notarius/docs/internal/pipeline.md

# Pipeline Internals

The implemented resolver and runner live in `internal/framework/pipeline`.
Their fixed workflow and ownership boundaries are defined by
[Architecture](../policy/architecture.md#system-shape). Configuration fields,
defaults, and selectable keys are defined in
[Configuration](../config.md#pipelines).

Resolution fixes the ordered steps, selected lanes, and all stage bindings;
preparation constructs every selected implementation before the runner begins
source work. After serial input parsing and plan selection or generation, the
runner materializes chunks and executes one step at a time. Within a step,
extract work uses one bounded run-wide worker pool in chunk-first, lane-second
order. Each lane's merge and normalize operations remain serial, and lanes in
the same step may overlap once their extracts are terminal. A later step cannot
start across its barrier until every earlier lane is terminal and each required
generated artifact has been accepted and handed off.

## Resolution

`internal/core/config.Config.Resolve` validates the loaded configuration,
selects the named profile, applies the runtime inputs supplied by the CLI, and
calls `pipeline.ResolvePipeline`.

`ResolvePipeline`:

1. selects the explicit ordered steps, or creates the implicit `default` step
   from the legacy top-level `artifacts` map;
2. selects and sorts artifact lanes within each step while enforcing global lane
   identity;
3. completes omitted bindings using the documented configuration defaults;
4. looks up each module and validator spec without constructing it;
5. for a typed extractor, derives its artifact kind, requires the codec, and
   selects exact-type merger, normalizer, and validator variants;
6. checks required and provided capabilities in workflow order;
7. resolves external and generated target-aware reference bindings and
   validates producer order, consumer slot declarations, and artifact-kind
   compatibility;
8. validates each selected module and validator option set through its registry
   entry; and
9. calculates a digest over the resolved structure, including step order, step
   IDs, lane membership, generated topology, producer and consumer identities,
   typed artifact kind and schema identity, and the effective validator policy
   in its resolved execution order.

Resolution returns a `ResolvedPipeline` containing ordered steps, lanes,
concrete bindings, validator chains, reference targets, and the digest. It does
not read external reference bytes or construct runtime modules. CLI lane and
reference selector syntax is defined in the [CLI reference](../cli.md#run).

The digest includes each resolved step's ID and lane membership, generated
producer/consumer topology, and each validator chain's stage, lane, owning
module, ordered validator bindings, execution classes, targets, and artifact
kinds. Changing step order, a dependency, a default chain, or an explicit
override therefore changes pipeline identity whenever it changes effective
execution policy.

## Reference Materialization

The CLI calls `MaterializeReferences` after resolution and before constructing
the LLM client or running the pipeline. For external bindings, the materializer
checks each binding against its resolved target declaration, reads and validates
the file, and builds both a `contracts.ReferenceSet` and provenance-only
metadata on the corresponding `ResolvedReferenceTarget`. A structured
generated binding is declaration-only at this point: its producer bytes do not
exist until the producer lane reaches an accepted normalized result.

Preparation delivers the materialized external set for each target through
`pipeline.BuildRequest`: chunkers and chunk validators receive the chunk target;
extractors and extract validators receive the lane extract target; mergers and
merge validators receive the lane merge target; and normalizers and normalize
validators receive the lane normalize target. Input and output builders receive
an empty set because those stages cannot declare references. Every builder gets
an isolated deep clone of its target set, so construction-time mutation cannot
change another builder, the resolved pipeline, or later runtime requests.

Prepared consumers do not need to be reconstructed when generated content is
available. At the step boundary, the runner encodes the accepted producer value
through its registered canonical codec, validates the generated bytes against
each target slot's kind, schema, media type, and size, and clones one immutable
reference item into the operation request. The item includes canonical digest,
size, and bounded producer provenance but no filesystem URI. A handoff failure
is a framework dependency error and prevents every consumer in that step from
starting.

The runner continues to clone the resulting set into the chunk, extract, merge,
or normalize request that owns the target. LLM-backed extensions may convert
those items into named prompt inputs. Reference content remains separate from
source evidence and source digests, whether the item came from a file or a
generated handoff.

Binding precedence, path resolution, accepted content, and media-type behavior
are configuration contracts; see [Configuration](../config.md#pipelines).
Durable provenance is defined in the
[JSON output contract](../integrations/json-output.md#manifestjson), while
runtime sensitive-data handling belongs in [Operations](../operations.md).

## Registries And Specs

`pipeline.Registries` holds option validators and run-local builders used during
resolution and preparation.
`pipeline.ModuleCatalog` exposes their specs during configuration validation and
resolution. Separate registries exist for every stage and for validators;
`ValidatorChainRegistry` stores production default-chain mappings. Both
containers also carry an `ArtifactCodecRegistry`. Generic registration records
one codec per stable artifact kind, validates its schema metadata and JSON
Schema, retains the exact schema digest and Go type, and safely encodes or
decodes framework-erased values with typed errors on incompatibility.

Typed extractor entries are keyed by module key and declare one artifact kind.
Merger, normalizer, and typed-validator variants are keyed by module or
validator key plus artifact kind. Chunk and serialized validators occupy
separate target namespaces; serialized registrations declare whether they
support chunks, artifacts, or both. Duplicate variants and exact Go-type
mismatches are rejected deterministically.

Lane-sensitive merger and normalizer spec discovery always supplies the
extractor's artifact kind, so variants under one reusable key may declare
different capabilities and reference slots. Kind-neutral registry inspection
selects the first registered artifact kind in sorted order.

Production composition registers the D&D spell-list, NPC-list, combat-turn-list,
NPC-interaction-list, and scene-description-list codecs and typed lane
variants, plus serialized JSON validators. Every artifact lane resolves through
the typed registries and a matching codec.

A `ModuleSpec` declares its stage plus required and provided capabilities.
Chunk, extract, merge, and normalize specs may also declare reference slots.
Registry implementations defensively copy spec metadata, reject duplicate keys,
and verify that a constructed implementation reports the registered key.
Builder registrations accept `ModuleDependencies` and cloned configuration
options through one `BuildRequest`. Builders decode those options and retain
typed values or injected dependencies in the constructed implementation.
Extractors declare their artifact kind, and merger, normalizer, and validator
resolution selects the matching typed variant.

A `ValidatorSpec` declares a validator key and execution class. Resolution uses
the execution class to reject incompatible profile bindings before execution.
The current production catalog and default chain are listed only in
[Configuration](../config.md#implemented-production-validators).

## Preparation And Runner Boundary

`pipeline.Prepare` receives a resolved pipeline, the registries, and shared
module dependencies. It constructs input; chunk and its validators; every
step's lane extract, merge, and normalize modules and validator chains in
resolved order; then output. It stops at the first error with pipeline, step,
stage, lane, module, and validator context as applicable. It never invokes an
operation method. Generated references are not available during preparation;
the operation request is the handoff boundary.

`PreparedPipeline` keeps private constructed executors and exposes cloned
resolved input, chunk, lane, and output identities. Prepared components may
implement `pipeline.CheckpointFingerprintProvider` to contribute explicit
semantic identities to checkpoint reuse. Preparation trims and validates each
non-secret name and value, prefixes it with the component's stage, lane,
module, and validator scope, rejects duplicates, and retains the resulting
sorted collection behind a defensive-copy accessor. Fingerprints must be
stable and must not contain source content, credentials, local paths,
timestamps, or other invocation-specific values.

`pipeline.RunInput` carries that prepared pipeline, raw source input, run identity and timing, optional
session and profile metadata, a chunk-plan store and mode, a checkpoint
execution policy, and checkpoint/debug collaborators. The runner
parses source bytes through the already constructed input adapter. Later stage
requests receive the generic source model; extract requests receive
chunk-scoped input material, while chunk, merge, and normalize requests retain
access to the original source material. Input, chunk, and output operation
requests do not carry raw module options. The chunk request also does not carry
an LLM client; an LLM-backed chunker receives the shared client during
preparation. Their operation requests retain run-specific source, reference,
profile, session, metadata, and step-handoff context as applicable. A generated
reference is cloned into each compatible consumer request and is never exposed
as a path.

Prepared lanes retain exact-type-checked erased operation closures. The runner
uses those closures to keep each value typed through extraction, validation,
merge, and normalization.

Source validation requires every unit to carry a canonical self-reference to
its containing document and its own unit ID. Explicit clone, checkpoint, and
debug boundaries retain that reference, and the canonical source digest covers
it deterministically. Chunks use the same source model and carry one canonical
reference spanning the first selected unit through the last.

`pipeline.RunOutput` carries the run manifest, accepted normalized serialized
artifacts with lane and normalizer provenance,
rejected results, warnings, checkpoint events, and logical files returned by the
output encoder. The CLI owns debug-summary and durable filesystem writes after
the runner returns.

## Execution Flow

The pipeline-wide coordinator owns the ordered step loop, generated-reference
sets at each barrier, and deterministic merging of step outcomes. For one step,
one run-local lane engine owns worker lifecycle, cancellation, dispatch,
continuation queues, and result collection. It initializes checkpoint state in
lane order, dispatches bounded extract work, advances terminal lanes through
serial merge and normalize work, selects failures by stable pipeline scope, and
merges lane-local outcomes back in resolved order. Completion timing never
becomes public ordering.

The runner:

1. validates its prepared input;
2. parses the raw input with the prepared adapter and validates the generic
   source document;
3. selects a stored plan or executes the configured chunker's `Plan` operation;
4. canonicalizes and materializes the plan, then validates the resulting
   chunks;
5. builds the framework-owned accepted chunk map from the accepted source,
   logical plan, and exact materialized chunks, then supplies it to the output
   request independently of output-module options;
6. executes each resolved step in configuration order. For one step, it
   dispatches extract jobs in source-chunk then resolved-lane order, starts a
   bounded lane continuation when all extracts for that lane are terminal, and
   waits for every lane to become terminal;
7. encodes and validates each accepted normalized producer artifact, then
   builds the immutable generated reference sets for the next step;
8. invokes the prepared output encoder only after every step succeeds and
   validates its logical file results;
9. returns the assembled manifest, outcomes, warnings, and files.

Within each artifact lane, it reuses the prepared extractor, merger, normalizer,
and validators while performing these transitions:

1. extract once per accepted chunk and add runner-owned lane, source, and chunk
   provenance;
2. validate each extract result and omit rejected results from merge input;
3. skip the rest of the lane when no extract result is accepted;
4. merge accepted extract results in their existing order;
5. validate the merge result and skip normalization on rejection;
6. normalize the accepted merge result;
7. validate and append the accepted normalized result.

At a step barrier, a lane with no accepted normalized output is still a regular
rejection unless a later generated binding names that lane as a required
producer. In that case the runner raises a deterministic dependency error and
does not start the consumer step. One accepted typed artifact may fan out to
multiple compatible target slots. Consumers in the same step may run
concurrently after the handoff; no work crosses the barrier early.

Module-provided warnings and payload warnings are promoted only from attempts
whose results are accepted and used.

## Chunk Plans And Reuse

`Chunker.Plan` returns a `source.ChunkPlan`: the canonical source digest,
ordered unit-ID ranges, and optional plan or range annotations. The framework
owns plan canonicalization and materialization. It creates the generic chunks
and therefore owns their IDs, indexes, source references, JSON content, units,
media type, and generic metadata. Plan and range annotations are independently
owned raw JSON and become `Chunk.PlanAnnotations` and `Chunk.Annotations`.

In `auto`, the runner looks up the source digest before invoking the chunker. A
valid hit is materialized and sent through the current run's configured chunk
validators; it does not invoke the chunk module, consume its retry budget, or
make a chunk-stage LLM call. A missing, invalid, or unmaterializable record
generates a candidate. `refresh` generates without lookup; `bypass` generates
without cache access. Generated plans are published only after the full chunk
validator chain approves them. A validator rejection is a regular rejected
pipeline outcome and never replaces a cached plan.

The store is source-addressed, not pipeline-addressed. Changes to pipeline
configuration, requested chunker, options, references, lanes, validators, or
LLM profile do not prevent a source-digest hit. The manifest records both the
currently requested chunker and the effective plan producer. Cache state and
paths are configured and operated outside the runner; see
[Configuration](../config.md#state-surfaces) and [Operations](../operations.md).

For an accepted plan, the runner also constructs the strict framework-owned
[Accepted Chunk Map](../integrations/chunk-map.md) before lane execution. It
uses the current resolved chunker as `requested_chunker` and the stored or
generated record as `producer`, preserving that distinction on reuse. Chunk
rejection supplies no map; later lane rejection does not discard it. Output
encoders receive a defensively owned serialized value and may explicitly
ignore it.

The extract job channel has the same capacity as the effective extract worker
count, so dispatch applies backpressure. A fixed continuation executor prevents
ready or checkpoint-reused lanes from creating one goroutine each. Workers and
continuations publish lane-local results; the coordinator is the only writer of
aggregate output and merges those results in resolved lane and source-chunk
order.

## Plan Canonicalization And Chunk Materialization

Plan canonicalization requires canonical JSON annotations, a matching source
digest, at least one range, existing ordered boundaries, and increasing range
starts. Ranges may overlap or leave gaps; a chunker may impose stricter policy.
Materialization deterministically reconstructs each range from the current
source document, deep-clones JSON-shaped source-unit metadata, and copies
annotations without interpreting their namespaces. Materialized chunks and
separate materializations do not share mutable unit metadata; unsupported or
cyclic metadata fails materialization with context.

Before lane execution, generic chunk validation checks the materialized chunks'
identities, order, source references, content, media type, units, and metadata.
No chunk checkpoint participates in plan selection: plan storage is the only
chunk-reuse mechanism. Extract, merge, and normalize checkpoints continue to
use materialized chunk digests as their dependencies.

## Validation And Retries

Chunk, extract, merge, and normalize results pass through the resolved validator
chain for their stage and module. Chunk validators receive canonical chunks;
typed validators receive the domain value; and serialized validators receive
canonical chunk JSON or artifact codec bytes. Validators execute in resolved
order and stop at the first error or rejection. An empty chain approves the
result.

Production D&D artifact chains keep generic JSON syntax validation first, then
run every rejecting domain validator before generic JSON Schema validation. The
domain validator therefore owns expected semantic diagnostics; the generic
schema validator remains the final rejecting representation backstop, before
warning-only relatedness validation. Explicitly configured validator chains
retain their configured order.

`runWithRetry` applies the effective retry policy around module execution and
its complete validation chain. A module or validator error becomes a framework
error when attempts are exhausted. A rejection becomes a recorded
`RejectedOutput` when attempts are exhausted. Cancellation stops retry
processing immediately.

Structured-completion adapters classify malformed or undecodable provider
output with the provider-neutral `contracts.ErrInvalidStructuredOutput` error.
A typed normalizer may turn that condition, or another unsafe proposal, into a
normalize retry directive with a module-supplied safe candidate, stable
diagnostic, and fallback warnings. The directive consumes the same configured
normalize retry budget: `retries` permits that many additional attempts after
the initial attempt. It neither creates a normalizer-local retry loop nor
records an accepted checkpoint for the discarded attempt.

Before adding a normalize retry directive to attempt debug data, the runner
requires a nonblank, valid UTF-8 reason code of at most 128 bytes and a
nonblank, valid UTF-8 message of at most 4,096 bytes. These are encoded-byte
limits. The framework rejects an invalid directive without truncating or
rewriting either field. It validates only this mechanical contract; normalizers
remain responsible for ensuring their otherwise valid diagnostics do not expose
source material, credentials, paths, names, or other sensitive content.

The framework treats a module-supplied candidate as opaque. The normalizer owns
its safety determination, and the configured normalizer validator chain remains
the acceptance boundary for the final fallback.

When a later normalize attempt succeeds, its candidate alone proceeds through
the usual validation and checkpoint path. When the final attempt still returns
a directive, the runner validates its supplied safe fallback through that same
normalizer validator chain before accepting or rejecting it. Ordinary
attempt-local warnings and fallback warnings remain unpromoted while another
attempt is available; only final exhaustion promotes the supplied fallback
warnings. Rejected output is a non-fatal pipeline outcome and does not advance.
Configuration owns retry counts and validator overrides; see
[Module Bindings](../config.md#module-bindings).

## Checkpoint And Debug Hooks

The runner depends on recorder and loader interfaces, using no-op
implementations when collaborators are absent. Each checkpointed workflow
boundary records a running, succeeded, or failed transition. Reuse decisions
are consulted in workflow order and accepted payloads are cloned before
entering the normal handoff path. Typed extract, merge, and normalize
checkpoints store codec bytes with artifact kind, schema ID, name, version and
exact digest, and media type. Reuse compares that identity with the prepared
codec and decodes through the codec; missing identity, mismatches, corrupt
bytes, and decode failures become explicit reuse misses and execute the lane
normally. Dependency fingerprints and debug content digests use the same stable
codec bytes that cross those boundaries.

That progressive extract, merge, and normalize reuse is the ordinary resume
path. A lane marked as a required predecessor for selective recomputation takes
a separate accepted-output path before extract scheduling. The loader reads the
existing successful normalize manifest and payload by step, lane, and
normalizer, without consulting extract or merge dependencies. It requires the
current non-empty checkpoint identity to match, so the invocation identity
still binds the input, resolved topology and configuration, references, runtime
overrides, profiles, and component fingerprints.

The runner decodes and canonically re-encodes each reusable artifact once with
the prepared codec, requiring exact kind, schema identity and digest, media
type, canonical bytes, content digest, and producer provenance. A valid accepted
producer becomes a runner-owned cloned normalized output, restores only
normalize-checkpoint warnings, and records one `accepted_artifact_reused`
normalize decision. It does not invoke or record extract, merge, normalize, or
their validators. Invalid or unavailable accepted state records its decision
and fails the producer step; the dependent step never starts and the producer
is not implicitly rerun. If a later required lane fails during initialization,
already hydrated terminal lanes remain in the failed output in resolved order.

Generated references add downstream dependencies containing the producer's
artifact kind, complete schema identity, media type, canonical content digest,
and size. Compatible accepted producer outputs may therefore feed a later step
without re-executing the producer. Forced lanes bypass accepted-output
hydration and execute normally. A missing, rejected, corrupt, incompatible, or
changed producer blocks its dependent while leaving independent work eligible
for reuse. The runner records bounded decision
categories: `reused`, `executed`, `forced_recompute`, and
`dependency_invalidated`. Operator meanings for the stable reason codes belong
to [Operations](../operations.md#resume-and-selective-recompute).

The CLI includes prepared-component fingerprints in the run-wide checkpoint
identity alongside resolved configuration, raw input, reference provenance,
runtime overrides, and LLM-profile fingerprints. Module metadata is not used
implicitly for cache identity: components opt in only with stable semantic
values that can change accepted output. Adding or changing a component
fingerprint intentionally produces a cold cache miss. Existing checkpoint
schemas and paths remain unchanged.

The CLI's `--recompute-step` policy forces the selected step and all transitive
dependents, but requires accepted normalized artifacts for every unselected
producer on which that closure depends. It changes execution policy only; it
does not alter persistent checkpoint identity.

Debug instrumentation wraps run, stage, attempt, validator, and structured LLM
boundaries. Every executed chunk, extract, merge, and normalize attempt writes
one terminal envelope for acceptance, validator rejection, module or validator
error, or applicable candidate or final serialization error. The envelope
contains its attempt-local warnings, any available candidate and rejection,
and terminal error text; normalize retry directives retain their attempt-local
candidate and diagnostic, while only the final safe fallback reaches validation.
Failures before a candidate exists omit that payload.
Only LLM calls made by the module operation belong to the module attempt.
Validator calls retain independent scopes under `validate/` and are not
duplicated into the module envelope. A failed terminal-envelope write is a
non-retryable framework error and is joined with any primary attempt error.
Debug data is never used as a checkpoint source. Typed artifact debug envelopes
are domain-neutral, redact sensitive metadata and bytes through the common
debug policy, and record codec identity plus schema and content digests.

Merge and normalize attempts serialize their in-memory candidate with the
codec's required candidate encoder before typed validation. Serialized
validators and attempt debug use that candidate representation, which carries
the codec media type and schema identity but is never checkpointed or passed
downstream. Only a validator-approved value is encoded through the strict final
codec and made eligible for a checkpoint or stage output.

Checkpoint identity, physical layout, reuse behavior, and debug artifact
handling are operator contracts in [Operations](../operations.md). Serialization
and recorder implementation are inventoried in
[Internal Overview](overview.md#run-state-components).

## Results And Failures

The runner owns manifest assembly and handoff summaries but not the durable JSON
schema. It records resolved module and lane provenance, validator chains,
source/reference identities, selected LLM profiles, normalized and rejected
summaries, status, and timing. Serialized artifact content remains outside the manifest.
Module metadata providers may add non-secret singleton or lane-scoped metadata.

Execution errors include stage, module, lane, or validator context. Once a
manifest exists, a failing run returns it with failed status and completion
time. Successful status reflects whether any result was rejected. The
durable manifest and logical file schemas are defined in the
[JSON output contract](../integrations/json-output.md).

On a framework failure, the runner cancels its derived context, stops submitting
new extract work, drains started tasks, and skips the output encoder. Parent
cancellation takes precedence. Otherwise context-cancellation fallout is
discarded when a substantive error exists, and the primary error is selected by
stage, resolved lane, and source chunk rather than completion time.

## Tests To Inspect

- `internal/core/config/effective_config_test.go`: config-to-resolution boundary.
- `internal/framework/pipeline/profile_test.go`: selection, defaults,
  capabilities, validator chains, and digest behavior.
- `internal/framework/pipeline/artifact_codec_registry_test.go`: typed codec
  metadata, registration, erasure safety, strict decoding, and cloning.
- `internal/framework/pipeline/typed_resolution_test.go`: heterogeneous typed
  lane resolution and preparation, target-specific validators,
  incompatibilities, ordering, and schema-sensitive pipeline identity.
- `internal/framework/pipeline/runner_concurrency_test.go`: bounded dispatch and
  continuations, reverse completion, stable errors, rejection, cancellation,
  retries, and independent provider-call limits.
- `internal/framework/pipeline/preparation_test.go`: option validation,
  construction order, dependency failures, and the before-source-work boundary.
- `internal/framework/pipeline/references_test.go`: target resolution and
  materialization.
- `internal/cli/run_contract_test.go`: production run transitions, retries,
  rejections, warnings, CLI recomputation controls, debug hooks, and manifests.
- `internal/cli/recompute_execution_contract_test.go`: filesystem-backed
  selective recomputation and accepted-producer recovery.
- `internal/cli/production_contract_test.go`: production composition and
  configuration-resolution smoke coverage.
- `internal/cli/example_contract_test.go`: maintained example resolution and
  execution ownership.
- `internal/modules/integration/*_test.go` and
  `internal/modules/seriatim/input/transcript/runner_test.go`: typed runner
  composition across concrete module families.
- `internal/framework/checkpoint/*_test.go`: checkpoint serialization and reuse
  collaborators.