) OpenAICompatibleClientConfig(profileID string) (llm.OpenAICompatibleClientConfig, error) {
-	trimmedID := strings.TrimSpace(profileID)
-	profile, ok := c.LLMProfile(trimmedID)
-	if !ok {
-		return llm.OpenAICompatibleClientConfig{}, fmt.Errorf("LLM profile %q is not configured", trimmedID)
-	}
-
-	provider := strings.TrimSpace(profile.Provider)
-	if provider == "" {
-		provider = providerOpenAICompatible
-	}
-	if provider != providerOpenAICompatible {
-		return llm.OpenAICompatibleClientConfig{}, fmt.Errorf("LLM profile %q provider %q is not supported", trimmedID, provider)
-	}
-
-	baseURL := strings.TrimSpace(profile.BaseURL)
-	if baseURL == "" {
-		return llm.OpenAICompatibleClientConfig{}, fmt.Errorf("LLM profile %q base URL must not be empty", trimmedID)
-	}
-	model := strings.TrimSpace(profile.Model)
-	if model == "" {
-		return llm.OpenAICompatibleClientConfig{}, fmt.Errorf("LLM profile %q model must not be empty", trimmedID)
-	}
-
-	return llm.OpenAICompatibleClientConfig{
-		BaseURL:        baseURL,
-		Model:          model,
-		APIKey:         profile.APIKey,
-		MaxRetries:     profile.MaxRetries,
-		RequestTimeout: time.Duration(profile.TimeoutSeconds) * time.Second,
-	}, nil
-}

internal/core/config/effective_config_contract_test.go

@@ -0,0 +1,584 @@
+package config
+
+import (
+	"context"
+	"encoding/json"
+	"strings"
+	"testing"
+
+	"gitea.maximumdirect.net/eric/notarius/internal/core/source"
+	"gitea.maximumdirect.net/eric/notarius/internal/framework/contracts"
+	"gitea.maximumdirect.net/eric/notarius/internal/framework/pipeline"
+)
+
+func TestEffectiveConfigRejectsEmptyAndUnknownPipelineIDs(t *testing.T) {
+	cfg := configForEffectiveTests(t, effectiveProfile())
+	for _, pipelineID := range []string{"", "missing"} {
+		name := pipelineID
+		if name == "" {
+			name = "empty"
+		}
+		t.Run(name, func(t *testing.T) {
+			_, err := cfg.Resolve(ResolveInput{PipelineID: pipelineID, Catalog: effectiveCatalog(t)})
+			if err == nil || !strings.Contains(err.Error(), "pipeline") {
+				t.Fatalf("Resolve(%q) error = %v, want pipeline context", pipelineID, err)
+			}
+		})
+	}
+}
+
+func TestEffectiveConfigResolvesTrimmedPipelineMapKeys(t *testing.T) {
+	profile := effectiveProfile()
+	profile.ID = " main "
+	cfg := Default()
+	cfg.Pipelines = map[string]pipeline.PipelineProfile{" main ": profile}
+	effective, err := cfg.Resolve(ResolveInput{PipelineID: "main", Catalog: effectiveCatalog(t)})
+	if err != nil {
+		t.Fatalf("Resolve() error = %v", err)
+	}
+	if effective.PipelineID != "main" || effective.ResolvedPipeline.ID != "main" {
+		t.Fatalf("resolved IDs = %q, %q", effective.PipelineID, effective.ResolvedPipeline.ID)
+	}
+}
+
+func TestEffectiveConfigOnlySelectsRequestedLanesWithoutMutatingSource(t *testing.T) {
+	profile := effectiveProfile()
+	profile.Artifacts["other"] = pipeline.ArtifactLaneProfile{Extract: pipeline.Binding("extract")}
+	cfg := configForEffectiveTests(t, profile)
+	effective, err := cfg.Resolve(ResolveInput{
+		PipelineID: "main",
+		Only:       []string{"other"},
+		Catalog:    effectiveCatalog(t),
+	})
+	if err != nil {
+		t.Fatalf("Resolve() error = %v", err)
+	}
+	if len(effective.ResolvedPipeline.Steps[0].ArtifactLanes) != 1 || effective.ResolvedPipeline.Steps[0].ArtifactLanes[0].ID != "other" {
+		t.Fatalf("resolved lanes = %#v", effective.ResolvedPipeline.Steps[0].ArtifactLanes)
+	}
+	if len(cfg.Pipelines["main"].Artifacts) != 2 {
+		t.Fatalf("source lanes were mutated: %#v", cfg.Pipelines["main"].Artifacts)
+	}
+
+	_, err = cfg.Resolve(ResolveInput{
+		PipelineID: "main",
+		Only:       []string{"missing"},
+		Catalog:    effectiveCatalog(t),
+	})
+	if err == nil || !strings.Contains(err.Error(), "lane \"missing\"") {
+		t.Fatalf("unknown lane error = %v", err)
+	}
+}
+
+func TestEffectiveConfigMaterializesDefaultBindingsThroughCatalog(t *testing.T) {
+	effective, err := resolveEffectiveProfile(t, effectiveProfile(), ResolveInput{})
+	if err != nil {
+		t.Fatalf("Resolve() error = %v", err)
+	}
+	resolved := effective.ResolvedPipeline
+	if resolved.Chunk.Module != pipeline.DefaultChunkModule || resolved.Output.Module != pipeline.DefaultOutputModule {
+		t.Fatalf("default pipeline bindings = %#v, %#v", resolved.Chunk, resolved.Output)
+	}
+	if len(resolved.Steps[0].ArtifactLanes) != 1 || resolved.Steps[0].ArtifactLanes[0].Merge.Module != pipeline.DefaultMergeModule || resolved.Steps[0].ArtifactLanes[0].Normalize.Module != pipeline.DefaultNormalizeModule {
+		t.Fatalf("default lane bindings = %#v", resolved.Steps[0].ArtifactLanes)
+	}
+}
+
+func TestEffectiveConfigPreservesPromptKitProfileSource(t *testing.T) {
+	tests := []struct {
+		name          string
+		profileSource PromptKitConfig
+	}{
+		{name: "profile directory", profileSource: PromptKitConfig{ProfileDir: "./profiles"}},
+		{name: "profile file", profileSource: PromptKitConfig{ProfileFile: "./profiles.yml"}},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			cfg := configForEffectiveTests(t, effectiveProfile())
+			cfg.PromptKit = tt.profileSource
+			effective, err := cfg.Resolve(ResolveInput{PipelineID: "main", Catalog: effectiveCatalog(t)})
+			if err != nil {
+				t.Fatalf("Resolve() error = %v", err)
+			}
+			if effective.Config.PromptKit != cfg.PromptKit {
+				t.Fatalf("effective PromptKit config = %#v, want %#v", effective.Config.PromptKit, cfg.PromptKit)
+			}
+		})
+	}
+}
+
+func TestEffectiveConfigOwnsPromptKitLocalBackend(t *testing.T) {
+	cfg := configForEffectiveTests(t, effectiveProfile())
+	cfg.PromptKit.LocalBackend = &PromptKitLocalBackendConfig{
+		Endpoint:         "http://localhost:8000/v1",
+		ConcurrencyLimit: 2,
+	}
+	effective, err := cfg.Resolve(ResolveInput{PipelineID: "main", Catalog: effectiveCatalog(t)})
+	if err != nil {
+		t.Fatalf("Resolve() error = %v", err)
+	}
+	if effective.Config.PromptKit.LocalBackend == nil {
+		t.Fatal("effective local backend = nil")
+	}
+	if effective.Config.PromptKit.LocalBackend == cfg.PromptKit.LocalBackend {
+		t.Fatal("effective local backend aliases input config")
+	}
+
+	cfg.PromptKit.LocalBackend.Endpoint = "http://changed-input.example/v1"
+	if effective.Config.PromptKit.LocalBackend.Endpoint != "http://localhost:8000/v1" {
+		t.Fatalf("input mutation changed effective config: %#v", effective.Config.PromptKit.LocalBackend)
+	}
+	effective.Config.PromptKit.LocalBackend.ConcurrencyLimit = 9
+	if cfg.PromptKit.LocalBackend.ConcurrencyLimit != 2 {
+		t.Fatalf("effective mutation changed input config: %#v", cfg.PromptKit.LocalBackend)
+	}
+}
+
+func TestEffectiveConfigResolutionFailuresRetainContext(t *testing.T) {
+	tests := []struct {
+		name   string
+		mutate func(*pipeline.PipelineProfile)
+		want   []string
+	}{
+		{
+			name: "unknown module",
+			mutate: func(profile *pipeline.PipelineProfile) {
+				profile.Input.Module = "missing-input"
+			},
+			want: []string{"pipeline \"main\"", "input"},
+		},
+		{
+			name: "missing capability",
+			mutate: func(profile *pipeline.PipelineProfile) {
+				profile.Chunk.Module = "needs-capability"
+			},
+			want: []string{"pipeline \"main\"", "chunk"},
+		},
+		{
+			name: "missing artifact variant",
+			mutate: func(profile *pipeline.PipelineProfile) {
+				profile.Artifacts["lane"] = pipeline.ArtifactLaneProfile{
+					Extract: pipeline.Binding("extract"),
+					Merge:   pipeline.Binding("other-merge"),
+				}
+			},
+			want: []string{"pipeline \"main\"", "lane \"lane\"", "merge"},
+		},
+		{
+			name: "invalid module options",
+			mutate: func(profile *pipeline.PipelineProfile) {
+				profile.Chunk = pipeline.ModuleBinding{Module: "generic", Options: map[string]any{"unknown": true}}
+			},
+			want: []string{"pipeline \"main\"", "chunk", "generic", "options"},
+		},
+		{
+			name: "invalid validator options",
+			mutate: func(profile *pipeline.PipelineProfile) {
+				lane := profile.Artifacts["lane"]
+				lane.Extract.Validators = pipeline.ValidatorOverride{
+					Set: true,
+					Validators: []pipeline.ModuleBinding{{
+						Module:  "option-validator",
+						Options: map[string]any{"invalid": true},
+					}},
+				}
+				profile.Artifacts["lane"] = lane
+			},
+			want: []string{"pipeline \"main\"", "lane \"lane\"", "extract", "option-validator", "options"},
+		},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			profile := effectiveProfile()
+			tt.mutate(&profile)
+			_, err := resolveEffectiveProfile(t, profile, ResolveInput{})
+			if err == nil {
+				t.Fatal("Resolve() error = nil, want failure")
+			}
+			for _, fragment := range tt.want {
+				if !strings.Contains(err.Error(), fragment) {
+					t.Fatalf("Resolve() error = %v, want context %q", err, fragment)
+				}
+			}
+		})
+	}
+}
+
+func TestEffectiveConfigLLMProfileOverrideChangesDigestAndOverridesValidators(t *testing.T) {
+	profile := effectiveProfile()
+	profile.Chunk.LLMProfile = "chunk-profile"
+	lane := profile.Artifacts["lane"]
+	lane.Extract.LLMProfile = "extract-profile"
+	lane.Merge.LLMProfile = "merge-profile"
+	lane.Normalize.LLMProfile = "normalize-profile"
+	lane.Extract.Validators = pipeline.ValidatorOverride{
+		Set: true,
+		Validators: []pipeline.ModuleBinding{{
+			Module:     "llm-validator",
+			LLMProfile: "validator-profile",
+		}},
+	}
+	profile.Artifacts["lane"] = lane
+
+	base, err := resolveEffectiveProfile(t, profile, ResolveInput{})
+	if err != nil {
+		t.Fatalf("base Resolve() error = %v", err)
+	}
+	overridden, err := resolveEffectiveProfile(t, profile, ResolveInput{LLMProfileOverride: "override-profile"})
+	if err != nil {
+		t.Fatalf("overridden Resolve() error = %v", err)
+	}
+	if base.ResolvedPipeline.Digest == overridden.ResolvedPipeline.Digest {
+		t.Fatal("LLM profile override did not change the pipeline digest")
+	}
+	resolved := overridden.ResolvedPipeline
+	if resolved.Chunk.LLMProfile != "override-profile" || resolved.Steps[0].ArtifactLanes[0].Extract.LLMProfile != "override-profile" ||
+		resolved.Steps[0].ArtifactLanes[0].Merge.LLMProfile != "override-profile" || resolved.Steps[0].ArtifactLanes[0].Normalize.LLMProfile != "override-profile" {
+		t.Fatalf("pipeline profile override was not applied: %#v", resolved)
+	}
+	validators := findEffectiveValidatorChain(resolved, pipeline.StageExtract, "lane")
+	if len(validators.Validators) != 1 || validators.Validators[0].Binding.LLMProfile != "override-profile" {
+		t.Fatalf("validator profile = %#v, want runtime override", validators)
+	}
+}
+
+func TestEffectiveConfigPipelineLLMProfileIsInheritedWithoutMutatingConfig(t *testing.T) {
+	profile := effectiveProfile()
+	profile.LLMProfile = " configured-profile "
+	effective, err := resolveEffectiveProfile(t, profile, ResolveInput{})
+	if err != nil {
+		t.Fatalf("Resolve() error = %v", err)
+	}
+	if got := effective.Config.Pipelines["main"].LLMProfile; got != " configured-profile " {
+		t.Fatalf("effective config pipeline llm profile = %q, want preserved programmatic value", got)
+	}
+	resolved := effective.ResolvedPipeline
+	if got := resolved.Chunk.LLMProfile; got != "configured-profile" {
+		t.Fatalf("resolved chunk profile = %q, want inherited profile", got)
+	}
+	if got := resolved.Steps[0].ArtifactLanes[0].Extract.LLMProfile; got != "configured-profile" {
+		t.Fatalf("resolved extract profile = %q, want inherited profile", got)
+	}
+}
+
+func TestEffectiveConfigValidatorOverridesRemainDistinctAndOrdered(t *testing.T) {
+	tests := []struct {
+		name  string
+		value pipeline.ValidatorOverride
+		want  []string
+	}{
+		{
+			name: "omitted uses default",
+			want: []string{"default-validator"},
+		},
+		{
+			name:  "explicit empty",
+			value: pipeline.ValidatorOverride{Set: true},
+			want:  nil,
+		},
+		{
+			name: "configured order",
+			value: pipeline.ValidatorOverride{
+				Set: true,
+				Validators: []pipeline.ModuleBinding{
+					pipeline.Binding("configured-a"),
+					pipeline.Binding("configured-b"),
+				},
+			},
+			want: []string{"configured-a", "configured-b"},
+		},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			profile := effectiveProfile()
+			lane := profile.Artifacts["lane"]
+			lane.Extract.Validators = tt.value
+			profile.Artifacts["lane"] = lane
+			effective, err := resolveEffectiveProfile(t, profile, ResolveInput{})
+			if err != nil {
+				t.Fatalf("Resolve() error = %v", err)
+			}
+			chain := findEffectiveValidatorChain(effective.ResolvedPipeline, pipeline.StageExtract, "lane")
+			got := make([]string, len(chain.Validators))
+			for i, validator := range chain.Validators {
+				got[i] = validator.Binding.Module
+			}
+			if len(got) != len(tt.want) {
+				t.Fatalf("validator chain = %#v, want %v", got, tt.want)
+			}
+			for i := range got {
+				if got[i] != tt.want[i] {
+					t.Fatalf("validator chain = %#v, want %v", got, tt.want)
+				}
+			}
+		})
+	}
+}
+
+func TestEffectiveConfigAndResolutionInputsDoNotAliasSource(t *testing.T) {
+	profile := effectiveProfile()
+	profile.Chunk.Options = map[string]any{"nested": map[string]any{"safe": "source"}}
+	profile.Chunk.References = pipeline.ExternalReferenceMap(map[string]string{"chunk-ref": "chunk.txt"})
+	lane := profile.Artifacts["lane"]
+	lane.Extract.Validators = pipeline.ValidatorOverride{
+		Set: true,
+		Validators: []pipeline.ModuleBinding{{
+			Module:  "configured-a",
+			Options: map[string]any{"nested": map[string]any{"safe": "validator-source"}},
+		}},
+	}
+	profile.Artifacts["lane"] = lane
+	cfg := configForEffectiveTests(t, profile)
+	only := []string{"lane"}
+	overrides := []pipeline.ReferenceBinding{{Stage: pipeline.StageChunk, SlotName: "chunk-ref", Source: "source.txt"}}
+	effective, err := cfg.Resolve(ResolveInput{
+		PipelineID:         "main",
+		Only:               only,
+		ReferenceOverrides: overrides,
+		Catalog:            effectiveCatalog(t),
+	})
+	if err != nil {
+		t.Fatalf("Resolve() error = %v", err)
+	}
+
+	effective.Config.Pipelines["main"].Chunk.Options["nested"].(map[string]any)["safe"] = "effective-config"
+	effective.ResolvedPipeline.Chunk.Options["nested"].(map[string]any)["safe"] = "resolved-pipeline"
+	effective.ResolvedPipeline.ChunkReferences.Bindings[0].Source = "resolved-reference"
+	effective.ResolvedPipeline.ValidatorChains[1].Validators[0].Binding.Options["nested"].(map[string]any)["safe"] = "resolved-validator"
+	effective.Only[0] = "mutated-only"
+	effective.ReferenceOverrides[0].Source = "mutated-override"
+
+	if got := cfg.Pipelines["main"].Chunk.Options["nested"].(map[string]any)["safe"]; got != "source" {
+		t.Fatalf("source config option was aliased: %v", got)
+	}
+	if got := cfg.Pipelines["main"].Chunk.References["chunk-ref"].Path; got != "chunk.txt" {
+		t.Fatalf("source config references were aliased: %v", got)
+	}
+	if only[0] != "lane" || overrides[0].Source != "source.txt" {
+		t.Fatal("resolution inputs were aliased")
+	}
+}
+
+func effectiveProfile() pipeline.PipelineProfile {
+	return pipeline.PipelineProfile{
+		ID:    "main",
+		Input: pipeline.Binding("input"),
+		Artifacts: map[string]pipeline.ArtifactLaneProfile{
+			"lane": {Extract: pipeline.Binding("extract")},
+		},
+	}
+}
+
+func configForEffectiveTests(t *testing.T, profile pipeline.PipelineProfile) Config {
+	t.Helper()
+	cfg := Default()
+	cfg.Pipelines = map[string]pipeline.PipelineProfile{"main": profile}
+	return cfg
+}
+
+func resolveEffectiveProfile(t *testing.T, profile pipeline.PipelineProfile, input ResolveInput) (EffectiveConfig, error) {
+	t.Helper()
+	cfg := configForEffectiveTests(t, profile)
+	if input.PipelineID == "" {
+		input.PipelineID = "main"
+	}
+	if input.Catalog.Inputs == nil {
+		input.Catalog = effectiveCatalog(t)
+	}
+	return cfg.Resolve(input)
+}
+
+func findEffectiveValidatorChain(resolved pipeline.ResolvedPipeline, stage pipeline.ModuleStage, laneID string) pipeline.ResolvedValidatorChain {
+	for _, chain := range resolved.ValidatorChains {
+		if chain.Stage == stage && chain.LaneID == laneID {
+			return chain
+		}
+	}
+	return pipeline.ResolvedValidatorChain{}
+}
+
+type effectiveArtifact struct {
+	Value string `json:"value"`
+}
+
+const effectiveArtifactKind contracts.ArtifactKind = "test/effective"
+
+type effectiveCodec struct{}
+
+func (effectiveCodec) Kind() contracts.ArtifactKind { return effectiveArtifactKind }
+func (effectiveCodec) Schema() contracts.ArtifactSchema {
+	return contracts.ArtifactSchema{
+		ID:         "effective-schema",
+		Name:       "Effective artifact",
+		Version:    "1",
+		JSONSchema: []byte(`{"type":"object"}`),
+	}
+}
+func (effectiveCodec) MediaType() string { return "application/json" }
+func (effectiveCodec) EncodeCandidate(value effectiveArtifact) ([]byte, error) {
+	return json.Marshal(value)
+}
+func (effectiveCodec) Encode(value effectiveArtifact) ([]byte, error) {
+	return json.Marshal(value)
+}
+func (effectiveCodec) Decode(content []byte) (effectiveArtifact, error) {
+	var value effectiveArtifact
+	err := json.Unmarshal(content, &value)
+	return value, err
+}
+
+type effectiveInput struct{ key string }
+
+func (m effectiveInput) Key() string { return m.key }
+func (m effectiveInput) Parse(context.Context, contracts.ParseRequest) (*source.SourceDocument, error) {
+	return &source.SourceDocument{}, nil
+}
+
+type effectiveChunker struct{ key string }
+
+func (m effectiveChunker) Key() string { return m.key }
+func (m effectiveChunker) ReferenceSlots() []contracts.ReferenceSlot {
+	if m.key == pipeline.DefaultChunkModule {
+		return []contracts.ReferenceSlot{{Name: "chunk-ref"}}
+	}
+	return nil
+}
+func (m effectiveChunker) Plan(context.Context, contracts.ChunkRequest) (contracts.ChunkPlanResult, error) {
+	return contracts.ChunkPlanResult{}, nil
+}
+
+type effectiveExtractor struct{ key string }
+
+func (m effectiveExtractor) Key() string                               { return m.key }
+func (m effectiveExtractor) ReferenceSlots() []contracts.ReferenceSlot { return nil }
+func (m effectiveExtractor) Extract(context.Context, contracts.TypedExtractionRequest) (contracts.TypedExtractionResult[effectiveArtifact], error) {
+	return contracts.TypedExtractionResult[effectiveArtifact]{}, nil
+}
+
+type effectiveMerger struct{ key string }
+
+func (m effectiveMerger) Key() string { return m.key }
+func (m effectiveMerger) Merge(context.Context, contracts.TypedMergeRequest[effectiveArtifact]) (contracts.TypedMergeResult[effectiveArtifact], error) {
+	return contracts.TypedMergeResult[effectiveArtifact]{}, nil
+}
+
+type effectiveNormalizer struct{ key string }
+
+func (m effectiveNormalizer) Key() string                               { return m.key }
+func (m effectiveNormalizer) ReferenceSlots() []contracts.ReferenceSlot { return nil }
+func (m effectiveNormalizer) Normalize(context.Context, contracts.TypedNormalizeRequest[effectiveArtifact]) (contracts.TypedNormalizeResult[effectiveArtifact], error) {
+	return contracts.TypedNormalizeResult[effectiveArtifact]{}, nil
+}
+
+type effectiveOutput struct{ key string }
+
+func (m effectiveOutput) Key() string { return m.key }
+func (m effectiveOutput) Encode(context.Context, contracts.OutputRequest) (contracts.OutputResult, error) {
+	return contracts.OutputResult{}, nil
+}
+
+type effectiveValidator struct {
+	name  string
+	class contracts.ExecutionClass
+}
+
+func (v effectiveValidator) Name() string                             { return v.name }
+func (v effectiveValidator) ExecutionClass() contracts.ExecutionClass { return v.class }
+func (v effectiveValidator) Validate(context.Context, contracts.TypedValidationRequest[effectiveArtifact]) (contracts.ValidationResult, error) {
+	return contracts.ValidationResult{Approved: true}, nil
+}
+
+func effectiveCatalog(t *testing.T) pipeline.ModuleCatalog {
+	t.Helper()
+	catalog := pipeline.ModuleCatalog{
+		Inputs:          pipeline.NewInputAdapterRegistry(),
+		Chunkers:        pipeline.NewChunkerRegistry(),
+		ArtifactCodecs:  pipeline.NewArtifactCodecRegistry(),
+		Extractors:      pipeline.NewExtractorRegistry(),
+		Mergers:         pipeline.NewMergerRegistry(),
+		Normalizers:     pipeline.NewNormalizerRegistry(),
+		Validators:      pipeline.NewValidatorRegistry(),
+		ValidatorChains: pipeline.NewValidatorChainRegistry(),
+		Outputs:         pipeline.NewOutputEncoderRegistry(),
+	}
+	if err := pipeline.RegisterArtifactCodec(catalog.ArtifactCodecs, effectiveCodec{}); err != nil {
+		t.Fatal(err)
+	}
+	if err := catalog.Inputs.RegisterWithSpec(pipeline.ModuleSpec{Key: "input", Stage: pipeline.StageInput, ExecutionClass: contracts.ExecutionClassDeterministic, Provides: []string{"source"}}, func() (contracts.InputAdapter, error) {
+		return effectiveInput{key: "input"}, nil
+	}); err != nil {
+		t.Fatal(err)
+	}
+	chunkSpec := pipeline.ModuleSpec{
+		Key:            pipeline.DefaultChunkModule,
+		Stage:          pipeline.StageChunk,
+		ExecutionClass: contracts.ExecutionClassLLMBacked,
+		Requires:       []string{"source"},
+		Provides:       []string{"chunk"},
+		ReferenceSlots: []contracts.ReferenceSlot{{Name: "chunk-ref"}},
+	}
+	chunkOptions := func(options map[string]any) error { return pipeline.RejectUnknownOptions(options, "size", "nested") }
+	if err := catalog.Chunkers.RegisterBuilderWithSpec(chunkSpec, chunkOptions, func(pipeline.BuildRequest) (contracts.Chunker, error) {
+		return effectiveChunker{key: pipeline.DefaultChunkModule}, nil
+	}); err != nil {
+		t.Fatal(err)
+	}
+	if err := catalog.Chunkers.RegisterBuilderWithSpec(pipeline.ModuleSpec{Key: "needs-capability", Stage: pipeline.StageChunk, ExecutionClass: contracts.ExecutionClassDeterministic, Requires: []string{"missing"}}, chunkOptions, func(pipeline.BuildRequest) (contracts.Chunker, error) {
+		return effectiveChunker{key: "needs-capability"}, nil
+	}); err != nil {
+		t.Fatal(err)
+	}
+	if err := pipeline.RegisterExtractor(catalog.Extractors, pipeline.ModuleSpec{Key: "extract", Stage: pipeline.StageExtract, ExecutionClass: contracts.ExecutionClassLLMBacked, ArtifactKind: effectiveArtifactKind, Requires: []string{"chunk"}, Provides: []string{"candidate"}}, func() (contracts.Extractor[effectiveArtifact], error) {
+		return effectiveExtractor{key: "extract"}, nil
+	}); err != nil {
+		t.Fatal(err)
+	}
+	if err := pipeline.RegisterMerger(catalog.Mergers, pipeline.ModuleSpec{Key: pipeline.DefaultMergeModule, Stage: pipeline.StageMerge, ExecutionClass: contracts.ExecutionClassLLMBacked, ArtifactKind: effectiveArtifactKind, Requires: []string{"candidate"}, Provides: []string{"merged"}}, func() (contracts.Merger[effectiveArtifact], error) {
+		return effectiveMerger{key: pipeline.DefaultMergeModule}, nil
+	}); err != nil {
+		t.Fatal(err)
+	}
+	if err := pipeline.RegisterMerger(catalog.Mergers, pipeline.ModuleSpec{Key: "other-merge", Stage: pipeline.StageMerge, ExecutionClass: contracts.ExecutionClassDeterministic, ArtifactKind: "other-kind", Requires: []string{"candidate"}, Provides: []string{"merged"}}, func() (contracts.Merger[effectiveArtifact], error) {
+		return effectiveMerger{key: "other-merge"}, nil
+	}); err != nil {
+		t.Fatal(err)
+	}
+	if err := pipeline.RegisterNormalizer(catalog.Normalizers, pipeline.ModuleSpec{Key: pipeline.DefaultNormalizeModule, Stage: pipeline.StageNormalize, ExecutionClass: contracts.ExecutionClassLLMBacked, ArtifactKind: effectiveArtifactKind, Requires: []string{"merged"}, Provides: []string{"normalized"}}, func() (contracts.Normalizer[effectiveArtifact], error) {
+		return effectiveNormalizer{key: pipeline.DefaultNormalizeModule}, nil
+	}); err != nil {
+		t.Fatal(err)
+	}
+	if err := catalog.Outputs.RegisterWithSpec(pipeline.ModuleSpec{Key: pipeline.DefaultOutputModule, Stage: pipeline.StageOutput, ExecutionClass: contracts.ExecutionClassDeterministic, Requires: []string{"normalized"}}, func() (contracts.OutputEncoder, error) {
+		return effectiveOutput{key: pipeline.DefaultOutputModule}, nil
+	}); err != nil {
+		t.Fatal(err)
+	}
+	for _, validator := range []struct {
+		key   string
+		class contracts.ExecutionClass
+	}{
+		{key: "default-validator", class: contracts.ExecutionClassDeterministic},
+		{key: "configured-a", class: contracts.ExecutionClassDeterministic},
+		{key: "configured-b", class: contracts.ExecutionClassDeterministic},
+		{key: "llm-validator", class: contracts.ExecutionClassLLMBacked},
+	} {
+		if err := pipeline.RegisterTypedValidatorBuilder(catalog.Validators, effectiveArtifactKind, pipeline.ValidatorSpec{Key: validator.key, ExecutionClass: validator.class}, func(options map[string]any) error {
+			return pipeline.RejectUnknownOptions(options, "nested")
+		}, func(pipeline.BuildRequest) (contracts.TypedValidator[effectiveArtifact], error) {
+			return effectiveValidator{name: validator.key, class: validator.class}, nil
+		}); err != nil {
+			t.Fatal(err)
+		}
+	}
+	if err := pipeline.RegisterTypedValidatorBuilder(catalog.Validators, effectiveArtifactKind, pipeline.ValidatorSpec{Key: "option-validator", ExecutionClass: contracts.ExecutionClassDeterministic}, func(options map[string]any) error {
+		return pipeline.RejectUnknownOptions(options, "allowed")
+	}, func(pipeline.BuildRequest) (contracts.TypedValidator[effectiveArtifact], error) {
+		return effectiveValidator{name: "option-validator", class: contracts.ExecutionClassDeterministic}, nil
+	}); err != nil {
+		t.Fatal(err)
+	}
+	if err := catalog.ValidatorChains.Register(pipeline.ValidatorChainMapping{Stage: pipeline.StageExtract, Module: "extract", Validators: []pipeline.ModuleBinding{pipeline.Binding("default-validator")}}); err != nil {
+		t.Fatal(err)
+	}
+	return catalog
+}

internal/core/config/effective_config_test.go

@@ -1,214 +0,0 @@
-package config
-
-import (
-	"strings"
-	"testing"
-	"time"
-
-	"gitea.maximumdirect.net/eric/notarius/internal/framework/pipeline"
-)
-
-func TestResolveRejectsEmptyAndUnknownPipelineID(t *testing.T) {
-	tests := []struct {
-		name       string
-		pipelineID string
-		want       string
-	}{
-		{name: "empty", pipelineID: " ", want: "pipeline id"},
-		{name: "unknown", pipelineID: "missing", want: "not configured"},
-	}
-
-	for _, tc := range tests {
-		t.Run(tc.name, func(t *testing.T) {
-			_, err := validConfig().Resolve(ResolveInput{PipelineID: tc.pipelineID, Catalog: fakeCatalog(t)})
-			if err == nil || !strings.Contains(err.Error(), tc.want) {
-				t.Fatalf("expected error containing %q, got %v", tc.want, err)
-			}
-		})
-	}
-}
-
-func TestResolveLaneFilteringSuccessAndFailure(t *testing.T) {
-	effective, err := validConfig().Resolve(ResolveInput{
-		PipelineID: " example ",
-		Only:       []string{" notes "},
-		Catalog:    fakeCatalog(t),
-	})
-	if err != nil {
-		t.Fatalf("Resolve: %v", err)
-	}
-
-	if effective.PipelineID != "example" {
-		t.Fatalf("unexpected pipeline ID: %q", effective.PipelineID)
-	}
-	if len(effective.ResolvedPipeline.ArtifactLanes) != 1 || effective.ResolvedPipeline.ArtifactLanes[0].ID != "notes" {
-		t.Fatalf("unexpected resolved lanes: %+v", effective.ResolvedPipeline.ArtifactLanes)
-	}
-	if effective.ResolvedPipeline.Digest == "" {
-		t.Fatalf("expected digest")
-	}
-
-	_, err = validConfig().Resolve(ResolveInput{
-		PipelineID: "example",
-		Only:       []string{"missing"},
-		Catalog:    fakeCatalog(t),
-	})
-	if err == nil || !strings.Contains(err.Error(), "selected artifact lane") {
-		t.Fatalf("expected invalid lane error, got %v", err)
-	}
-}
-
-func TestResolveUsesTrimmedPipelineMapKeys(t *testing.T) {
-	cfg := validConfig()
-	cfg.Pipelines[" example "] = cfg.Pipelines["example"]
-	delete(cfg.Pipelines, "example")
-
-	effective, err := cfg.Resolve(ResolveInput{PipelineID: "example", Catalog: fakeCatalog(t)})
-	if err != nil {
-		t.Fatalf("Resolve: %v", err)
-	}
-	if effective.PipelineID != "example" {
-		t.Fatalf("unexpected pipeline ID: %q", effective.PipelineID)
-	}
-}
-
-func TestResolveSurfacesUnknownModuleKeyThroughCatalog(t *testing.T) {
-	cfg := validConfig()
-	lane := cfg.Pipelines["example"].Artifacts["events"]
-	lane.Extract = pipeline.Binding("missing/extract")
-	cfg.Pipelines["example"].Artifacts["events"] = lane
-
-	_, err := cfg.Resolve(ResolveInput{PipelineID: "example", Catalog: fakeCatalog(t)})
-	if err == nil || !strings.Contains(err.Error(), "missing/extract") || !strings.Contains(err.Error(), "events") {
-		t.Fatalf("expected unknown module error with lane context, got %v", err)
-	}
-}
-
-func TestResolveSurfacesMissingCapabilityThroughCatalog(t *testing.T) {
-	_, err := validConfig().Resolve(ResolveInput{
-		PipelineID: "example",
-		Catalog: fakeCatalog(t, pipeline.ModuleSpec{
-			Key:      "json",
-			Stage:    pipeline.StageOutput,
-			Requires: []string{"missing-capability"},
-		}),
-	})
-	if err == nil || !strings.Contains(err.Error(), "missing capability") || !strings.Contains(err.Error(), "json") {
-		t.Fatalf("expected missing capability error, got %v", err)
-	}
-}
-
-func TestResolveDigestChangesWhenEffectiveConfigChanges(t *testing.T) {
-	cfg := validConfig()
-	first, err := cfg.Resolve(ResolveInput{PipelineID: "example", Catalog: fakeCatalog(t)})
-	if err != nil {
-		t.Fatalf("Resolve first: %v", err)
-	}
-
-	lane := cfg.Pipelines["example"].Artifacts["events"]
-	lane.Extract.Options = map[string]any{"temperature": 0.2}
-	cfg.Pipelines["example"].Artifacts["events"] = lane
-	second, err := cfg.Resolve(ResolveInput{PipelineID: "example", Catalog: fakeCatalog(t)})
-	if err != nil {
-		t.Fatalf("Resolve second: %v", err)
-	}
-
-	if first.ResolvedPipeline.Digest == second.ResolvedPipeline.Digest {
-		t.Fatalf("expected digest to change, got %q", first.ResolvedPipeline.Digest)
-	}
-}
-
-func TestResolveLLMProfileOverrideAppliesBeforeDigest(t *testing.T) {
-	cfg := validConfig()
-	cfg.LLMProfiles["runtime"] = LLMProfile{Provider: "openai-compatible"}
-
-	base, err := cfg.Resolve(ResolveInput{PipelineID: "example", Catalog: fakeCatalog(t)})
-	if err != nil {
-		t.Fatalf("Resolve base: %v", err)
-	}
-	effective, err := cfg.Resolve(ResolveInput{
-		PipelineID:         "example",
-		Catalog:            fakeCatalog(t),
-		LLMProfileOverride: "runtime",
-	})
-	if err != nil {
-		t.Fatalf("Resolve override: %v", err)
-	}
-
-	if base.ResolvedPipeline.Digest == effective.ResolvedPipeline.Digest {
-		t.Fatalf("expected digest to change after LLM profile override")
-	}
-	for _, binding := range resolvedBindings(effective.ResolvedPipeline) {
-		if binding.LLMProfile != "runtime" {
-			t.Fatalf("binding profile = %q, want runtime", binding.LLMProfile)
-		}
-	}
-
-	_, err = cfg.Resolve(ResolveInput{
-		PipelineID:         "example",
-		Catalog:            fakeCatalog(t),
-		LLMProfileOverride: "missing",
-	})
-	if err == nil || !strings.Contains(err.Error(), "LLM profile override") {
-		t.Fatalf("expected override profile error, got %v", err)
-	}
-}
-
-func resolvedBindings(resolved pipeline.ResolvedPipeline) []pipeline.ModuleBinding {
-	bindings := []pipeline.ModuleBinding{resolved.Input, resolved.Chunk, resolved.Output}
-	for _, lane := range resolved.ArtifactLanes {
-		bindings = append(bindings, lane.Extract, lane.Merge, lane.Normalize)
-		bindings = append(bindings, lane.Validators...)
-	}
-	return bindings
-}
-
-func TestOpenAICompatibleClientConfigRejectsIncompleteDefaultProfile(t *testing.T) {
-	cfg := Default()
-
-	_, err := cfg.OpenAICompatibleClientConfig("default")
-	if err == nil || !strings.Contains(err.Error(), "base URL") {
-		t.Fatalf("expected incomplete profile error, got %v", err)
-	}
-}
-
-func TestOpenAICompatibleClientConfigSuccess(t *testing.T) {
-	cfg := validConfig()
-	profile := cfg.LLMProfiles["default"]
-	profile.APIKey = "secret"
-	profile.TimeoutSeconds = 45
-	profile.MaxRetries = 4
-	cfg.LLMProfiles["default"] = profile
-
-	llmCfg, err := cfg.OpenAICompatibleClientConfig(" default ")
-	if err != nil {
-		t.Fatalf("OpenAICompatibleClientConfig: %v", err)
-	}
-
-	if llmCfg.BaseURL != "https://example.invalid/v1" || llmCfg.Model != "test-model" || llmCfg.APIKey != "secret" {
-		t.Fatalf("unexpected client config strings: %+v", llmCfg)
-	}
-	if llmCfg.MaxRetries != 4 {
-		t.Fatalf("unexpected max retries: %d", llmCfg.MaxRetries)
-	}
-	if llmCfg.RequestTimeout != 45*time.Second {
-		t.Fatalf("unexpected timeout: %s", llmCfg.RequestTimeout)
-	}
-}
-
-func TestOpenAICompatibleClientConfigRejectsUnknownAndUnsupportedProfiles(t *testing.T) {
-	_, err := validConfig().OpenAICompatibleClientConfig("missing")
-	if err == nil || !strings.Contains(err.Error(), "not configured") {
-		t.Fatalf("expected unknown profile error, got %v", err)
-	}
-
-	cfg := validConfig()
-	profile := cfg.LLMProfiles["default"]
-	profile.Provider = "unsupported"
-	cfg.LLMProfiles["default"] = profile
-
-	_, err = cfg.OpenAICompatibleClientConfig("default")
-	if err == nil || !strings.Contains(err.Error(), "provider") {
-		t.Fatalf("expected unsupported provider error, got %v", err)
-	}
-}

internal/core/config/env.go

@@ -6,7 +6,6 @@ import (
 	"strconv"
 	"strings"
 
-	"gitea.maximumdirect.net/eric/notarius/internal/core/diagnostics"
 	"gitea.maximumdirect.net/eric/notarius/internal/framework/pipeline"
 )
 
@@ -30,43 +29,6 @@ func (c *Config) applyEnvOverridesWithLookup(lookup func(string) (string, bool))
 	if c == nil {
 		return fmt.Errorf("config must not be nil")
 	}
-	if c.LLMProfiles == nil {
-		c.LLMProfiles = map[string]LLMProfile{}
-	}
-
-	defaultProfile := c.LLMProfiles[pipeline.DefaultLLMProfile]
-	if raw, ok := lookup("NOTARIUS_LLM_DEFAULT_API_KEY"); ok {
-		defaultProfile.APIKey = raw
-	}
-	if raw, ok := lookup("NOTARIUS_LLM_DEFAULT_BASE_URL"); ok {
-		defaultProfile.BaseURL = strings.TrimSpace(raw)
-	}
-	if raw, ok := lookup("NOTARIUS_LLM_DEFAULT_MODEL"); ok {
-		defaultProfile.Model = strings.TrimSpace(raw)
-	}
-	if raw, ok := lookup("NOTARIUS_LLM_DEFAULT_TIMEOUT_SECONDS"); ok {
-		value, err := parseIntEnv("NOTARIUS_LLM_DEFAULT_TIMEOUT_SECONDS", raw)
-		if err != nil {
-			return err
-		}
-		defaultProfile.TimeoutSeconds = value
-	}
-	if raw, ok := lookup("NOTARIUS_LLM_DEFAULT_MAX_RETRIES"); ok {
-		value, err := parseIntEnv("NOTARIUS_LLM_DEFAULT_MAX_RETRIES", raw)
-		if err != nil {
-			return err
-		}
-		defaultProfile.MaxRetries = value
-	}
-	if raw, ok := lookup("NOTARIUS_LLM_DEFAULT_MAX_CONCURRENCY"); ok {
-		value, err := parseIntEnv("NOTARIUS_LLM_DEFAULT_MAX_CONCURRENCY", raw)
-		if err != nil {
-			return err
-		}
-		defaultProfile.MaxConcurrency = value
-	}
-	c.LLMProfiles[pipeline.DefaultLLMProfile] = defaultProfile
-
 	if raw, ok := lookup("NOTARIUS_TOTAL_LLM_CONCURRENCY"); ok {
 		value, err := parseIntEnv("NOTARIUS_TOTAL_LLM_CONCURRENCY", raw)
 		if err != nil {
@@ -74,11 +36,60 @@ func (c *Config) applyEnvOverridesWithLookup(lookup func(string) (string, bool))
 		}
 		c.Concurrency.TotalLLM = value
 	}
-	if raw, ok := lookup("NOTARIUS_WORK_DIR"); ok {
+	if raw, ok := lookup("NOTARIUS_STAGE_WORKERS_EXTRACT"); ok {
-		c.Diagnostics.WorkDir = strings.TrimSpace(raw)
+		value, err := parseIntEnv("NOTARIUS_STAGE_WORKERS_EXTRACT", raw)
+		if err != nil {
+			return err
+		}
+		if c.Concurrency.StageWorkers == nil {
+			c.Concurrency.StageWorkers = make(map[string]int)
+		}
+		c.Concurrency.StageWorkers["extract"] = value
+		c.Concurrency.extractWorkersConfigured = true
 	}
-	if raw, ok := lookup("NOTARIUS_DIAGNOSTICS_RETENTION"); ok {
+	c.Concurrency.recomputeStageWorkerDefaults()
-		c.Diagnostics.Retention = diagnostics.RetentionMode(strings.TrimSpace(raw))
+	if raw, ok := lookup("NOTARIUS_OUTPUT_DIR"); ok {
+		c.Output.Directory = strings.TrimSpace(raw)
+		if c.Output.Directory == "" {
+			return fmt.Errorf("NOTARIUS_OUTPUT_DIR: must not be empty")
+		}
+		if strings.ContainsRune(c.Output.Directory, '\x00') {
+			return fmt.Errorf("NOTARIUS_OUTPUT_DIR: must not contain NUL")
+		}
+	}
+	if raw, ok := lookup("NOTARIUS_CACHE_CHUNK_PLANS_MODE"); ok {
+		mode, err := pipeline.ParseChunkCacheMode(raw)
+		if err != nil {
+			return fmt.Errorf("NOTARIUS_CACHE_CHUNK_PLANS_MODE: %w", err)
+		}
+		c.Cache.ChunkPlans.Mode = mode
+	}
+	if raw, ok := lookup("NOTARIUS_CACHE_CHUNK_PLANS_DIR"); ok {
+		c.Cache.ChunkPlans.Directory = cleanOptionalPath(raw)
+		if c.Cache.ChunkPlans.Directory == "" {
+			return fmt.Errorf("NOTARIUS_CACHE_CHUNK_PLANS_DIR: must not be empty")
+		}
+		if strings.ContainsRune(c.Cache.ChunkPlans.Directory, '\x00') {
+			return fmt.Errorf("NOTARIUS_CACHE_CHUNK_PLANS_DIR: must not contain NUL")
+		}
+	}
+	if raw, ok := lookup("NOTARIUS_CACHE_CHECKPOINTS_DIR"); ok {
+		c.Cache.Checkpoints.Directory = cleanOptionalPath(raw)
+		if c.Cache.Checkpoints.Directory == "" {
+			return fmt.Errorf("NOTARIUS_CACHE_CHECKPOINTS_DIR: must not be empty")
+		}
+		if strings.ContainsRune(c.Cache.Checkpoints.Directory, '\x00') {
+			return fmt.Errorf("NOTARIUS_CACHE_CHECKPOINTS_DIR: must not contain NUL")
+		}
+	}
+	if raw, ok := lookup("NOTARIUS_DEBUG_DIR"); ok {
+		c.Debug.Directory = strings.TrimSpace(raw)
+		if c.Debug.Directory == "" {
+			return fmt.Errorf("NOTARIUS_DEBUG_DIR: must not be empty")
+		}
+		if strings.ContainsRune(c.Debug.Directory, '\x00') {
+			return fmt.Errorf("NOTARIUS_DEBUG_DIR: must not contain NUL")
+		}
 	}
 	return nil
 }

internal/core/config/env_contract_test.go

@@ -0,0 +1,239 @@
+package config
+
+import (
+	"errors"
+	"reflect"
+	"strings"
+	"testing"
+
+	"gitea.maximumdirect.net/eric/notarius/internal/framework/pipeline"
+)
+
+func TestPrecedenceFileValuesOverrideBuiltInDefaults(t *testing.T) {
+	cfg := applyFileConfig(t, `version: 4
+concurrency:
+  total_llm: 4
+  stage_workers:
+    extract: 2
+output:
+  directory: ./file-output
+cache:
+  chunk_plans:
+    directory: ./file-plans
+    mode: refresh
+  checkpoints:
+    directory: ./file-checkpoints
+debug:
+  directory: ./file-debug
+`)
+	if cfg.Concurrency.TotalLLM != 4 || cfg.Concurrency.StageWorkers["extract"] != 2 ||
+		cfg.Output.Directory != "./file-output" || cfg.Cache.ChunkPlans.Directory != "file-plans" ||
+		cfg.Cache.ChunkPlans.Mode != pipeline.ChunkCacheRefresh || cfg.Cache.Checkpoints.Directory != "file-checkpoints" ||
+		cfg.Debug.Directory != "./file-debug" {
+		t.Fatalf("file values did not override defaults: %#v", cfg)
+	}
+}
+
+func TestPrecedenceOperationalEnvironmentOverridesFileValues(t *testing.T) {
+	cfg := applyFileConfig(t, `version: 4
+concurrency:
+  total_llm: 2
+  stage_workers:
+    extract: 1
+output:
+  directory: ./file-output
+cache:
+  chunk_plans:
+    directory: ./file-plans
+    mode: refresh
+  checkpoints:
+    directory: ./file-checkpoints
+debug:
+  directory: ./file-debug
+`)
+	env := map[string]string{
+		"NOTARIUS_TOTAL_LLM_CONCURRENCY":  "8",
+		"NOTARIUS_STAGE_WORKERS_EXTRACT":  "6",
+		"NOTARIUS_OUTPUT_DIR":             "/env/output",
+		"NOTARIUS_CACHE_CHUNK_PLANS_MODE": "bypass",
+		"NOTARIUS_CACHE_CHUNK_PLANS_DIR":  "/env/plans",
+		"NOTARIUS_CACHE_CHECKPOINTS_DIR":  "/env/checkpoints",
+		"NOTARIUS_DEBUG_DIR":              "/env/debug",
+	}
+	if err := cfg.ApplyEnvOverridesWithLookup(lookupValues(env)); err != nil {
+		t.Fatal(err)
+	}
+	if cfg.Concurrency.TotalLLM != 8 || cfg.Concurrency.StageWorkers["extract"] != 6 ||
+		cfg.Output.Directory != "/env/output" || cfg.Cache.ChunkPlans.Directory != "/env/plans" ||
+		cfg.Cache.ChunkPlans.Mode != pipeline.ChunkCacheBypass || cfg.Cache.Checkpoints.Directory != "/env/checkpoints" ||
+		cfg.Debug.Directory != "/env/debug" {
+		t.Fatalf("environment values did not override file values: %#v", cfg)
+	}
+}
+
+func TestPrecedenceExtractWorkersFollowEffectiveConcurrencyUnlessExplicit(t *testing.T) {
+	tests := []struct {
+		name       string
+		file       string
+		env        map[string]string
+		wantTotal  int
+		wantWorker int
+	}{
+		{
+			name:       "default follows environment total",
+			file:       "version: 4\n",
+			env:        map[string]string{"NOTARIUS_TOTAL_LLM_CONCURRENCY": "5"},
+			wantTotal:  5,
+			wantWorker: 5,
+		},
+		{
+			name:       "file worker is retained",
+			file:       "version: 4\nconcurrency:\n  total_llm: 3\n  stage_workers:\n    extract: 2\n",
+			env:        map[string]string{"NOTARIUS_TOTAL_LLM_CONCURRENCY": "6"},
+			wantTotal:  6,
+			wantWorker: 2,
+		},
+		{
+			name: "environment worker is retained",
+			file: "version: 4\nconcurrency:\n  total_llm: 2\n",
+			env: map[string]string{
+				"NOTARIUS_TOTAL_LLM_CONCURRENCY": "6",
+				"NOTARIUS_STAGE_WORKERS_EXTRACT": "4",
+			},
+			wantTotal:  6,
+			wantWorker: 4,
+		},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			cfg := applyFileConfig(t, tt.file)
+			if err := cfg.ApplyEnvOverridesWithLookup(lookupValues(tt.env)); err != nil {
+				t.Fatal(err)
+			}
+			if cfg.Concurrency.TotalLLM != tt.wantTotal || cfg.Concurrency.StageWorkers["extract"] != tt.wantWorker {
+				t.Fatalf("concurrency = %#v, want total %d and extract %d", cfg.Concurrency, tt.wantTotal, tt.wantWorker)
+			}
+		})
+	}
+}
+
+func TestPrecedenceEmptyFileCacheDirectoriesDeferPerUserResolution(t *testing.T) {
+	cfg := applyFileConfig(t, `version: 4
+cache:
+  chunk_plans:
+    directory: ""
+  checkpoints:
+    directory: ""
+`)
+	if err := cfg.Validate(); err != nil {
+		t.Fatalf("empty file cache directories should be valid: %v", err)
+	}
+	if cfg.Cache.ChunkPlans.Directory != "" || cfg.Cache.Checkpoints.Directory != "" {
+		t.Fatalf("empty cache directories were not preserved for deferred resolution: %#v", cfg.Cache)
+	}
+	resolver := func() (string, error) { return "/user/cache", nil }
+	chunkPlans, err := DefaultChunkPlanRoot(resolver)
+	if err != nil {
+		t.Fatal(err)
+	}
+	checkpoints, err := DefaultCheckpointRoot(resolver)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if chunkPlans != "/user/cache/notarius/chunk-plans" || checkpoints != "/user/cache/notarius/checkpoints" {
+		t.Fatalf("deferred cache roots = %q, %q", chunkPlans, checkpoints)
+	}
+}
+
+func TestDefaultCacheRootsRejectInvalidUserCacheResolvers(t *testing.T) {
+	tests := []struct {
+		name     string
+		resolver func() (string, error)
+		want     string
+	}{
+		{name: "nil resolver", want: "must not be nil"},
+		{
+			name: "resolver failure",
+			resolver: func() (string, error) {
+				return "", errors.New("cache home unavailable")
+			},
+			want: "resolve user cache directory",
+		},
+		{name: "empty directory", resolver: func() (string, error) { return "  ", nil }, want: "must not be empty"},
+	}
+	families := []struct {
+		name string
+		root func(func() (string, error)) (string, error)
+	}{
+		{name: "chunk plans", root: DefaultChunkPlanRoot},
+		{name: "checkpoints", root: DefaultCheckpointRoot},
+	}
+
+	for _, family := range families {
+		for _, tt := range tests {
+			t.Run(family.name+"/"+tt.name, func(t *testing.T) {
+				_, err := family.root(tt.resolver)
+				if err == nil || !strings.Contains(err.Error(), tt.want) {
+					t.Fatalf("error = %v, want substring %q", err, tt.want)
+				}
+			})
+		}
+	}
+}
+
+func TestEnvEmptyDirectoryOverridesAreErrors(t *testing.T) {
+	tests := []string{
+		"NOTARIUS_OUTPUT_DIR",
+		"NOTARIUS_CACHE_CHUNK_PLANS_DIR",
+		"NOTARIUS_CACHE_CHECKPOINTS_DIR",
+		"NOTARIUS_DEBUG_DIR",
+	}
+	for _, name := range tests {
+		t.Run(name, func(t *testing.T) {
+			cfg := Default()
+			err := cfg.ApplyEnvOverridesWithLookup(lookupValues(map[string]string{name: " \t"}))
+			if err == nil || !strings.Contains(err.Error(), name) {
+				t.Fatalf("error = %v, want responsible environment variable", err)
+			}
+		})
+	}
+}
+
+func TestEnvInvalidIntegersAndChunkCacheModesReportTheirNames(t *testing.T) {
+	tests := map[string]string{
+		"NOTARIUS_TOTAL_LLM_CONCURRENCY":  "not-an-integer",
+		"NOTARIUS_STAGE_WORKERS_EXTRACT":  "not-an-integer",
+		"NOTARIUS_CACHE_CHUNK_PLANS_MODE": "not-a-cache-mode",
+	}
+	for name, value := range tests {
+		t.Run(name, func(t *testing.T) {
+			cfg := Default()
+			err := cfg.ApplyEnvOverridesWithLookup(lookupValues(map[string]string{name: value}))
+			if err == nil || !strings.Contains(err.Error(), name) {
+				t.Fatalf("error = %v, want responsible environment variable", err)
+			}
+		})
+	}
+}
+
+func TestEnvRemovedProviderVariablesAreIgnored(t *testing.T) {
+	before := Default()
+	cfg := Default()
+	removed := map[string]string{
+		"NOTARIUS_LLM_DEFAULT_ENDPOINT": "ignored-provider-setting",
+		"NOTARIUS_LLM_DEFAULT_MODEL":    "ignored-provider-setting",
+	}
+	if err := cfg.ApplyEnvOverridesWithLookup(lookupValues(removed)); err != nil {
+		t.Fatal(err)
+	}
+	if !reflect.DeepEqual(cfg, before) {
+		t.Fatalf("removed provider variables changed configuration: %#v", cfg)
+	}
+}
+
+func lookupValues(values map[string]string) func(string) (string, bool) {
+	return func(name string) (string, bool) {
+		value, ok := values[name]
+		return value, ok
+	}
+}