Prepare the v2.2.0 release

.woodpecker/release.yml

@@ -3,12 +3,12 @@ when:
 
 steps:
   validate-release:
-    image: golang:1.25.5
+    image: golang:1.26.7
     commands:
       - ./scripts/check-release-candidate.sh "$CI_COMMIT_TAG"
 
   build-release-assets:
-    image: golang:1.25.5
+    image: golang:1.26.7
     depends_on:
       - validate-release
     commands:

.woodpecker/shuffle.yml

@@ -3,6 +3,6 @@ when:
 
 steps:
   shuffled-race-tests:
-    image: golang:1.25
+    image: golang:1.26
     commands:
       - go test -race -shuffle=on -count=3 ./...

.woodpecker/verify.yml

@@ -3,44 +3,44 @@ when:
 
 steps:
   tests:
-    image: golang:1.25
+    image: golang:1.26
     commands:
       - go test ./...
 
   race-tests:
-    image: golang:1.25
+    image: golang:1.26
     depends_on: tests
     commands:
       - go test -race ./...
 
   static-analysis:
-    image: golang:1.25
+    image: golang:1.26
     depends_on: tests
     commands:
       - go vet ./...
 
   build:
-    image: golang:1.25
+    image: golang:1.26
     depends_on: tests
     commands:
       - go build ./...
 
   documentation-and-examples:
-    image: golang:1.25
+    image: golang:1.26
     depends_on: tests
     commands:
       - go test ./internal/doccheck
       - go test ./internal/config -run '^TestExamplesLoadAndValidate$'
 
   cross-build:
-    image: golang:1.25
+    image: golang:1.26
     depends_on: [race-tests, static-analysis, build, documentation-and-examples]
     commands:
       - |
         set -eu
         output_dir="$(mktemp -d)"
         trap 'rm -rf "$output_dir"' EXIT
-        for target in linux/amd64 linux/arm64 darwin/amd64 darwin/arm64 windows/amd64 windows/arm64; do
+        for target in linux/amd64 linux/arm64 darwin/arm64; do
           goos="${target%/*}"
           goarch="${target#*/}"
           CGO_ENABLED=0 GOOS="$goos" GOARCH="$goarch" go build -o "$output_dir/narratio-$goos-$goarch" ./cmd/narratio

docs/cli.md

@@ -20,7 +20,7 @@ Top-level commands:
 - `publish <session_id>`: force-run publish.
 - `clean <session_id>` or `clean --all`: remove local work/spool state.
 - `session <subcommand>`: session helper commands.
-- `config <subcommand>`: validate, display, source-trace, or compare resolved pipeline configuration.
+- `config <subcommand>`: validate, display, source-trace, catalog, or compare resolved pipeline configuration.
 
 Session subcommands:
 
@@ -45,6 +45,7 @@ Most session-aware commands accept:
 - `--session-id <session_id>`
 - `--previous-session-id <session_id>`
 - `--profile <name>`
+- `--secrets-dir <directory>`
 
 Rules:
 
@@ -53,10 +54,13 @@ Rules:
 - if both positional `<session_id>` and `--session-id` are provided, values must match.
 - `--previous-session-id` is a strict expectation: the selected session file
   must contain the same `previous_session_id`.
-- `--profile` selects a declared pipeline profile. It may be supplied once;
+- `--profile` selects `profiles/<name>.yml` relative to the pipeline root. It
-  an explicit empty or unknown value fails configuration resolution. When it is
+  may be supplied once; omitted, empty, and the exact name `default` select the
-  omitted, a declared `default_profile` is used. The same selection applies to
+  base configuration, while an unknown named profile fails. The same selection
-  all common-flag commands, including `regenerate-artifacts`.
+  applies to all common-flag commands, including `regenerate-artifacts`.
+- `--secrets-dir` selects ConfigKit credential-file bootstrap explicitly. A
+  relative value uses caller CWD. Without the flag, a present `secrets/`
+  directory beside the resolved root pipeline is used; an absent one is skipped.
 - `clean --all` cannot be combined with campaign/session selectors.
 - notification delivery is currently limited to the configured `noop` mode; see
   the [configuration reference](./config.md#notifications).
@@ -78,19 +82,24 @@ Commands with additional positionals keep their command-specific order:
 
 ## Command Reference
 
-### `config validate`, `config show`, `config sources`, and `config diff`
+### `config validate`, `config show`, `config sources`, `config artifacts`, and `config diff`
 
 ```bash
-narratio config validate [--config <pipeline.yml>] [--campaign <id> | --campaign-file <campaign.yml>] [--profile <name>]
+narratio config validate [--config <pipeline.yml>] [--campaign <id> | --campaign-file <campaign.yml>] [--profile <name>] [--secrets-dir <directory>]
-narratio config show [--config <pipeline.yml>] [--campaign <id> | --campaign-file <campaign.yml>] [--profile <name>]
+narratio config show [--config <pipeline.yml>] [--campaign <id> | --campaign-file <campaign.yml>] [--profile <name>] [--secrets-dir <directory>]
-narratio config sources [--config <pipeline.yml>] [--campaign <id> | --campaign-file <campaign.yml>] [--profile <name>]
+narratio config sources [--config <pipeline.yml>] [--campaign <id> | --campaign-file <campaign.yml>] [--profile <name>] [--secrets-dir <directory>]
-narratio config diff <left-profile> <right-profile> [--config <pipeline.yml>] [--campaign <id> | --campaign-file <campaign.yml>]
+narratio config artifacts [--config <pipeline.yml>] [--campaign <id> | --campaign-file <campaign.yml>] [--profile <name>] [--secrets-dir <directory>]
+narratio config diff <left-profile> <right-profile> [--config <pipeline.yml>] [--campaign <id> | --campaign-file <campaign.yml>] [--secrets-dir <directory>]
 ```
 
 These commands resolve the selected profile, defaults, ordinary paths, and—if
 a campaign is selected—the campaign-owned party. They neither discover or load
-a session nor create a workspace, manifest, run, lock, adapter, remote
+a session nor create a workspace, manifest, run, lock, adapter, or remote
-connection, or credential environment.
+connection. Like every pipeline-loading command, they bootstrap the selected or
+conventional secrets directory before resolving configuration; `config show`
+never includes bootstrap reports or secret values in its YAML output.
+Typed credential references, including `notarius.llm_api_key`, must therefore
+resolve even though these inspection commands never construct an adapter.
 
 Campaign selection is optional for a pipeline without party-driven artifact
 families. A pipeline with `scriptorium.artifact_families` needs a selected or
@@ -101,20 +110,31 @@ Session, range, force, and artifact-execution flags are not accepted.
 `config validate` writes a concise root-path, selected-profile (or `none`), and
 effective-digest summary after successful complete validation. `config show`
 writes one deterministic, secret-free YAML document containing defaulted and
-expanded concrete configuration. It omits composition declarations, artifact
+expanded concrete configuration. It omits ConfigKit directives, artifact
 family declarations, and runtime provenance.
 
 `config sources` reports the same fully validated resolution without printing
 effective values. Its header identifies the root, ordered imports, selected
-profile and overlay, selected campaign, party mode/source, and digest. The
+profile and overlay, selected campaign, party source, and digest. The
 remaining tab-separated records are sorted as `path`, `role`, and `source`.
 Roles distinguish root, import, profile, centralized default, campaign, party,
-legacy-player, and generated family ownership. A generated party member has
+and generated family ownership. A generated party member has
 one family record and one party record at the same logical path. The output
 never reads or prints secret values.
 
-`config diff` resolves both supplied profile names from one parsed root source
+`config artifacts` prints the concrete catalog used by analysis, ordered by
-set and compares their fully resolved, secret-free effective mappings. It does
+artifact key. Its tab-separated header is `key`, `source`, `definitions`,
+`enabled`, `output`, `dependencies`, `family`, `character`, `family_source`, and
+`party_source`. Definitions and effective dependencies are sorted comma-separated
+lists, and control characters in fields are escaped. Family-member rows identify
+their family, character, defining family file, and party file. An empty catalog
+prints only the header. Ordinary artifacts can be inspected without a campaign;
+party-driven families retain the campaign requirement. See
+[Scriptorium Artifact Entries](./config.md#scriptorium-artifact-entries) for the
+configuration layout.
+
+`config diff` resolves both supplied profile names independently and compares
+their fully resolved, secret-free effective mappings. It does
 not accept `--profile`; the two positional names must be distinct, declared
 profiles. When party-driven families are present, both profiles must resolve to
 the same selected campaign and party. Use `--campaign-file` if profile-specific
@@ -153,17 +173,22 @@ Behavior:
 
 - evaluates one inclusive contiguous range of the canonical stage order;
 - defaults an omitted `--from` to `prepare` and an omitted `--through` to
-  `notify`, so omitting both retains full-pipeline behavior;
+  `cleanup`, so omitting both retains full-pipeline behavior;
 - rejects unknown endpoints and a `--from` endpoint after `--through`;
 - runs `render` before `extract`; an omitted or disabled Notarius
   configuration records an explicit `notarius_disabled` self-skip;
+- resolves every selected typed credential during the initial configuration
+  load, including a Notarius credential when extract is disabled, outside the
+  selected range, or later skipped as resumable;
 - skips already-succeeded stages unless `--force` is set or a stage-specific
   resume check finds its durable result obsolete;
 - applies `--force` only to stages in the selected range;
 - rejects repeated `--from`, `--through`, or `--force` options, including
   `--name=value` spellings;
 - continues interrupted or partially completed sessions by running non-succeeded stages;
-- writes session and run manifests.
+- writes the session manifest as the sole durable progress ledger. Each
+  invocation also creates its run-local directory and returns an immediate
+  command summary; neither creates another durable progress record.
 - reports the resolved profile (or `none`) and effective configuration digest.
 
 When `--artifacts` is present, the selected range must contain `analyze` or
@@ -187,6 +212,10 @@ artifacts and any prerequisites required by those targets; without
 never run. Common session/configuration options and repeatable artifact values
 pass through unchanged.
 
+The forced extraction rebuilds the complete configured Notarius bundle,
+disables Notarius checkpoint loading, and requests chunk-plan refresh.
+`--artifacts` narrows only the analysis work; it never selects Notarius lanes.
+
 Because the expansion owns `--force`, `--from`, and `--through`, callers cannot
 supply those options. The shared `run` parser reports them as duplicate
 singleton flags. The alias has no private execution options or behavior, and
@@ -211,6 +240,7 @@ Valid stage names:
 - `analyze`
 - `publish`
 - `notify`
+- `cleanup`
 
 Rules:
 
@@ -228,6 +258,10 @@ Equivalent to:
 narratio run-stage analyze <session_id> --force [...common config flags]
 ```
 
+This selects only `analyze`: it does not invoke Notarius, replace extraction,
+or remove existing extract outputs. `--artifacts` narrows the forced analysis
+targets and their required analysis prerequisites.
+
 ### `publish`
 
 ```bash
@@ -240,6 +274,10 @@ Equivalent to:
 narratio run-stage publish <session_id> --force [...common config flags]
 ```
 
+This command selects only `publish`; it does not invoke cleanup. Use a normal
+full run to include configured cleanup, or `run-stage cleanup` to retry cleanup
+from valid publication evidence.
+
 ### `clean`
 
 ```bash
@@ -253,8 +291,8 @@ Behavior:
 - `--all` removes all local session work and spool state;
 - cache remains unless `--clear-cache` is provided.
 
-See [Operations: Cleanup](./operations.md#cleanup) for deletion scope and
+See [Operations: Cleanup](./operations.md#cleanup) for manual and pipeline-stage
-post-publish cleanup behavior.
+deletion scope.
 
 ### `session plan`
 
@@ -279,7 +317,8 @@ execution; the preview does not predict output identities.
 narratio session validate <session_id> [...common config flags]
 ```
 
-Read-only preflight checks for config validity, required inputs, audio mode, previous-session requirements, publish outputs, and effective locks.
+Read-only preflight checks for config validity, required inputs, audio mode,
+previous-session requirements, publish outputs, and independent remote locks.
 
 ### `session status`
 
@@ -288,7 +327,12 @@ narratio session status <session_id> [...common config flags]
 ```
 
 Prints local manifest state and, when storage is available, status for the
-pointer-selected remote commit and its declared published outputs.
+pointer-selected remote commit and its declared published outputs. Current
+publish metadata reports `published`, `carried_locked`, `locked_unpublished`,
+`optional_missing`, and `unselected`; carried output lines include the run
+derived from their exact declared key, the key itself, and its declared
+destination. A separately displayed `dest` is the destination requested by the
+current configuration and may differ while an older declaration is locked.
 
 ### `session init`
 
@@ -307,6 +351,7 @@ Options:
 
 - `--config <pipeline.yml>`
 - `--campaign <id>` or `--campaign-file <campaign.yml>`
+- `--secrets-dir <directory>`
 - `--previous-session-id <id>`
 - `--date <YYYY-MM-DD>`
 - `--title <text>`
@@ -346,7 +391,8 @@ narratio session artifacts <session_id> [--remote] [...common config flags]
 Lists effective built-in, configured Scriptorium, and configured extraction
 sources; reports planned, available, unavailable, and published state without
 reading payload bodies; and includes publish rules, lock state, and optional
-remote published-state availability.
+remote published-state availability. Remote inspection uses exact commit
+declarations and displays the five current publish outcomes when recorded.
 
 ### `session locks`
 
@@ -358,12 +404,14 @@ narratio session locks remove <session_id> <source> [...common config flags]
 
 Behavior:
 
-- list mode reports the effective merge of static and remote locks;
+- list mode validates and reports independent remote locks sorted by source;
-- add/remove mutate only remote locks;
+- add creates one lock and reports an existing lock unless `--force` is used;
-- static locks from pipeline config cannot be removed by CLI commands.
+- `--force` performs one generation-matched replacement and reports a
+  concurrent change for operator retry; and
+- remove is idempotent and reports when the source was not locked.
 
 See [Operations: Publish Locks](./operations.md#publish-locks) for lock storage
-and precedence.
+and publication behavior.
 
 ## `--artifacts` Selection Rules
 

docs/config.md

@@ -44,13 +44,16 @@ the durable copied session input.
 
 ### Read-only effective pipeline inspection
 
-`narratio config validate`, `narratio config show`, and `narratio config
+`narratio config validate`, `narratio config show`, `narratio config sources`,
-sources` use the same `--config`, `--campaign`, `--campaign-file`, and
+and `narratio config artifacts` use the same `--config`, `--campaign`,
+`--campaign-file`, and
 `--profile` selection rules as pipeline commands, but do not select, discover,
-or load a session. They do not read credential values or create runtime state.
+or load a session. They resolve configured typed credential references as part
+of the ordinary ConfigKit load but expose only their environment-variable names;
+they do not create runtime state.
 `narratio config diff <left-profile> <right-profile>` uses the same pipeline and
-campaign selectors, resolves each named profile independently from one parsed
+campaign selectors, resolves each named profile with an independent ConfigKit
-root source set, and does not accept a separate `--profile` flag.
+load, and does not accept a separate `--profile` flag.
 
 Campaign selection is optional only when the resolved pipeline has no
 `scriptorium.artifact_families`. When families are declared, Narratio selects a
@@ -59,11 +62,12 @@ then parses the campaign-owned party and expands concrete artifacts and any
 family publish rules before validation. `config validate` prints the resulting
 root, profile, and effective digest. `config show` emits the normalized
 effective pipeline YAML, with defaults and concrete expansion included but
-composition and family declarations omitted. `config sources` prints a stable
+ConfigKit directives and family declarations omitted. This review projection is
+not promised to be a reloadable authoring document. `config sources` prints a stable
 source projection instead of effective values: root/import/profile/default
 ownership plus campaign/party and generated-family records. Canonical derived
-players trace to the party; a legacy configured players file is explicitly
+players trace to the party. `config artifacts` lists the concrete artifact map,
-marked as a legacy player source. The [CLI reference](cli.md#config-validate-config-show-config-sources-and-config-diff)
+effective dependencies, and defining sources. The [CLI reference](cli.md#config-validate-config-show-config-sources-config-artifacts-and-config-diff)
 owns command syntax and output conventions.
 
 `config diff` compares normalized field values rather than YAML text or source
@@ -77,14 +81,15 @@ difference.
 ### Migrating to the maintained bundle
 
 Use the [production/testing bundle](../examples/production-testing/pipeline.yml)
-as the complete copyable migration reference. Split stable pipeline settings
+as the complete copyable migration reference. Keep production settings in the
-into explicit additive imports, place production/testing differences in one
+base, split them into explicit additive imports, and place testing differences
-selected overlay, and retain a production `default_profile`. Convert campaign
+in the fixed `profiles/testing.yml` overlay. Put ordinary artifacts and families
-rosters to [canonical party input](integrations/party.md), remove a separate
+in the recommended one-definition files shown by that bundle. Convert campaign
-`players_file`, then express character work as families. Inspect the result
+rosters to [canonical party input](integrations/party.md), then express character
-with `config validate`, `config show`, and `config sources`; use `config diff`
+work as families. Inspect the result with `config validate`, `config show`,
-to review profiles before running a session. Unversioned parties and their
+`config sources`, and `config artifacts`; use `config diff default testing` to
-`players_file` remain a clearly bounded legacy compatibility path.
+review the overlay before running a session. Unversioned party documents are
+rejected.
 
 ### Identity segments
 
@@ -97,51 +102,51 @@ remote state with an unsafe legacy identity must be migrated before use.
 
 ## Validation and Merge Rules
 
-- YAML decode is strict (`KnownFields(true)`) and accepts exactly one document:
+- ConfigKit performs strict typed YAML decoding and accepts exactly one document;
-  unknown fields or trailing documents fail load.
+  unknown fields, duplicate keys, and trailing documents fail load.
 - A pipeline file may explicitly import additive YAML fragments through the
-  root-only `composition.imports` list. Imported files contribute fields to one
+  root-only `configkit.imports` list. Imported files contribute fields to one
   logical pipeline document; they do not override fields supplied by the root
   or another import.
-- A root pipeline may declare named profiles. Exactly one profile is selected
+- An omitted profile, an empty profile option, and the exact name `default`
-  by an option-aware caller or by `composition.default_profile`; a caller's
+  select the base configuration. Any other name selects
-  explicit selection takes precedence. Declaring profiles without either form
+  `profiles/<name>.yml` relative to the pipeline root.
-  of selection is an error.
+- Campaign, session, and canonical-party documents are strict single-source
+  ConfigKit schemas. They do not enable imports, profiles, or `configkit`
+  directives.
 - Configured timeout and retry-delay durations must be positive. An omitted
   artifact timeout continues to inherit its configured Scriptorium timeout.
 - Session files must be concrete; unresolved `{{ ... }}` placeholders fail load.
+- Session and party inputs inspected before typed loading are limited to 8 MiB.
 - Pipeline defaults are applied before validation.
+- Defaults apply only to absent fields. Explicit empty strings, zero values,
+  `false`, and `null` retain ConfigKit's typed meaning and are then checked by
+  Narratio domain validation; they are not silently replaced by defaults.
+  Explicit null optional stage sections stay absent, while required sections
+  fail validation.
 - Campaign and session identities must agree.
 - Required stable files (`speakers_file`, `autocorrect_file`, `glossary_file`,
-  `party_file`) and the optional `spell_catalog_file` resolve from session
+  `party_file`) are campaign-owned; the first three may be overridden by the
-  overrides when provided, otherwise from campaign defaults. An empty or
+  session. The optional `spell_catalog_file` also inherits a campaign value
-  omitted session spell-catalog value inherits the campaign value.
+  when omitted by the session.
-- `party_file` is classified when pipeline and campaign configuration are
+- `party_file` is parsed when pipeline and campaign configuration are combined.
-  combined. A versioned [canonical party](integrations/party.md) is
+  It must be a versioned [canonical party](integrations/party.md), is
-  campaign-owned, derives the players input internally, and forbids both a
+  campaign-owned and derives the players input internally. `players_file` and
-  separate `players_file` and a session `party_file` override. An unversioned
+  session `party_file` are unsupported configuration keys.
-  party remains a bounded legacy input and requires `players_file`; its normal
-  campaign/session overrides continue to apply.
 - Exactly one audio mode must be configured in session input:
   - local (`audio_dir` or `audio_files`), or
   - S3 (`audio_s3.prefix`).
 
-### Pipeline composition
+### Pipeline imports and profiles
 
 Large pipeline configurations may be split into explicitly named fragments and
-may declare one overlay per selectable profile:
+may use fixed-path profile overlays:
 
 ```yaml
-composition:
+configkit:
   imports:
-    - config/storage.yml
+    - path: config/storage.yml
-    - config/integrations.yaml
+    - path: config/integrations.yaml
-  default_profile: production
-  profiles:
-    production:
-      overlay: profiles/production.yml
-    testing:
-      overlay: profiles/testing.yml
 
 campaigns:
   root: /usr/local/share/narratio/campaigns
@@ -155,36 +160,29 @@ Imports are resolved relative to the directory containing the root pipeline
 file and are loaded in declaration order. Narratio does not scan directories or
 infer fragments. Each import must be a confined regular `.yml` or `.yaml` file:
 absolute paths, traversal, symlinks, directories, duplicate files, and an
-import of the root pipeline itself are rejected. Only the root pipeline may
+import of the root pipeline itself are rejected. Imported files cannot declare
-contain `composition`; nested composition is rejected.
+another `configkit` control section.
 
-Composition is additive. A map may be extended by multiple files when every
+Base merging is additive. A map may be extended by multiple files when every
 leaf is distinct, but a scalar, list, or map/list/scalar kind cannot be claimed
 more than once, even when the repeated values are identical. Conflict errors
-name the full field path and every source that claimed it. The assembled YAML is
+name the full field path and every source that claimed it. Narratio additionally
-then decoded against the normal strict pipeline schema and defaults are applied
+rejects an artifact or artifact-family key declared by more than one base file,
-once.
+even when those files contribute disjoint children.
 
-Profile names are case-sensitive, non-empty, trimmed, and cannot contain
+Profile names are case-sensitive and use ConfigKit's safe fixed-path naming
-control characters. If `profiles` is present, it must contain at least one
+rules. Only the selected overlay is opened; a malformed unselected overlay does
-entry and every entry must contain only an `overlay` path. An explicit profile
+not affect a base or different-profile load. Maps merge recursively, overlay
-selection overrides `default_profile`; unknown and explicitly empty selections
+scalars replace base scalars, and overlay lists replace base lists completely.
-fail. Narratio never selects the first profile implicitly and does not read a
+Explicit `false`, zero, empty strings, empty containers, and null values remain
-profile selection from the environment.
+explicit inputs and are not refilled by defaults. Profiles cannot inherit,
+stack, import files, or declare another ConfigKit control section.
 
-Every declared overlay is resolved relative to the root pipeline directory and
+During ConfigKit loading, Narratio applies fresh typed defaults, recorded
-must satisfy the same confined regular-YAML-file rules as an import. Narratio
+entry-level defaults and normalization, read-only file-local validation, and
-parses every declared overlay even when it is not selected, then applies only
+ordinary path resolution before computing a deterministic effective
-the selected one. Maps merge recursively, overlay scalars replace base scalars,
-and overlay lists replace base lists completely. Explicit `false`, zero, empty
-lists, and empty maps remain meaningful. YAML null cannot delete a value, and
-kind changes are rejected. Profiles cannot inherit from or stack with other
-profiles, and overlays cannot import files or declare profiles.
-
-After composition, Narratio strictly decodes the result, applies centralized
-defaults once, resolves ordinary paths, and computes a deterministic effective
 configuration digest. The digest represents the normalized, secret-free
-runtime pipeline mapping; it excludes composition declarations, source
+runtime pipeline mapping; it excludes ConfigKit directives, source
 provenance, profile identity, and raw environment secret values. Equivalent
 effective mappings therefore have the same digest regardless of how fields are
 split among the root and imports.
@@ -213,7 +211,6 @@ inputs:
   speakers_file: ./speakers.yml
   autocorrect_file: ./autocorrect.yml
   glossary_file: ./glossary.yml
-  players_file: ./players.yml
   party_file: ./party.yml
 ```
 
@@ -228,15 +225,31 @@ inputs:
 ## Secrets Handling
 
 - Do not place raw secrets in YAML.
-- Use env var names in config (for example `pipeline.audita.llm_api_key_env`).
+- Credential fields contain environment-variable names, represented internally
-- Optionally load credential files from `pipeline.secrets.env_dir`. Each valid
+  as ConfigKit typed secret references (for example
-  environment-variable filename supplies one value; trailing CR/LF is removed.
+  `pipeline.notarius.llm_api_key: NOTARIUS_LLM_API_KEY`). A configured reference
+  must resolve to a nonempty environment value during load.
+- Every surviving reference is resolved during ordinary pipeline loading,
+  regardless of whether its integration is disabled, outside the selected
+  stage range, or eligible for resume. Omit the field or set it to `null` when
+  that deployment should not require the credential.
+- Every command that loads a pipeline accepts `--secrets-dir <path>`. When the
+  flag is omitted, Narratio uses a present `secrets/` directory beside the
+  resolved root `pipeline.yml`; an absent conventional directory is ignored.
+  Explicit relative paths are resolved from the caller's working directory,
+  and a missing or empty explicit selection is an error.
+- Each immediate regular file whose name is a valid environment-variable name
+  supplies one value. ConfigKit removes trailing CR/LF, permits a symlink as the
+  selected directory, rejects non-hidden symlink entries, and limits input to
+  1 MiB per file, 8 MiB total, and 1024 entries. Dot-prefixed entries (including
+  files, directories, and symlinks) are ignored and omitted from bootstrap
+  reports, but still count toward the directory entry limit.
 - An existing process environment value takes precedence over a credential file.
-- Credential directories and files must not be symlinks and must be regular,
+- On POSIX, operators must provision the directory with no group/other access
-  bounded files (at most 8 KiB per value). On POSIX, provision the directory
+  (normally `0700`) and files with no group/other access (normally `0600`);
-  with no group/other access (normally `0700`) and files with no group/other
+  ConfigKit validates type and bounds but does not enforce permission modes.
-  access (normally `0600`).
+- Bootstrap happens once per command, before any configuration load or adapter
-- Commands that need storage/auth load filesystem secrets before constructing adapters.
+  construction. Safe reports contain filenames, paths, and statuses, never values.
 
 ## Publish Configuration Summary
 
@@ -259,9 +272,6 @@ publish:
     - source: narratio.artifact.session_recap
       dest: artifacts/session_recap.md
       required: true
-  locks:
-    - source: narratio.artifact.session_recap
-      reason: manual post-publish edits
 ```
 
 Rules:
@@ -272,7 +282,8 @@ Rules:
   a rule names that source; the Notarius index and complete bundle are not
   publish sources.
 - `outputs[].required` defaults to `true`.
-- static locks (`pipeline.publish.locks`) merge with remote locks (`{session_prefix}/locks.yml`), with static locks taking precedence on duplicates.
+- publish locks are independent remote session objects managed with
+  `narratio session locks`; they are not pipeline configuration fields.
 
 ## Full Schema
 
@@ -280,24 +291,21 @@ Rules:
 
 | Field | Type | Required | Default / Rule |
 | --- | --- | --- | --- |
-| `composition.imports[]` | list of strings | No | explicit additive pipeline fragments relative to the root pipeline directory; `.yml` or `.yaml` regular files only |
+| `configkit.imports[].path` | string | No | explicit additive pipeline fragment relative to the root pipeline directory; `.yml` or `.yaml` regular files only |
-| `composition.default_profile` | string | Conditional | selected when profiles exist and no caller explicitly selects one; must name a declared profile |
-| `composition.profiles.<name>.overlay` | string | Conditional | required for every declared profile; one confined `.yml` or `.yaml` overlay relative to the root pipeline directory |
 | `pipeline.workspace.root` | string | No | `/var/lib/narratio` |
-| `pipeline.workspace.cleanup_after_publish` | bool | No | `false` |
+| `pipeline.workspace.cleanup_after_publish` | bool | No | `false`; selects the published run's workspace target for final cleanup and requires effective `pipeline.publish.enabled=true` and `pipeline.publish.upload_run=true` |
 | `pipeline.campaigns.root` | string | No | `/usr/local/share/narratio/campaigns` |
 | `pipeline.campaigns.default_campaign_id` | string | No | empty |
-| `pipeline.secrets.env_dir` | string | No | empty |
 | `pipeline.storage.backend` | string | No | `local`; supported values are `local` and `s3` (case-insensitive) |
 | `pipeline.storage.s3.bucket` | string | Conditional | required when backend is `s3` and S3 session-audio or publish upload is enabled |
 | `pipeline.storage.s3.root_prefix` | string | No | `dnd` |
 | `pipeline.storage.s3.region` | string | No | empty |
 | `pipeline.storage.s3.endpoint` | string | No | empty |
 | `pipeline.storage.s3.force_path_style` | bool | No | `false` |
-| `pipeline.storage.s3.access_key_id_env` | string | No | `OBJECT_STORAGE_KEY_ID` |
+| `pipeline.storage.s3.access_key_id` | secret reference | Conditional | unset; must be paired with `secret_access_key` |
-| `pipeline.storage.s3.secret_access_key_env` | string | No | `OBJECT_STORAGE_KEY` |
+| `pipeline.storage.s3.secret_access_key` | secret reference | Conditional | unset; must be paired with `access_key_id` |
 | `pipeline.spool.root` | string | No | `/var/spool/narratio` |
-| `pipeline.spool.delete_audio_after_publish` | bool | No | `false` |
+| `pipeline.spool.delete_audio_after_publish` | bool | No | `false`; selects the published run's spool-audio target for final cleanup and requires effective `pipeline.publish.enabled=true` and `pipeline.publish.upload_run=true` |
 | `pipeline.cache.root` | string | No | `/var/cache/narratio` |
 | `pipeline.cache.s3_audio` | bool | No | `true` |
 | `pipeline.publish.enabled` | bool | No | `true` |
@@ -306,9 +314,6 @@ Rules:
 | `pipeline.publish.outputs[].source` | string | Yes (per rule) | must reference built-in or configured artifact source |
 | `pipeline.publish.outputs[].dest` | string | Conditional | derived if omitted and source supports derivation |
 | `pipeline.publish.outputs[].required` | bool | No | `true` |
-| `pipeline.publish.locks[]` | list | No | empty |
-| `pipeline.publish.locks[].source` | string | Yes (per lock) | must reference supported publish source |
-| `pipeline.publish.locks[].reason` | string | No | empty |
 | `pipeline.whisperx.transcribe_url` | string | Yes | absolute `http` or `https` URL |
 | `pipeline.whisperx.language` | string | No | `en` |
 | `pipeline.whisperx.timeout` | duration | No | `30m` |
@@ -326,7 +331,7 @@ Rules:
 | `pipeline.seriatim.env.filler_max_duration` | float | No | unset |
 | `pipeline.audita.binary` | string | No | `audita` |
 | `pipeline.audita.timeout` | duration | No | `3h` |
-| `pipeline.audita.llm_api_key_env` | string | No | empty |
+| `pipeline.audita.llm_api_key` | secret reference | No | unset |
 | `pipeline.audita.modules[]` | list[string] | No | empty |
 | `pipeline.audita.base_url` | string | No | empty |
 | `pipeline.audita.model` | string | No | empty |
@@ -354,10 +359,14 @@ Rules:
 | `pipeline.trim.seriatim.report` | bool | No | `false` |
 | `pipeline.notarius.enabled` | bool | No | `false` |
 | `pipeline.notarius.binary` | string | No | `notarius` |
+| `pipeline.notarius.llm_api_key` | secret reference | No | unset; exact environment-variable name forwarded to Notarius |
+| `pipeline.notarius.config_profile` | string | No | unset; exact Notarius ConfigKit profile selector |
 | `pipeline.notarius.config_path` | string | Conditional | required when enabled; relative paths resolve from the pipeline file directory |
 | `pipeline.notarius.pipeline_id` | string | Conditional | required when enabled |
 | `pipeline.notarius.timeout` | duration | No | `3h`; must be positive |
 | `pipeline.notarius.working_directory` | string | No | directory containing resolved `config_path`; relative paths resolve from the pipeline file directory |
+| `pipeline.notarius.resume` | bool | No | `false`; request compatible Notarius checkpoint reuse during ordinary executed extraction |
+| `pipeline.notarius.chunk_cache` | string | No | empty; when set, exactly `auto`, `bypass`, or `refresh` |
 | `pipeline.notarius.references` | map[string]string | No | empty; maps normalized Notarius selectors to supported prepared Narratio source IDs; maximum 256 entries |
 | `pipeline.notarius.outputs` | map | Conditional | at least one entry when enabled |
 | `pipeline.render.enabled` | bool | No | `true` |
@@ -374,6 +383,73 @@ Rules:
 | `pipeline.scriptorium.artifact_families` | map | No | empty; expands one ordinary artifact per canonical party character |
 | `pipeline.notification.mode` | string | No | `noop`; the only supported notification mode until a provider is implemented |
 
+### Notarius Configuration Profile
+
+`pipeline.notarius.config_profile` optionally selects a Notarius ConfigKit
+configuration profile when extraction executes:
+
+```yaml
+notarius:
+  config_profile: downstream-economy
+```
+
+When the field is omitted or `null`, Narratio supplies no downstream profile
+selector and preserves Notarius's existing configuration-selection behavior.
+When present, Narratio forwards the exact string to Notarius, including an
+empty string or `default`, which explicitly select Notarius's composed base
+configuration. Notarius owns selector interpretation and validation; Narratio
+does not restrict profile names, resolve profile files, or rewrite the value.
+
+This selection is independent of Narratio's own `--profile` selection. A
+Narratio overlay may explicitly select a differently named Notarius overlay,
+but Narratio never forwards its own profile name automatically. It is also
+independent of `pipeline.notarius.llm_api_key`: an overlay may change either or
+both fields according to the deployed Notarius configuration. See the
+[Notarius integration contract](./integrations/notarius.md) for invocation and
+failure behavior.
+
+### Notarius Credential
+
+`pipeline.notarius.llm_api_key` optionally names the one environment variable
+to forward to the Notarius child process:
+
+```yaml
+notarius:
+  llm_api_key: OPENROUTER_API_KEY
+```
+
+The field contains the variable name, not a credential value. That name must
+match `api_key_env` in the Notarius/PromptKit profile selected by the deployed
+Notarius pipeline; Narratio does not infer or remap provider names. Omission or
+`null` forwards nothing. A non-null selected reference must resolve at pipeline
+load time through the process environment or existing secrets-directory
+bootstrap, even when Notarius is disabled, extract is outside the requested
+range, or a completed extraction can be reused.
+
+The Notarius child receives only Narratio's minimal execution environment plus
+this exact selected name and loaded value. Other credentials are not forwarded.
+Changing the value, name, presence, or absence does not invalidate extract's
+semantic resume evidence. The public name remains part of safe effective
+inspection, so renaming it may appear in `config diff` and change the effective
+configuration digest; rotating only its value does neither.
+
+See the [Notarius integration contract](./integrations/notarius.md) for the
+delivery and redaction boundary.
+
+### Notarius Cache Controls
+
+`pipeline.notarius.resume` defaults to `false`. When true, an executed ordinary
+extraction passes `--resume`; the deployed Notarius configuration must already
+enable checkpoint recording and should select an explicit persistent checkpoint
+root. Narratio does not inspect or generate that configuration.
+
+`pipeline.notarius.chunk_cache` is empty by default, leaving Notarius's
+configured mode in control. A nonempty value must be `auto`, `bypass`, or
+`refresh` and is passed as `--chunk_cache`. Forced extraction takes precedence
+over both fields: it omits `--resume` and passes `--chunk_cache refresh`.
+Neither field changes Narratio's semantic extraction fingerprint, so changing
+them alone does not rerun an otherwise complete durable extraction.
+
 ### Notarius Reference Bindings
 
 `pipeline.notarius.references` maps a Notarius CLI selector to a prepared
@@ -403,8 +479,8 @@ It validates only selector structure and the prepared source vocabulary;
 Notarius owns target-slot declarations and media compatibility.
 
 Before extraction, Narratio resolves every binding from the current prepared
-session manifest and streams it into a verified invocation-local snapshot whose
+session manifest and streams it into a verified, content-addressed session
-absolute path is passed to Notarius. Missing, unsafe, empty,
+snapshot whose absolute path is passed to Notarius. Missing, unsafe, empty,
 changed-during-copy, or checksum-inconsistent prepared evidence fails with
 guidance to force `prepare`. Bindings are sorted by normalized selector and are
 part of extraction fingerprint and resume identity. See the
@@ -439,7 +515,7 @@ For each `pipeline.scriptorium.artifacts.<name>`:
 | Field | Type | Required | Rule |
 | --- | --- | --- | --- |
 | `enabled` | bool | No | `false` if omitted |
-| `depends_on[]` | list[string] | No | must reference configured artifact keys; no self-reference; configured graph must be acyclic |
+| `depends_on[]` | list[string] | No | explicit ordering-only edges; must reference configured artifact keys, may not self-reference, and join inferred input edges in one acyclic graph |
 | `render_debug` | bool | No | per-artifact override |
 | `prompt_id` | string | Conditional | required when artifact is enabled |
 | `profile_id` | string | No | empty |
@@ -453,8 +529,8 @@ Narratio adds `session_id=narratio-session-<session_id>` to every Scriptorium re
 Without `--artifacts`, analyze executes enabled configured artifacts. With an
 explicit `--artifacts` list, the exact named configured artifacts are the
 one-invocation targets even if their `enabled` values are false. Analyze closes
-those targets over `depends_on`: a current prerequisite is reused, while a
+those targets over the effective dependency graph: a current prerequisite is reused, while a
-stale, missing, failed, or legacy prerequisite is rebuilt before its dependent.
+stale, missing, or failed prerequisite is rebuilt before its dependent.
 Unrelated artifacts are not executed. Named targets and any prerequisite that
 may require rebuilding must therefore have valid executable fields. This
 override affects analyze planning only; publish uses the list only to filter
@@ -471,12 +547,22 @@ For each artifact input `pipeline.scriptorium.artifacts.<name>.inputs.<input_nam
 loading. Use the canonical `source` identifier to select the input; Narratio
 does not provide adapter-specific input passthrough fields.
 
+Each `narratio.artifact.<name>` input source automatically contributes an
+effective dependency on that configured artifact. Narratio unions these inferred
+edges with explicit `depends_on` entries, then trims, deduplicates, and sorts the
+result during final configuration resolution. Authors use `depends_on` only for
+ordering constraints that are not already expressed by an artifact input.
+Use [`narratio config artifacts`](./cli.md#config-validate-config-show-config-sources-config-artifacts-and-config-diff)
+to inspect keys, definition sources, outputs, and effective dependencies without
+loading a session or creating runtime state.
+
 ### Scriptorium Artifact Families
 
 `pipeline.scriptorium.artifact_families` declares a shared artifact template
 for every canonical campaign character. Configuration resolution expands each
 family into ordinary `pipeline.scriptorium.artifacts` entries before analyze
-planning or Scriptorium invocation. A legacy party cannot be used for a family.
+planning or Scriptorium invocation. Party parsing rejects unsupported document
+schemas before family expansion.
 
 For each `pipeline.scriptorium.artifact_families.<name>`:
 
@@ -501,9 +587,10 @@ Their resolved values are strings. A member variable may not reuse a static
 other Scriptorium artifact.
 
 Within a family only, an input source may use
-`narratio.member_artifact.<family>`. The referenced family must be named in
+`narratio.member_artifact.<family>`. Resolution rewrites the source to the
-that family's `member_dependencies`; resolution rewrites the source to the
+corresponding ordinary `narratio.artifact.<family>_<character_id>` source and
-corresponding ordinary `narratio.artifact.<family>_<character_id>` source.
+infers the concrete member dependency. Use `member_dependencies` only for
+additional same-character ordering constraints that are not already inputs.
 This syntax is rejected in explicit artifacts and never reaches runtime stages
 or Scriptorium.
 
@@ -524,7 +611,6 @@ integration.
 | `inputs.speakers_file` | string | Yes | stable input default |
 | `inputs.autocorrect_file` | string | Yes | stable input default |
 | `inputs.glossary_file` | string | Yes | stable input default |
-| `inputs.players_file` | string | Conditional | required only with an unversioned legacy `party_file`; forbidden for a canonical party |
 | `inputs.party_file` | string | Yes | stable campaign party source; relative paths resolve from `campaign.yml` |
 | `inputs.spell_catalog_file` | string | No | optional spell-catalog overlay default; required when a Notarius reference selects `narratio.input.spell_catalog` |
 
@@ -540,8 +626,6 @@ integration.
 | `inputs.speakers_file` | string | No | overrides campaign stable input |
 | `inputs.autocorrect_file` | string | No | overrides campaign stable input |
 | `inputs.glossary_file` | string | No | overrides campaign stable input |
-| `inputs.players_file` | string | No | legacy-party override; forbidden for a canonical party |
-| `inputs.party_file` | string | No | legacy-party override; forbidden for a canonical campaign party |
 | `inputs.spell_catalog_file` | string | No | overrides the optional campaign spell catalog; empty or omitted inherits the campaign value |
 | `inputs.audio_dir` | string | Conditional | local audio mode |
 | `inputs.audio_files[]` | list[string] | Conditional | local audio mode |

docs/development.md

@@ -10,6 +10,11 @@ polished transcripts and generated artifacts. Start with the
 [Architecture](policy/architecture.md) for normative system boundaries, and the
 [Internal Overview](internal/overview.md) for implemented component ownership.
 
+## Go Prerequisite
+
+Maintained development and verification require Go 1.26. Use the official Go
+1.26 toolchain before running repository commands.
+
 ## What To Read
 
 | When working on | Read | Why |
@@ -53,7 +58,7 @@ in the [Release Procedure](release.md).
 
 Woodpecker also runs `go test -race -shuffle=on -count=3 ./...` on its scheduled
 job to expose ordering and repeatability defects. Current runners cross-compile
-for macOS and Windows, but do not provide native macOS or Windows execution.
+the supported Linux amd64 and arm64 targets and the best-effort macOS arm64
-Those cross-builds establish compilation only, not platform-equivalent runtime
+target. The macOS cross-build establishes compilation only, not native runtime
-evidence. Add native checks only when official runner labels and successful
+evidence. Add a native check only when an official runner label and successful
 native-run evidence are available.

docs/integrations/audita.md

@@ -41,7 +41,6 @@ Construction fails for invalid static config values, including:
 
 Run fails for:
 - missing required request paths;
-- missing required credential env var when configured (`llm_api_key_env`);
 - subprocess execution failure;
 - invalid processed transcript JSON (`segments` array required);
 - invalid report JSON when reporting is enabled.
@@ -53,6 +52,9 @@ Failure results still include output/log/config/exit metadata for diagnostics.
 
 ## Deterministic Behavior
 - CLI args are built from runner config + request in a fixed order.
+- A configured typed `llm_api_key` reference is resolved during pipeline load;
+  the captured value is supplied only as the child's `AUDITA_LLM_API_KEY`
+  override and is never re-read from the parent environment.
 - Generated invocation YAML (`audita.generated.v1`) is emitted when requested.
 - Manifest writes are stage-owned; adapter itself is stateless.
 

docs/integrations/notarius.md

@@ -7,14 +7,14 @@ lanes from the final trimmed Seriatim transcript. Narratio owns invocation,
 safe bundle discovery, lane selection, and its own artifact metadata. Notarius
 owns pipeline definitions, lane schemas, the receipt, and bundle formats.
 
-Canonical Notarius v0.6.0 references:
+Canonical Notarius v0.7.0 references:
 
-- [CLI reference](https://gitea.maximumdirect.net/eric/notarius/src/tag/v0.6.0/docs/cli.md)
+- [CLI reference](https://gitea.maximumdirect.net/eric/notarius/src/tag/v0.7.0/docs/cli.md)
-- [Subprocess consumer contract](https://gitea.maximumdirect.net/eric/notarius/src/tag/v0.6.0/docs/consumers/subprocess.md)
+- [Subprocess consumer contract](https://gitea.maximumdirect.net/eric/notarius/src/tag/v0.7.0/docs/consumers/subprocess.md)
-- [D&D pipeline and lane contracts](https://gitea.maximumdirect.net/eric/notarius/src/tag/v0.6.0/docs/consumers/dnd-pipeline.md)
+- [D&D pipeline and lane contracts](https://gitea.maximumdirect.net/eric/notarius/src/tag/v0.7.0/docs/consumers/dnd-pipeline.md)
-- [Run-result receipt](https://gitea.maximumdirect.net/eric/notarius/src/tag/v0.6.0/docs/integrations/run-result.md)
+- [Run-result receipt](https://gitea.maximumdirect.net/eric/notarius/src/tag/v0.7.0/docs/integrations/run-result.md)
-- [JSON output bundle](https://gitea.maximumdirect.net/eric/notarius/src/tag/v0.6.0/docs/integrations/json-output.md)
+- [JSON output bundle](https://gitea.maximumdirect.net/eric/notarius/src/tag/v0.7.0/docs/integrations/json-output.md)
-- [D&D spell-catalog overlay](https://gitea.maximumdirect.net/eric/notarius/src/tag/v0.6.0/docs/integrations/dnd-spell-catalog-overlays.md)
+- [D&D spell-catalog overlay](https://gitea.maximumdirect.net/eric/notarius/src/tag/v0.7.0/docs/integrations/dnd-spell-catalog-overlays.md)
 
 The [complete Narratio example](../../examples/pipeline.full.annotated.yml)
 records the exact current constraints for all ten D&D lanes. Treat the linked
@@ -25,19 +25,38 @@ duplicate the complete schemas.
 
 When `pipeline.notarius.enabled` is true, Narratio resolves the executable,
 configuration path, input path, output directory, and working directory to
-absolute paths. Narratio requires the Notarius v0.6.0 CLI contract when
+absolute paths. Configured references require support for the `--reference`
-references are configured and invokes each binding as a separate argument
+contract introduced in Notarius v0.6.0. Cache controls require support for the
-before `--json`:
+selected `--resume` and `--chunk_cache` flags. Notarius v0.7.0 is the tested
+cache integration baseline, not an unconditional minimum version; Narratio
+does not impose a runtime version gate. Each reference binding is a separate
+argument before `--json`. When `pipeline.notarius.config_profile` is present,
+its exact value is another argument pair before `--json`:
 
 ```text
-notarius run <pipeline_id> --config <config_path> --input <trimmed_json> --output-dir <staging_dir> [--reference <selector>=<verified_snapshot_path>]... --json
+notarius run <pipeline_id> --config <config_path> --input <trimmed_json> --output-dir <staging_dir> [--resume] [--chunk_cache <mode>] [--reference <selector>=<verified_snapshot_path>]... [--config-profile <name>] --json
 ```
 
-Reference paths are absolute invocation-local snapshots streamed from the
+Notarius v0.7.0 is a known supporting version for `--config-profile`; using the
-manifest-verified canonical files prepared inside the current Narratio session
+option requires a compatible Notarius executable, but deployments that omit the
-workspace. Narratio verifies snapshot checksum and size before and after the
+field acquire no additional version requirement. Omitted or null configuration
-subprocess, and passes only configured bindings, ordered lexically by normalized
+adds no flag. Supplied empty, `default`, and named selectors are forwarded
-selector, as direct argument-vector entries without shell interpretation. A CLI
+without normalization so Notarius can apply its own ConfigKit semantics. This
+option selects a Notarius configuration overlay, not a PromptKit profile through
+`--llm-profile`, and Narratio never substitutes its own selected profile name.
+A missing or invalid downstream overlay is an ordinary Notarius subprocess
+failure: Narratio does not retry without the selector or fall back to another
+profile.
+
+Reference paths are absolute stable snapshots streamed from manifest-verified
+canonical files prepared inside the current Narratio session workspace. They
+live at `cache/notarius/references/<sha256>/<prepared-basename>` and retain the
+full lowercase content digest. Narratio verifies snapshot checksum and size
+before and after the subprocess, reconstructs missing or corrupt regular files
+from a still-valid prepared source, and passes only configured bindings,
+ordered lexically by normalized selector, as direct argument-vector entries
+without shell interpretation. Symlinks, directories, other non-regular
+conflicts, and prepared-source mismatches fail rather than being deleted. A CLI
 binding takes precedence over a matching external path in Notarius
 configuration. Narratio never emits `--without-reference`.
 
@@ -75,9 +94,41 @@ It does not pass a Narratio session ID or run `notarius config validate`
 automatically; the configured working directory and Narratio's minimal child
 environment apply to the subprocess.
 
+Ordinary extraction maps `pipeline.notarius.resume: true` to `--resume` and a
+nonempty `pipeline.notarius.chunk_cache` to `--chunk_cache <mode>`. Forced
+extraction always omits `--resume` and supplies `--chunk_cache refresh`.
+Notarius owns checkpoint enablement, recording, identity, and both cache roots;
+Narratio neither reads nor synthesizes its configuration. Use explicit
+persistent roots in Notarius when reuse must span subprocess invocations.
+Changing a reference path/digest or other upstream identity input correctly
+causes a cold miss. Chunk refresh bypasses lookup and publishes the accepted
+replacement for the active source; it does not delete unrelated plans or
+checkpoint state.
+
+When `pipeline.notarius.llm_api_key` is absent or null, Narratio adds no
+credential to that environment. When configured, the field names an environment
+variable whose already-loaded nonempty value is forwarded under that exact
+name. Operators must use the same name as the `api_key_env` in the selected
+Notarius/PromptKit profile. Narratio does not inspect the Notarius configuration
+to infer a provider or destination name, and it does not forward unrelated
+credentials from its parent environment or secrets directory.
+
+The selected name is registered with Narratio's shared streaming redactor.
+Credential bytes written by the child to captured stdout, captured stderr, or
+stderr-derived failure diagnostics are replaced with `<redacted>`, including
+custom names not known to the built-in sensitive-name list. Narratio does not
+serialize the credential into arguments, stage metadata, manifests, or
+generated configuration.
+
+This captured-stream protection is not a filesystem sandbox. Notarius remains
+responsible for keeping credentials out of files it writes directly, including
+bundle artifacts and its own diagnostics. Narratio does not scan or rewrite
+arbitrary downstream files and cannot promise redaction of transformed or
+encoded credentials.
+
 ## Accepted Result
 
-Narratio's supported invocation baseline is Notarius v0.6.0. The accepted
+Narratio's tested invocation baseline is Notarius v0.7.0. The accepted
 receipt remains `notarius.run-result.v2`; reference flags do not change the
 receipt or ten-lane output contract. The receipt
 must identify the configured pipeline, and its `index_file` must be exactly
@@ -102,6 +153,8 @@ version, and, when configured, module key. Missing, duplicate, rejected, or
 incompatible required lanes fail extraction even if Notarius exited zero. A
 configured lane whose v2 validation summary is `rejected` or `incomplete` also
 fails extraction.
+An accepted reused producer may report `producer_attempt_count: 0`; Narratio
+preserves that valid summary rather than treating it as a missing field.
 Unconfigured lanes may remain in the preserved bundle but do not become
 selectable Narratio sources.
 
@@ -130,3 +183,18 @@ Configuration fields and defaults are in [Configuration](../config.md).
 Operator paths, rerun procedures, and bundle retention are in
 [Operations](../operations.md). See [Troubleshooting](../troubleshooting.md)
 for failure recovery.
+
+## Opt-in Integration Test
+
+The production-boundary cache test requires an absolute path to a locally
+built Notarius v0.7.0-compatible executable. It uses only a local
+OpenAI-compatible `httptest` provider and temporary cache/config/output roots:
+
+```sh
+NARRATIO_TEST_NOTARIUS_BINARY=/absolute/path/to/notarius \
+  GOWORK=off go test ./internal/stage \
+  -run '^TestNotariusCacheReuseIntegration$' -count=1 -v
+```
+
+When the variable is unset, the normal test suite skips this test and never
+searches `PATH`, a sibling checkout, the network, or a real provider.

docs/integrations/party.md

@@ -1,8 +1,8 @@
 # Canonical Party Input
 
-`party.yml` is a campaign-owned roster input. Narratio recognizes the
+`party.yml` is a campaign-owned roster input. It must be a single, versioned
-versioned `narratio.party.v1` document below when it resolves a pipeline,
+`narratio.party.v1` document when Narratio resolves a pipeline, campaign, and
-campaign, and session together.
+session together.
 
 ```yaml
 schema_version: narratio.party.v1
@@ -34,19 +34,17 @@ non-empty, trimmed display strings without control characters. Character names
 and aliases must be unique across the full roster under Unicode-aware
 case-insensitive comparison; player names may repeat.
 
-The document has exactly one YAML document and accepts no unknown fields. A
+The document has exactly one YAML document and accepts no unknown fields.
-wrong or malformed `schema_version` is an error.
+`schema_version` is required: a missing, wrong, or malformed value is an error.
 
-## Legacy migration boundary
+## Migration
 
-An unversioned party input remains supported only as opaque legacy reference
+Unversioned party inputs are rejected. Replace each legacy party document with
-material while campaigns migrate. It requires a separate `players_file` and
+the versioned structure above, move each roster entry under `characters`, and
-retains the existing session override behavior. It cannot be mixed with a
+remove separate player and session-party settings: those keys are unsupported.
-canonical party: canonical campaigns must omit `players_file`, and sessions
-must not override their party or players inputs.
 
-Use the canonical document for new campaigns. The configuration rules and
+The configuration rules and source-relative path behavior are defined in the
-source-relative path behavior are defined in the [Configuration Reference](../config.md).
+[Configuration Reference](../config.md).
 
 ## Derived players document
 

docs/integrations/whisperx.md

@@ -17,11 +17,20 @@ Narratio sends an HTTP `POST` to the configured transcription URL using
 The server must return a `2xx` response whose body is valid JSON. Narratio does
 not currently require a more specific response schema at this boundary.
 
-The transcription URL must be an absolute `http` or `https` URL. The audio body
+The transcription URL must be an absolute `http` or `https` URL. Each
-is streamed through a fresh multipart writer for every attempt, so its memory
+application attempt creates one bounded multipart framing and streams the audio
-use is bounded by the transport buffer rather than by the complete audio file.
+from an open file between that framing; it does not buffer the complete audio
+or spool a multipart copy. The file field precedes the language field.
 WhisperX response acquisition is capped at 10 MiB.
 
+The initial request body and any transport-requested replay use the same
+multipart boundary and framing. Every body opens an independent audio handle
+and replay starts at the beginning. Narratio closes all bodies created for the
+attempt on response, failure, timeout, or cancellation, including when a
+transport abandons an earlier body. Genuine source read/close failures remain
+request failures; deliberate early closure of an abandoned original does not
+poison a valid replay.
+
 ## Request And Result Contract
 
 Each adapter request identifies a speaker, a readable audio file, and the
@@ -43,6 +52,11 @@ transcript output.
 - Narratio waits the configured retry delay between attempts and aborts that
   wait when the parent context is canceled.
 
+Transport replay occurs inside one application attempt and does not consume an
+entry from the configured retry budget or increment the reported attempt
+count. The standard HTTP redirect policy remains in effect; body replay can
+also support body-preserving redirects.
+
 ## Validation And Failure Semantics
 
 Client construction rejects a missing or non-HTTP(S) absolute transcription URL,
@@ -55,12 +69,19 @@ failure to install the output causes the transcription to fail. Errors include
 attempt context, and the result retains attempts, final status when available,
 and elapsed duration for diagnostics.
 
+Replay makes a fresh request body available to the HTTP transport; it does not
+provide an exactly-once server guarantee and does not prove recovery from every
+proxy or protocol-level failure. A server may have received work before the
+client learns that a request failed. WhisperX operations should therefore be
+safe to retry for the same input when practical.
+
 ## Determinism And Concurrency
 
 Each audio request has stable multipart field names, and successful bytes are
 installed atomically. The transcribe stage may process speaker files in
-parallel, bounded by the configured concurrency. It records results in stable
+parallel, bounded by the configured concurrency. Each validated speaker is
-speaker order after all work completes; any speaker failure fails the stage.
+checkpointed independently; any speaker failure still fails the aggregate
+stage, while committed earlier speakers remain reusable on an ordinary retry.
 
 ## Related Canonical Docs
 

docs/internal/adapters.md

@@ -35,8 +35,9 @@ Adapters do not own:
 
 ## Default Wiring
 
-`internal/app/runner.go` initializes default adapters when not injected and
+The application composes default adapters per invocation when they were not
-only when the selected execution plan needs them:
+injected through the stage environment, and only when the selected execution
+plan needs them:
 
 - WhisperX HTTP client for `transcribe`.
 - Seriatim subprocess runner for `merge`, `normalize`, `trim`, or `render`.
@@ -49,8 +50,10 @@ only when the selected execution plan needs them:
 Remote publish locks are loaded only for a selected, enabled publish that
 uploads a run. Shared session lifecycle setup still applies to every selected
 range, but an unselected integration is neither initialized nor validated by
-runner composition. Each selected stage retains its own fail-fast configuration
+application-owned composition. Constructor, clock, randomness, remote-lock,
-and input validation.
+and post-publish deletion boundaries are scoped to the invocation; stages still
+receive their collaborators through the stage environment. Each selected stage
+retains its own fail-fast configuration and input validation.
 
 `session plan` is outside production adapter composition. It performs
 resume validation and models selected transitions against cloned manifest
@@ -62,31 +65,44 @@ Notarius is composed only when extraction is enabled; the extract stage owns
 prepared reference resolution, receipt, bundle, and configured-lane policy.
 The adapter validates the ordered selector/absolute-path pairs and is the sole
 owner of serializing them as repeated `--reference` arguments before `--json`.
+It also serializes an effective resume request as `--resume` and a nonempty
+validated chunk mode as `--chunk_cache <mode>`. The extract stage, not the
+adapter, resolves force precedence before constructing the request.
+Its request may also carry one already-resolved opaque ConfigKit secret
+reference. Only the Notarius adapter reads that value, rejects an unresolved or
+empty programmatic reference, and maps the exact configured name/value into the
+shared subprocess override and sensitive-name fields. It never rereads the
+parent environment. A nil reference preserves the previous invocation.
 
-Object-store construction goes through `newCommandObjectStore`, which loads
+Application-owned object-store construction loads configured filesystem
-configured filesystem secrets before adapter initialization.
+secrets before invoking the invocation's storage factory. Production commands
+use the concrete storage adapter constructor; focused tests can supply an
+in-memory backend without changing package state.
 
 ## Failure Semantics
 
 - Constructor errors fail stage execution setup early.
 - Runtime adapter errors propagate to stage code and then manifest failure handling.
 - Subprocess adapters persist stage logs/generated configs through stage-managed paths.
-- Shared subprocess execution starts an owned process group on Linux/macOS or a
+- Shared subprocess execution keeps private lifecycle mechanics separate from
-  kill-on-close job object on Windows. Every terminal path disposes of that
+  Narratio's environment, diagnostics, redaction, and error policy. It starts
-  owned tree before returning. After a natural leader exit, Unix checks for
+  an owned process group on Linux and macOS. Every terminal path disposes of
-  remaining group members and uses bounded graceful then forceful termination;
+  that owned tree before returning. After a natural leader exit, the runner checks
-  Windows closes the job so kill-on-close applies. Cancellation, deadlines, and
+  for remaining group members and uses bounded graceful then forceful termination.
+  Cancellation, deadlines, and
   diagnostic limits use the same terminal disposal path without losing their
   original result classification. Child environments contain only the execution
-  baseline and adapter-specified values; configured credentials are explicit
+  baseline and adapter-specified values; configured credentials are resolved
-  sensitive values. Stdout and stderr are redacted while streaming into separate
+  once by ConfigKit and passed explicitly as sensitive values. Adapters do not
+  re-read their named environment references. Stdout and stderr are redacted while streaming into separate
   8 MiB diagnostic captures; a bounded wait closes a stream retained by a
   departed leader's descendant. Unsupported platforms reject owned command
   execution.
 
 ## Implementation And Tests
 
-- Composition: `internal/app/runner.go`, `internal/app/object_store.go`
+- Composition: `internal/app/runner_composition.go`,
+  `internal/app/runner_dependencies.go`, `internal/app/object_store.go`
 - Shared subprocess mechanics: `internal/adapters/subprocess`
 - Focused adapters: `internal/adapters/{whisperx,seriatim,audita,scriptorium,notarius,storage,notify}`
 - `internal/adapters/whisperx/http_test.go`

docs/internal/artifacts.md

@@ -20,8 +20,8 @@ The internal registry recognizes these stable built-in source IDs:
 - `narratio.transcript.final_trimmed_markdown`
 - `narratio.bounds.session`
 
-Registry entries bind each ID to its producer, output kind, canonical fallback,
+Registry entries bind each ID to its producer, output kind, and content
-and content validator. The focused stage documents own their input/output flow;
+validator. The focused stage documents own their input/output flow;
 [Configuration](../config.md) owns where operators may select these IDs.
 
 ## Configured, Extraction, And Previous-Session Sources
@@ -58,10 +58,12 @@ Configured definitions are always registered. Without an explicit selection,
 the effective analyze set contains enabled definitions. With `--artifacts`, the
 exact named configured definitions become the effective set for that invocation,
 regardless of their `enabled` value. The effective-set resolver itself does not
-expand dependencies; the analyze work planner closes those targets over their
+expand dependencies. Configuration has already materialized the sorted union of
-configured prerequisite graph. Availability is separate from executability.
+explicit ordering edges and configured-artifact input edges, and the analyze
-Configuration may normalize a family selection into its concrete generated
+work planner closes targets over that one concrete prerequisite graph.
-members before this resolver runs. The effective set retains optional family
+Availability is separate from executability. Configuration may normalize a
+family selection into its concrete generated members before this resolver runs.
+The effective set retains optional family
 and character origin metadata, but its keys, catalog sources, and runtime
 lookups remain concrete configured-artifact identities.
 Family publish policies are likewise expanded into ordinary configured-source
@@ -69,8 +71,9 @@ publish rules during configuration resolution.
 Configured outputs, including non-executable prerequisites, become available
 only when the versioned analyze state identifies a current result whose source,
 contract, canonical configured path, size, and checksum match a confined
-no-follow regular file. An incidental canonical file and a legacy aggregate
+no-follow regular file. An incidental canonical file is unavailable, and a
-analyze output are unavailable.
+manifest with an unsupported aggregate-only representation is rejected before
+catalog construction.
 Extraction entries are registered from configuration and become available only
 after compatible extraction evidence is hydrated.
 
@@ -91,8 +94,13 @@ Current provenance values:
 
 Built-ins:
 
-1. manifest producer outputs (when present)
+- resolve only from the current manifest's succeeded registered producer;
-2. canonical session-path fallback
+- require exactly one output with the registered source ID and output kind, a
+  safe session-rooted path, producer run ID, canonical checksum, and valid
+  content; and
+- verify the registered checksum against the durable bytes. Incidental files,
+  canonical filenames, duplicate outputs, failed/stale producer records, and
+  malformed evidence are unavailable.
 
 Configured sources (`narratio.artifact.*`):
 
@@ -103,7 +111,7 @@ Configured sources (`narratio.artifact.*`):
   record for the exact configured key and source ID, a complete contract, the
   configured canonical relative path, positive stored size, and stored
   checksum matching bytes read from a confined no-follow regular file; and
-- treat non-current statuses, legacy or malformed records, removed keys,
+- treat non-current statuses, malformed records, removed keys,
   unsafe or missing files, and size/checksum mismatches as unavailable without
   rewriting manifest state. Catalog construction iterates current
   configuration, so removed or renamed records are not advertised.
@@ -135,9 +143,12 @@ Extraction sources (`narratio.extraction.*`):
 
 Previous-session sources (`narratio.previous_session.artifact.*`):
 
-- resolve only from local `previous/` cache state;
+- resolve only from one matching `previous_artifact` input record in the
-- prefer manifest-backed previous-input paths;
+  current manifest, including its exact source ID and checksum;
-- fallback to existing previous-cache filesystem paths.
+- are hydrated only from the unique `published_output` declaration for their
+  canonical source in the selected remote commit; and
+- never infer a path from a previous manifest, current configuration, basename,
+  or existing cache file.
 
 Source absence is evaluated by the consuming artifact input. An optional input
 is omitted from that invocation; a required input fails resolution. This is
@@ -165,25 +176,34 @@ Validation by content type:
 Artifacts package owns shared remote current-state loading mechanics used by
 restore, status and validation checks, and previous-cache planning.
 
+The selected session manifest is accepted only with
+`schema_version: narratio.session-manifest.v1` and complete stage identity.
+An unsupported local or downloaded manifest must be replaced by regeneration,
+verified restore, or republication; artifact resolution never fills in missing
+authority markers.
+
 For a new-protocol current state, the pointer-selected immutable commit is the
 complete restore authority. Callers receive its declared object identities and
-must not supplement them by listing mutable session prefixes. The legacy reader
+must not supplement them by listing mutable session prefixes.
-is intentionally separate and remains migration-only support.
+
+A `published_output` declaration must use a canonical publish-compatible source
+ID and an exact `runs/{run_id}/published/{destination}` key in the same session.
+Only this declaration type may refer to an earlier run; other declared objects
+remain confined to the selected commit's run.
 
 The reader opens each small control object directly and enforces owner-specific
 limits before decoding: 64 KiB for the mutable commit pointer, 4 MiB for the
-immutable commit manifest, and 8 MiB for the selected session manifest. Legacy
+immutable commit manifest, and 8 MiB for the selected session manifest. These
-compatibility applies a 4 KiB limit to `current/run_id.txt` and the same 8 MiB
+are exposed as `MaxCurrentCommitPointerBytes`,
-manifest limit to `current/manifest.json`. These are exposed as
+`MaxRemoteCommitManifestBytes`, and `MaxRemoteSessionManifestBytes`.
-`MaxCurrentCommitPointerBytes`, `MaxRemoteCommitManifestBytes`,
-`MaxRemoteSessionManifestBytes`, `MaxLegacyCurrentRunPointerBytes`, and
-`MaxLegacyCurrentManifestBytes`.
 
 Each read uses the generation and size metadata returned with its opened body.
 Actual bytes remain subject to a limit-plus-one read even if size metadata is
 absent or inaccurate. Immutable selections then retain their declared-size,
 checksum, generation, and identity checks. No current-state control object is
-downloaded through a temporary file.
+downloaded through a temporary file. Artifact and previous-session downloads
+validate their complete sibling temporary file against available committed
+identity before replacing a canonical local path.
 
 Core helpers:
 
@@ -191,10 +211,9 @@ Core helpers:
 - `ValidateCurrentStateIdentity`
 - `RemoteCommitManifest` and `CurrentCommitPointer`
 
-Typed missing-state errors:
+Typed missing-state error:
 
-- `CurrentRunPointerMissingError` (`ErrCurrentRunPointerMissing`)
+- `CurrentCommitPointerMissingError` (`ErrCurrentCommitPointerMissing`)
-- `CurrentManifestMissingError` (`ErrCurrentManifestMissing`)
 
 Identity validation supports caller-provided expectations:
 
@@ -208,6 +227,21 @@ Caller policy is intentionally outside artifacts helpers:
 - some callers downgrade missing state to status/findings;
 - some callers skip optional behavior when state is missing.
 
+## Remote Lock Objects
+
+The artifacts package defines the independent remote-lock payload and identity
+contract. Each canonical publish source maps to one object below the validated
+session `locks/` prefix; its filename is the lowercase SHA-256 digest of the
+exact source followed by `.lock`.
+
+Lock payloads are strict versioned JSON containing the canonical source, an
+optional reason, and a UTC creation time. Reads are bounded to
+`MaxRemoteLockObjectBytes` (64 KiB), reject unknown fields and trailing values,
+and verify that the payload source matches the key digest. Exact lookup reports
+presence explicitly, prefix listing validates every object and sorts by source,
+creation uses an absence precondition, forced replacement uses one observed
+generation, and removal is idempotent.
+
 ## Key Path Helpers
 
 `internal/artifacts/paths.go` and S3-key helpers define canonical helpers for:
@@ -220,8 +254,13 @@ Caller policy is intentionally outside artifacts helpers:
 New publication creates run-scoped immutable objects, including
 `runs/{run_id}/commit.json` and `runs/{run_id}/session-manifest.json`. The sole
 mutable selector is `current/commit-pointer.json`; readers verify its selected
-commit and declared object generations/checksums. Legacy current-pair loading
+commit and declared object generations/checksums.
-is confined to `current_state_legacy.go` for migration only.
+
+A `published_output` declaration may identify its commit's run or an earlier
+immutable run in the same campaign/session namespace. The key parser validates
+that namespace, rejects mutable and malformed keys, and derives the originating
+run from the exact declaration. All other commit artifacts remain confined to
+the commit's own run prefix.
 
 Campaign, session, and Narratio run IDs are validated as portable opaque
 segments at configuration and artifact boundaries before they can be used in a
@@ -252,7 +291,6 @@ physical layout.
   `internal/artifacts/prepared_input.go`
 - Current state: `internal/artifacts/current_state.go`,
   `internal/artifacts/current_state_commit.go`,
-  `internal/artifacts/current_state_legacy.go`
 - Paths and keys: `internal/artifacts/paths.go`,
   `internal/artifacts/s3_keys.go`
 - Previous requirements: `internal/artifacts/previous_requirements.go`

docs/internal/command-restore.md

@@ -11,6 +11,11 @@ physical restore scope belong in
 Restore separates remote authority, local conflict policy, and filesystem
 mutation so each remains testable independently.
 
+The public command supplies production defaults to one private,
+invocation-scoped dependency value. Discovery, planning, and execution use that
+same value, so focused sequencing tests can replace one boundary without
+mutable package hooks.
+
 ## Discovery Contract
 
 Discovery delegates current-state pointer and manifest loading to
@@ -34,18 +39,25 @@ Planner behavior:
 
 - a new-protocol restore uses only the selected commit's declared artifact set;
   each action carries that artifact's immutable key, checksum, size, and
-  generation. Coherent legacy state remains on the isolated compatibility path;
+  generation;
 - remote-to-local mapping is traversal-safe;
 - actions are sorted by local relative path and then remote key;
+- planning opens the selected session root and inspects every target by its
+  normalized session-relative path. Missing ancestors produce a missing target;
+  an ancestor or leaf symlink is a non-regular conflict and is never opened for
+  hashing;
 - force converts differing eligible regular files from conflicts to downloads;
-  directories and other non-regular targets remain conflicts.
+  directories, symlinks, and other non-regular targets remain conflicts.
 
 For a non-dry-run restore, planning/classification happens only after acquiring
 the session lock. Runner manifest/reuse checks acquire that same lock first.
 
 Previous-cache readiness is resolved through `previouscache.Resolve` for restore,
-prepare, status, and validation. A committed source is selected only by its
+prepare, status, and validation. Each requirement selects the unique
-exact source identity; legacy fallback remains isolated and rejects ambiguity.
+`published_output` for its canonical source in the pointer-selected commit,
+then strips only that run's immutable `published/` prefix for its safe local
+cache path. It never probes candidate objects or derives paths from manifests,
+configuration, or basenames.
 
 ## Execution Contract
 
@@ -53,10 +65,14 @@ Execution order and safety:
 
 - non-manifest downloads happen before manifest install;
 - `manifest.json` installs last;
-- downloads use sibling temp files plus atomic rename;
+- downloads use sibling temp files, validate the complete object, then use an
+  atomic rename;
 - manifest replacement is validated before rename;
 - each committed object is verified against its declared checksum, size, and
   generation before installation;
+- installation reopens the selected session root, creates only validated
+  missing directory components with workspace modes, and rejects symlinked or
+  non-regular ancestors and leaves before replacement;
 - a committed manifest already verified during discovery is retained for the
   matching restore action and revalidated before installation, avoiding a
   second body transfer;
@@ -67,15 +83,17 @@ Execution order and safety:
 - restored manifest local references are rebased beneath the selected local
   session root. Unsafe relative references and producer-machine absolute paths
   outside the manifest's producer session root are rejected; producer-local
-  spool/cache and cleanup locations are not restored as authority.
+  spool paths and run-local diagnostic references are not restored as
+  authority. Cleanup remains ordinary restored stage state and is not reset or
+  synthesized by restore.
 
 Audio restore path:
 
 - uses `audio.MaterializeS3Audio`;
 - integrates spool and S3 audio cache paths;
-- reuses cached audio only when its no-follow regular file, content digest, and
+- reuses cached audio only when its opened regular file, content digest, and
   identity sidecar all match the selected remote object version; otherwise it
-  refreshes through the durable download path.
+  refreshes through the validated download path.
 
 ## Reporting Contract
 
@@ -92,9 +110,7 @@ Audio restore path:
   loaded at discovery; later pointer changes cannot add objects or substitute a
   different run into its plan;
 - a verified `current/commit-pointer.json` and its selected immutable commit
-  establish new-protocol remote commitment; coherent legacy
+  establish remote commitment; when absent, republish the session;
-  `current/run_id.txt` plus `current/manifest.json` remains read-only migration
-  support;
 - restore does not execute pipeline stages.
 
 ## Implementation And Tests
@@ -107,4 +123,5 @@ Audio restore path:
 - Tests: `internal/app/restore_discovery_test.go`,
   `internal/app/restore_plan_test.go`,
   `internal/app/restore_execution_test.go`,
+  `internal/app/restore_recovery_test.go`,
   `internal/app/restore_workflow_test.go`

docs/internal/configuration.md

@@ -8,29 +8,21 @@ pipeline-loading boundary implemented by `internal/config`.
 
 `LoadPipeline` assembles and validates a pipeline in this order:
 
-1. Parse the root YAML into a presence-aware composition tree. The tree retains
+1. Give ConfigKit a fresh typed `PipelineConfig` default value, the root file,
-   source names, full field paths, node kinds, declaration order, and explicit
+   enabled imports, and the caller's profile selection. Omitted, empty, and
-   zero, false, empty-map, and empty-list values.
+   exact `default` selections use the base; named profiles use the fixed
-2. Remove the root-only `composition` envelope and validate its explicit
+   `profiles/<name>.yml` convention.
-   `imports`, `default_profile`, and named `profiles` declarations. A load
+2. ConfigKit parses the root and declared imports, additively merges base
-   option retains the difference between omitted and explicitly empty profile
+   sources, applies only the selected overlay, and strictly decodes the result.
-   selection.
+3. The definition resolver applies recorded entry-level and inherited defaults,
-3. Open each import relative to the root pipeline directory through the
+   runs deterministic domain normalization, records derived publish destinations,
-   confined regular-file boundary. Imports must use a `.yml` or `.yaml`
+   and resolves ordinary paths relative to the root pipeline file.
-   extension and cannot traverse, use symlinks, repeat a file, import the root,
+4. The definition validator performs read-only checks that require only the
-   or contain another composition envelope.
+   pipeline file. Checks requiring canonical-party family expansion remain at
-4. Resolve and structurally parse every declared profile overlay through the
+   the existing explicit pipeline/campaign resolution boundary.
-   same confined regular-file boundary. Missing or malformed unselected
+5. Narratio inspects public ConfigKit metadata to reject artifact and family keys
-   overlays fail the load. Overlays cannot contain a composition envelope.
+   declared by multiple base files, adapts source ownership records, attaches
-5. Additively merge the root body and imports. Distinct map leaves compose;
+   runtime-only state, and computes the normalized effective digest.
-   repeated scalar or list paths and node-kind disagreements are conflicts.
-6. Select exactly one declared profile from an explicit option or the default,
-   then recursively merge its overlay. Overlay leaves replace base leaves,
-   lists are atomic replacements, and null or kind changes fail.
-7. Emit deterministic canonical YAML and strictly decode it into
-   `PipelineConfig`.
-8. Apply pipeline defaults once, resolve ordinary relative pipeline paths from
-   the root pipeline file, and digest the normalized effective mapping.
 
 This ordering preserves monolithic configuration behavior. Moving a field to
 an imported fragment changes its source ownership, not its path base, default,
@@ -52,109 +44,152 @@ download retain that exact pipeline object and its private provenance. Removing
 a temporary downloaded session file therefore cannot invalidate the resolved
 pipeline or campaign context.
 
+Campaign and session files use independent ConfigKit definitions with imports
+and profiles disabled. Campaign files normally load directly from their source;
+session files are first read through the 8 MiB bounded-read boundary because the
+application must reject unresolved template placeholders before passing an exact
+byte snapshot to ConfigKit. Local and remotely downloaded session bytes then use
+the same typed loader and retain the existing post-load identity expectations,
+stable-input precedence, and source-relative path semantics.
+
 The application also has a separate read-only inspection resolver for `config
-validate`, `config show`, and `config sources`. It uses the same production root/profile and
+validate`, `config show`, `config sources`, and `config artifacts`. It uses the same production root/profile and
 campaign selection functions, but never routes through session discovery,
-remote-session download, secret loading, adapter composition, workspace
+remote-session download, adapter composition, workspace
 initialization, manifest access, or cleanup. A pipeline with retained artifact
 family declarations must resolve its selected campaign before ordinary pipeline
 validation, which expands its canonical-party members and generated publish
 rules. A pipeline without those declarations may be validated by itself.
+Secret-directory bootstrap remains invocation-scoped and happens before the
+ordinary ConfigKit load because explicit typed references must resolve even for
+inspection.
 
 `MarshalEffectivePipeline` is the configuration-owned projection for `config
-show`. It serializes the typed, defaulted effective mapping through the
+show`. It serializes one explicit typed safe projection with the existing YAML
-deterministic composition renderer, then removes resolution-only artifact
+encoder. The projection has no slots for resolution-only artifact family
-family declarations. The result contains no composition envelope or private
+declarations or private provenance. The result has one trailing newline;
-provenance fields and has one trailing newline; commands do not marshal runtime
+commands do not marshal runtime objects directly.
-objects directly.
 
 `EffectivePipelineSources` and `EffectiveCampaignSources` provide the separate
 safe provenance projection for `config sources`. Pipeline ownership begins with
-the complete logical field paths retained during composition and classifies
+the complete logical field paths projected from the schema and classifies
 each contributor as root, import, profile, or centralized default. The
 projection replaces generated concrete member paths with paired family and
 canonical-party records, and does the same for generated publish rules.
 Campaign records identify campaign-owned fields and party inputs; canonical
-derived players point to the party source, while legacy players retain a
+derived players point to the party source. The application command only joins these sorted
-dedicated legacy-player role. The application command only joins these sorted
 records with selection metadata and never reparses configuration files.
 
-`config diff` uses a paired profile loader that parses the root, imports, and
+`EffectiveArtifactCatalog` projects the same finalized concrete artifact map for
-declared overlays once, then clones the additive base before independently
+`config artifacts`. It derives source identity from each map key, carries sorted
-selecting, decoding, defaulting, and finalizing each profile. When campaign
+effective dependencies, aggregates ConfigKit definition contributors, and adds
+family/character/party provenance for generated members. It does not inspect
+files named after artifacts or infer another runtime graph.
+
+`config diff` uses a thin paired loader that performs two independent ConfigKit
+loads. When campaign
 resolution is needed, the command loads one selected campaign and party and
 expands both effective pipelines from that same party value. The configuration
-owner projects each normalized effective mapping into sorted logical paths;
+owner projects each normalized effective mapping directly from that same typed
+projection into sorted logical paths using JSON number preservation;
 mapping leaves are compared individually while sequence values remain atomic.
 Values are compact deterministic JSON representations for command output, not
 raw YAML fragments, ownership records, or secret material. A differing digest
 with no projected difference is treated as an internal consistency error.
 
-Campaign context construction also reads and classifies the campaign-owned
+Campaign context construction reads the campaign-owned party source through
-party source through `ParseParty`. A canonical party retains its raw bytes and
+`ParseParty`. The required canonical party retains its raw bytes and normalized
-normalized roster in runtime-only `ResolvedParty` provenance, while a legacy
+roster in runtime-only `ResolvedParty` provenance. Canonical resolution creates
-party remains opaque. Canonical resolution creates a virtual
+a deterministic `derived_from_party` players projection. The party is the only
-`derived_from_party` players input and rejects competing campaign or session
+configured roster authority; player and session-party fields are unsupported.
-players files and session party overrides. The compact legacy compatibility
-path resolves the effective campaign/session party and players files together.
 
 ## Canonical Party Domain
 
-`ParseParty` is the package-owned boundary for classifying a party source.
+`ParseParty` is the package-owned boundary for parsing a party source. Its
-When a top-level `schema_version` is present, it strictly validates the
+single-source ConfigKit definition disables imports and profiles, requires
-`narratio.party.v1` contract into ordered character domain values. The
+`schema_version` through `VersionRule`, and strictly decodes the
-canonical value retains a separate exact byte copy of its source so consumers
+character-keyed typed schema before applying Narratio's party-domain checks.
-can materialize the authored party document without reserializing it. Its
+Successful metadata coordinates recover authored character-map order without a
-`PlayersYAML` method deterministically derives the versioned players-only
+second YAML parse; class and alias slice order comes directly from typed decode.
-projection.
+The canonical value retains a separate exact byte copy of its source so
+consumers can materialize the authored party document without reserializing it.
+Its `PlayersYAML` method deterministically derives the ID-sorted, versioned
+players-only projection.
 
-An unversioned source is classified by the small legacy compatibility boundary
+An unversioned source is rejected before resolution, family expansion, or
-in `party_legacy.go`; it deliberately exposes no parsed roster information.
+prepare-stage materialization.
-That boundary exists solely to isolate removable compatibility behavior from
+
-the canonical parser.
+## Validation Ownership
+
+Configuration validation remains within `internal/config`, with the public
+entry points and pipeline-domain sequencing separated from the private owners
+for core pipeline/session rules, external adapters, publish/storage, Notarius,
+Scriptorium, and shared scalar/path checks. These are package-local ownership
+boundaries rather than extension points or a general validation framework.
+
+Validation is read-only, fails fast, and does not define which of several
+independent invalid fields is reported first. Map-backed declarations are
+nevertheless traversed deterministically. Reusable domain normalization is an
+explicit preceding operation for both loaded and programmatically constructed
+pipelines. Notarius normalization commits canonical scalar, reference, and output
+values only after the complete normalization succeeds; its validator then checks
+that candidate without changing it. Scriptorium validates artifact declarations
+and outputs, dependencies and inputs, variables, referenced outputs, and the final
+dependency graph in deterministic artifact order.
+
+Diagnostics identify the relevant semantic field. Merge conflicts also
+retain every claiming source needed to correct an additive ownership conflict.
+Complete prose, punctuation, formatting, and private validation choreography are
+not compatibility contracts.
 
 ## Diagnostics And Runtime Metadata
 
-Syntax, duplicate-key, composition, conflict, and schema failures include the
+Syntax, duplicate-key, merge, profile, and schema failures include the
 relevant source name and full field path. Additive conflicts report every
 claiming source so operators can repair the split without repeatedly
 rediscovering additional conflicts.
 
 The loaded pipeline retains private runtime metadata for the absolute root
 path, ordered imports, selected profile name and selection source, selected
-overlay, contributing sources, effective digest, and leaf ownership. Base
+overlay, original ConfigKit metadata and source snapshots, effective digest,
+and leaf ownership. Base
 leaves retain their root/import owners, replaced leaves belong to the selected
 overlay, and centrally supplied values use the synthetic `default` owner. This
 metadata does not participate in YAML decoding or alter the public
 configuration model.
 
-The effective digest is SHA-256 over deterministic canonical YAML produced from
+The effective digest is SHA-256 over a versioned deterministic JSON encoding of
-the defaulted `PipelineConfig`. Runtime Notarius paths remain absolute for
+the same typed safe projection used by show and diff. Runtime Notarius paths remain absolute for
 execution, but the digest substitutes their normalized logical values captured
 before root-relative resolution, so relocating an equivalent configuration
-bundle does not change provenance. Because composition and resolution metadata
+bundle does not change provenance. Because ConfigKit and resolution metadata
 are private, the digest excludes source layout, profile name, and ownership.
-Configuration stores environment variable names rather than resolving raw
+Configuration stores credentials as ConfigKit typed references. A successful
-credentials, so raw secret values are neither loaded nor hashed.
+load resolves their values inside the opaque typed boundary, while this safe
+projection serializes only `EnvName()` identities. Credential-value changes are
+therefore never hashed; reference-name changes remain ordinary provenance.
+The optional Notarius credential follows the same path even for disabled or
+resumable extraction; normalization copies the opaque reference without reading
+or resolving it again. Direct YAML serialization has a configuration-owned safe
+projection so only the reference name can leave this boundary.
 `recomputePipelineEffectiveDigest` is the single package-owned refresh point
 for later runtime expansion.
 
 ## Test Surfaces
 
-`composition_test.go` protects the presence and merge algebra independently of
+`pipeline_configkit_test.go` exercises real imports and fixed-path profiles,
-the public schema. `pipeline_composition_test.go` exercises explicit imports,
+base selection, unselected-overlay isolation, source attribution, and Narratio's
-confinement, conflicts, strict decoding, metadata, and root-relative path
+unique artifact/family base-declaration rule. Existing focused configuration
-behavior through `LoadPipeline`. `pipeline_profiles_test.go` covers selection,
+tests protect defaults, explicit-value semantics, strict decoding, validation,
-all-overlay validation, overlay behavior, provenance, option propagation, and
+metadata, and root-relative path behavior. Application configuration-loader tests protect the
-effective-digest stability. Application configuration-loader tests protect the
 single-read boundary by changing the pipeline file after its initial load and
 confirming local session resolution retains the original pipeline. Other
 configuration tests continue to protect defaults and validation after assembly.
-`party_test.go` protects the versioned party schema, domain invariants, and
+`party_test.go` protects the versioned typed party schema, domain invariants,
-deterministic players projection without involving campaign or runtime wiring.
+authored character order, retained bytes, and deterministic players projection
+without involving campaign or runtime wiring.
 `party_resolution_test.go` protects campaign-owned party loading, canonical
-input restrictions, legacy overrides, source provenance, and virtual players
+input restrictions, source provenance, and virtual players input selection.
-input selection.
 
 ## Artifact Family Resolution
 

docs/internal/fileops.md

@@ -1,71 +1,55 @@
 # Internal: File Operations
 
-`internal/fileops` owns the narrow mechanics for durable replacement of one
+`internal/fileops` owns the narrow mechanics for atomically visible replacement of one
 byte file. Callers keep ownership of serialization, validation, cancellation,
 and destination-directory policy.
 
-## Destination Confinement
+## Destination Paths
 
-Before it creates, replaces, or installs a destination file, `fileops` opens
+Ordinary workspace replacement uses standard operating-system path semantics.
-each ancestor from the filesystem root and rejects symbolic links or components
+Callers with a root-scoped safety contract own that validation at their boundary.
-that change during traversal. The resulting parent-directory handle is retained
-for sibling temporary-file creation and rename, so a later pathname swap cannot
-redirect the replacement. Existing destination symlinks are replaced as leaf
-entries; their targets are never followed.
 
-Remote object acquisition uses a writer supplied by the storage owner. The
+Remote object acquisition is owned by its domain caller. The caller writes a
-writer receives a `fileops`-owned, already-open sibling temporary file rather
+sibling temporary file, validates the complete result, and passes that file to
-than a mutable destination path. Callers still own remote object selection,
+`InstallSiblingTempFile` only when it is ready to become canonical. Callers own
-validation, conflict handling, and final mode.
+remote object selection, validation, conflict handling, and temporary-file
+cleanup.
 
-Directory promotion keeps the verified destination parent open while it creates
+Directory bundles are not promoted by `internal/fileops`; extraction owns its
-the temporary tree, copies regular source entries, and performs the platform
+validated Notarius bundle rename.
-no-replace rename. Platforms without a verified handle-relative atomic
-no-replace primitive reject promotion before writing a temporary tree.
 
 ## Cleanup Contract
 
-`RemoveAllUnderRoot` accepts an explicit root and a proper descendant. It opens
+`RemoveAllUnderRoot` accepts an explicit root and a proper descendant, opens
-the root and each target ancestor without following symlinks, then removes the
+that root with `os.OpenRoot`, and removes the normalized root-relative target
-tree through those directory handles. It rejects root deletion and any symlink
+with `Root.RemoveAll`. It rejects root, traversal, and out-of-root targets; a
-encountered in the target path or tree; repeated removal of a missing target is
+missing target is successful. `RemoveFileUnderRoot` additionally requires an
-successful. Command and post-publish policy remains owned by `internal/app`.
+existing leaf to be regular. Command and post-publish policy remains owned by
+`internal/app`.
 
-## Confined Reads
+## Bounded Reads
 
-`ReadRegularFileUnderRoot` is the no-follow, bounded read primitive for a
+`ReadFileBounded` opens a normal path and reads at most its caller-supplied
-caller-selected root and relative file path; `ReadRegularFile` is its
+limit. Credential mode policy and environment precedence remain owned by
-path-based convenience wrapper. They verify every ancestor through directory
+`internal/app`.
-handles and admit only a stable regular-file handle. Callers enforce their own
-byte limits and access policy. Credential mode policy and environment
-precedence remain owned by `internal/app`.
 
 ## Replacement Contract
 
 `ReplaceFileAtomic` requires an existing destination directory. It creates a
 sibling temporary file, writes the complete byte sequence, applies the
-caller-supplied mode, syncs and closes the file, runs an optional pre-rename
+caller-supplied mode, closes the file, runs an optional pre-rename check, and
-check, replaces the destination with a rename, then syncs the containing
+replaces the destination with a rename.
-directory.
 
 The pre-rename check is the last point at which a caller can cancel without
 installing a new destination. A failure before the rename leaves the old
-destination unchanged and removes the temporary file; any cleanup failure is
+destination unchanged and removes the temporary file. Successful replacement
-returned alongside the primary failure. A failure after the rename may leave
+provides atomic visibility; it makes no crash- or power-loss durability claim.
-the new file visible, but it is not reported as crash-durable.
 
 Replacement follows the operating system's same-filesystem rename semantics.
 If a platform cannot replace an existing destination, the operation returns an
 error and never removes the old file as an emulation step.
 
-## Directory-Sync Support
+`WriteFileAtomic` and the copy helpers create the destination parent with the
-
+repository's workspace permissions. `InstallSiblingTempFile` requires the
-Linux and macOS attempt to sync the destination directory. Windows opens the
+caller to provide an existing temporary file in the existing destination
-directory with backup semantics and flushes its buffers. If either operation
+directory.
-is unavailable for the platform, directory handle, or filesystem,
-`ErrDirectorySyncUnsupported` is returned. Narratio does not treat that result
-as successful crash-durable replacement.
-
-`WriteFileAtomic`, copy helpers, and downloaded temporary-file installation
-retain their compatibility behavior of creating the destination parent with
-the repository's workspace permissions before using this contract.

docs/internal/manifest.md

@@ -2,29 +2,38 @@
 
 ## Purpose
 
-Explain the session-progress and invocation-audit models implemented by
+Explain the session-progress model implemented by `internal/manifest`. Physical
-`internal/manifest`. Physical manifest placement belongs in
+manifest placement belongs in
 [Operations](../operations.md#local-state-layout).
 
 ## Session Manifest
 
 `manifest.Manifest` records:
 
+- required `schema_version: narratio.session-manifest.v1`
 - identity (`session_id`, `campaign`, `run_id`)
 - local path metadata (`local_workdir`, `local_spool_dir`)
 - remote identity metadata (`s3_bucket`, `s3_session_prefix`, `s3_run_prefix`)
 - `inputs` records
 - durable `artifacts` records
-- per-stage `stages` map
+- per-stage `stages` map, including the ordinary cleanup stage record
-- an optional `post_publish_cleanup` obligation, which binds a committed run,
-  remote commit identity, and each exact root-confined local target to its
-  completion evidence
 
 Session, campaign, and run identities in local and downloaded manifests must be
 portable opaque segments. Unsafe legacy identities are rejected with migration
 guidance rather than being normalized into a different workspace or remote
 namespace.
 
+Decoding is strict at every typed manifest field. Unknown fields are rejected;
+the schema version remains `narratio.session-manifest.v1` and there is no
+compatibility reader that silently drops retired state. Cleanup progress uses
+only the cleanup entry in `stages`.
+
+Every persisted session manifest has a non-nil `stages` map. Each entry is
+non-nil and its embedded `name` exactly matches its map key; loading never
+invents either marker. A missing or unsupported schema version, or incomplete
+stage identity, is unsupported local state: move it aside and regenerate it,
+or restore a verified current committed snapshot when that is authoritative.
+
 Prepare records independent `party` and `players` input checksums. In canonical
 party mode, the party record retains its campaign source identity while the
 players record uses `derived_from_party`; raw roster content is never embedded
@@ -41,13 +50,52 @@ The model admits these stage states:
 - `stale`
 - `interrupted`
 
+### Transcribe-owned speaker state
+
+A `transcribe` stage record may carry `transcribe_state_version: 1` and a
+`transcribe_speakers` map keyed by the existing derived speaker ID. The version
+is the authority marker: a versioned empty map means that no earlier speaker
+completion is reusable. Absence of both fields is accepted as legacy
+aggregate-only state so current Narratio can read older manifests.
+
+Each map entry represents one completed speaker only. It binds a versioned
+fingerprint to the session-relative audio path, audio SHA-256 and size, one
+canonical `transcript_raw` output with its path, SHA-256 and size, and the
+producer run ID. Paths must be canonical and session-relative, checksums must
+be lowercase SHA-256 values, and the record and output producer identities
+must agree. Other stages cannot carry this transcribe-owned state.
+Audio size is nonnegative; WhisperX determines whether that audio can be
+transcribed. Output size remains positive and the transcript must be valid JSON.
+
+Transcribe reconciles this collection against the current prepared audio set
+and observable WhisperX semantics. Missing or corrupt regular outputs become
+work; unsafe or non-regular evidence is rejected. It saves the reconciled
+collection before HTTP dispatch and saves a detached candidate after each
+validated completion. Only after a save succeeds is the candidate installed as
+live authority. Canonical bytes produced before an unsuccessful save are not a
+completion record.
+
+Aggregate transcribe status still describes the whole invocation. A failed,
+running, stale, or interrupted aggregate may therefore retain reusable speaker
+records, while aggregate outputs are declared only after the current
+collection is complete. Ordinary retry reconciles and reuses those records.
+Force first checkpoints an empty current collection, preventing later retries
+from reviving pre-force results.
+
+Older aggregate-only transcribe success is conservatively regenerated once
+when transcribe is selected because it does not bind individual outputs to the
+audio bytes used. The session-manifest schema identifier remains
+`narratio.session-manifest.v1`; this additive state is accepted by current
+Narratio, but older strict-decoding binaries may reject manifests after these
+fields have been written.
+
 ### Analyze-owned artifact state
 
-The `analyze` stage record may carry `analyze_state_version: 1` and an
+Every `analyze` stage record carries `analyze_state_version: 1` and an
 `analyze_artifacts` map keyed by normalized configured artifact key. The
 version is the authority marker: version 1 with no entries is a valid evaluated
-empty set, while an absent version is legacy aggregate-only state and provides
+empty set. An absent or unsupported version is unsupported local state and is
-no current configured-artifact evidence.
+rejected rather than being treated as aggregate-only authority.
 
 Each analyze artifact record has one disposition:
 
@@ -72,18 +120,15 @@ do not replace the concrete configured key or infer family membership from a
 name, so older records without them remain valid.
 
 The session-stage collection is the reconciled authority across invocations.
-The corresponding collection on an invocation's `analyze` stage record is an
+These records remain analyze-owned data inside the fixed stage; they are not
-audit of only the artifacts evaluated or attempted by that run. These records
+dynamic stages or generic subtasks.
-remain analyze-owned data inside the fixed stage; they are not dynamic stages
-or generic subtasks.
 
-The stage result contract has one analyze-specific projection boundary. On
+Analyze checkpoints one complete validated candidate collection after each
-success, the runner validates and deep-copies the complete reconciled session
+successful artifact and before exposing that artifact to later work. It saves a
-collection and the invocation subset. Aggregate session outputs are rebuilt in
+detached manifest candidate and installs its version, collection, and updated
-configured-key order from current session records only; invocation outputs are
+timestamp into live state only after the store reports success. Aggregate stage
-limited to current records produced by that invocation's run ID. Ordinary
+outputs are ordinary result outputs rebuilt in configured-key order from current
-stage outputs cannot accompany this projection, so there is one source of
+session records only.
-artifact authority.
 
 Successful incremental execution replaces only evaluated artifact records and
 preserves valid unrelated current records. Rebuilt outputs are compared by
@@ -92,43 +137,31 @@ the same recomputed fingerprint to remain current, while a changed identity
 removes output authority from every unselected transitive dependent by marking
 it stale. A partial analyze invocation can therefore succeed while unrelated
 configured records remain stale. Existing canonical files never create current
-records without validated execution and projection.
+records without validated execution and a successful checkpoint.
 
 Aggregate analyze status is deliberately coarser than this collection. Resume
 validation may skip a succeeded aggregate record when the selected artifact
 closure is current even if unrelated records are stale. Conversely, a stale
 aggregate record may cross the ordinary runner boundary and perform zero
 Scriptorium calls when reconciliation proves every selected artifact current;
-the successful projection then restores the aggregate status.
+ordinary successful stage completion then restores the aggregate status.
 
-Analyze may return a projection together with an error. That restricted result
+On an artifact error, analyze installs a complete retry collection on the live
-cannot carry ordinary outputs, skip state, aggregate logs, generated configs,
+manifest before returning. The active record is failed without output, current
-or metadata. The runner persists only the validated per-artifact collections,
+transitive dependents are stale, unrelated current records survive, and only
-then marks the aggregate analyze and run state failed and invalidates delivery
+earlier checkpointed completions remain current. The runner's ordinary failure
-dependents conservatively. Unrelated current records survive because the
+path marks aggregate analyze failed, clears aggregate outputs, invalidates
-session projection is complete. A malformed projection is not applied, and a
+delivery dependents, and saves that collection. A terminal session-save failure
-failed session projection save restores the prior per-artifact authority before
+is joined with the original adapter, validation, filesystem, cancellation, or
-terminal failure persistence.
+checkpoint cause.
 
-The incremental executor constructs this restricted projection at each
+If an artifact checkpoint fails, analyze rolls its working collection back to
-scheduled artifact boundary. The active record is failed without output,
+the last committed collection before constructing the active failed record.
-current transitive dependents are stale, unrelated current records survive, and
+Newly materialized canonical bytes therefore remain incidental and cannot
-only earlier validated and materialized completions remain current in the
+become current authority. A later ordinary run reconciles the durable records,
-invocation subset. Session failure state is persisted before invocation failure
+reuses earlier current artifacts, and resumes with the failed or stale work.
-state. If either terminal save fails, its persistence error is joined with the
+The same recovery applies after process interruption: the aggregate record may
-original adapter, validation, or filesystem cause; a failed projection save
+be non-successful, but every completed artifact checkpoint remains reusable.
-does not turn incidental canonical bytes into manifest authority.
-
-## Run Manifest
-
-`manifest.RunManifest` is created for each invocation and records:
-
-- invocation identity and `force` flag
-- the selected profile (when any) and secret-free effective configuration digest
-- requested stages
-- per-stage action (`run` or `skip`)
-- per-stage status
-- overall run status (`running`, `succeeded`, `failed`)
 
 ## Remote Commit Manifest
 
@@ -145,38 +178,51 @@ binds that object by checksum, size, and generation. Readers strictly reject
 unknown fields, version mismatches, pointer/commit identity mismatches, and
 objects that do not match their declaration.
 
-The reader retains a temporary, clearly isolated compatibility path for a
+An absent commit pointer means no remote state is published. Republish to
-coherent legacy `current/manifest.json` plus `current/run_id.txt` pair. That
+establish a current immutable snapshot.
-path is removable after migration and is never used to write new state.
 
 ## Persistence Semantics
 
 `manifest.LocalStore`:
 
-- validates loaded documents;
+- accepts only the exact supported session-manifest schema version;
-- normalizes missing maps/stage records;
+- rejects unknown fields and multiple JSON values;
-- writes through a sibling temporary file, syncing the completed file and
+- validates stage-map identity without synthesizing missing maps, records, or
-  destination directory after atomic replacement;
+  names;
+- retains only representation-preserving collection normalization;
+- writes through a sibling temporary file and atomic replacement visibility;
 - updates `updated_at` on save.
 
-If the operating system or filesystem cannot sync a directory, save returns an
+Manifest replacement provides atomic visibility, not a crash- or power-loss
-explicit error instead of claiming crash-durable replacement. A returned error
+durability guarantee. Callers reload after a returned save error before retrying.
-after the rename can therefore leave the new manifest visible but not confirmed
+A transcribe or analyze checkpoint error leaves the prior manifest collection
-durable; callers must reload it before retrying.
+as authority. The runner then attempts to persist the aggregate failure; if
+that terminal save also fails, its error is joined with the originating failure
+and the next invocation must reconcile from the last readable manifest.
 
 ## Execution Semantics
 
-The application runner marks an executing stage running and then succeeded or
+Planning and execution share one pure pre-execution decision over the recorded
-failed in both manifests, persisting each transition. On success it records
+stage status, force selection, and normalized resume-validation result. That
-outputs, logs, generated configuration references, metadata, and—when the
+decision identifies run versus reuse and any required stale and dependent
-stage implements the optional contract—a versioned semantic-configuration
+invalidation transition. Planning applies it only to a cloned manifest.
-fingerprint. Artifact
+Execution remains the sole owner of observed stage results and persistence.
+
+For each selected stage, the application runner computes semantic evidence,
+performs any eligible resume validation, applies the pre-execution decision,
+and persists required invalidation before either reusing the existing success
+or beginning an attempt. The same one-stage execution owner marks the attempt
+running before invoking the stage and then records and persists its failed,
+skipped, or successful terminal state before a later stage may start. On
+success it records outputs, logs, generated configuration references, metadata,
+and—when the stage implements the optional contract—a versioned
+semantic-configuration fingerprint. Artifact
 records may include optional contract and external provenance objects; old
 manifests remain compatible when those fields are absent. A successful forced
 rerun marks only succeeded transitive dependent session-stage records stale.
 The application owns a fixed dependency relation distinct from execution order;
 dependents are returned in canonical order. Render and extract therefore never
-stale one another, while either can stale analyze, publish, and notify.
+stale one another, while either can stale analyze, publish, notify, and cleanup.
 
 Starting an execution clears the current session-stage record's prior outputs,
 logs, generated configuration references, metadata, and semantic fingerprint.
@@ -184,40 +230,36 @@ Failed and skipped
 transitions enforce the same clearing rule directly, while success repopulates
 only fields returned by the new result. Marking a record stale does not clear
 those details because resume validation and diagnosis may still require them
-before execution begins. Invocation run manifests remain immutable audit
+before execution begins.
-records of their own outcomes.
 
-Aggregate lifecycle clearing deliberately preserves the analyze-owned
+Aggregate lifecycle clearing deliberately preserves both the transcribe-owned
-per-artifact collection. This lets later reconciliation replace only evaluated
+speaker collection and the analyze-owned per-artifact collection. This lets
-entries without erasing unrelated current results. Other stages retain their
+later reconciliation replace only non-current entries without erasing durable
-existing aggregate-only lifecycle behavior and are forbidden from carrying the
+completions. Other stages retain aggregate-only lifecycle behavior and are
-analyze-specific fields.
+forbidden from carrying either collection.
 
 A stage may explicitly return a skipped disposition and stable reason. The
-runner persists that outcome in both manifests, clears older outputs for the
+runner persists that outcome in the session manifest, clears older outputs,
-session-stage record along with older logs, generated configuration references,
+logs, generated configuration references, and metadata, then applies any
-and metadata, then applies any bounded details from the current skip and
+bounded details from the current skip and continues. This self-skip is distinct from deciding not to execute an
-continues. This self-skip is distinct from deciding not to execute an
 already-succeeded stage and is reconsidered on later runs. Skipped results
 cannot contain outputs. An intentional self-skip records the current semantic
 fingerprint because it is a completed, reusable stage result; failed or
 interrupted work never promotes one.
 
-When an already-succeeded stage is skipped, the invocation run manifest records
+When an already-succeeded stage is skipped, the session manifest deliberately
-the `skip` action and reason. The session manifest deliberately retains its
+retains its existing succeeded record because it remains the cross-invocation
-existing succeeded record because it remains the cross-invocation progress
+progress authority. If a stage supplies semantic configuration evidence, reuse first
-authority. If a stage supplies semantic configuration evidence, reuse first
 requires the persisted positive schema version and lowercase SHA-256 digest to
-match the current resolved stage semantics. Missing legacy evidence, malformed
+match the current resolved stage semantics. Missing semantic evidence, malformed
 evidence, or a mismatch makes the stage and its fixed transitive dependents
-stale. The invocation skip copies the matched fingerprint for provenance but
+stale. The existing stage-specific resume
-does not rewrite session authority. The existing stage-specific resume
 validator runs only after this semantic check succeeds; both checks are
-required. Extraction and analyze have resume validators and may reject an
+required. Extraction, analyze, and cleanup have resume validators and may
-otherwise eligible skip when their selected durable evidence is obsolete; the
+reject an otherwise eligible skip when their selected durable evidence is
-runner marks the aggregate record stale and executes it. Analyze's validator
+obsolete; the runner marks the aggregate record stale and executes it.
-can still accept a partial selection when only unrelated artifact records are
+Analyze's validator can still accept a partial selection when only unrelated
-stale.
+artifact records are stale.
 
 Implemented reuse coverage is deliberately split between aggregate semantic
 evidence and focused durable validators:
@@ -225,7 +267,7 @@ evidence and focused durable validators:
 | Work | Reuse authority | Focused owners |
 | --- | --- | --- |
 | prepare | aggregate semantic fingerprint | [prepare](stage-prepare.md) |
-| transcribe | aggregate semantic fingerprint | [transcribe](stage-transcribe.md), [WhisperX](../integrations/whisperx.md) |
+| transcribe speakers | aggregate semantic fingerprint plus per-speaker input/output reconciliation | [transcribe](stage-transcribe.md), [WhisperX](../integrations/whisperx.md) |
 | merge | aggregate semantic fingerprint | [merge](stage-merge.md), [Seriatim](../integrations/seriatim.md) |
 | polish | aggregate semantic fingerprint | [polish](stage-polish.md), [Audita](../integrations/audita.md) |
 | normalize | aggregate semantic fingerprint | [normalize](stage-normalize.md), [Seriatim](../integrations/seriatim.md) |
@@ -235,6 +277,7 @@ evidence and focused durable validators:
 | analyze artifacts | per-artifact fingerprint, reconciliation, and output validator | [analyze](stage-analyze.md), [Scriptorium](../integrations/scriptorium.md) |
 | publish | aggregate semantic fingerprint plus immediate lock/commit checks | [publish](stage-publish.md), [storage adapter](adapters.md) |
 | notify | aggregate delivery-mode fingerprint | [pipeline overview](overview.md), [configuration](../config.md#notifications) |
+| cleanup | aggregate semantic fingerprint plus publish-evidence and target-absence validator | [cleanup](stage-cleanup.md), [workspace](workspace.md) |
 
 These contracts record resolved choices Narratio can observe, not operational
 runner tuning. External model, module, prompt, profile, and configuration-file
@@ -242,10 +285,8 @@ contents that a tool privately loads remain outside the contract when their
 configured identifier is unchanged; operators must force the affected work
 after such a private content change.
 
-Session manifest is the authoritative stage-progress ledger across invocations.
+The session manifest is the authoritative stage-progress ledger across
-Run manifest is invocation-scoped audit state.
+invocations. It retains the most recently resolved invocation's bounded
-
-Both manifests retain the most recently resolved invocation's bounded
 configuration provenance. It identifies the selected profile name and source
 (`default` or `cli`) plus the effective configuration digest, but never a raw
 secret or profile content. This provenance is informational: it does not
@@ -254,9 +295,9 @@ invalidates only stages whose semantic configuration changed. When a private
 external-tool model, module, prompt, or profile changes behind an unchanged
 configured identifier, use `--force` for the affected work.
 
-`session plan` computes the same current fingerprint and applies the same
+`session plan` computes the same current fingerprint and applies the shared
-comparison and invalidation rules to a cloned manifest. It predicts the runner
+decision and invalidation rules to a cloned manifest. It predicts the runner
-decision without persisting session or invocation state. The shared helper
+decision without persisting session state. The semantic fingerprint owner
 hashes deterministic JSON from stage-owned typed structs; stage providers must
 exclude secrets, complete effective-configuration dumps, and operational
 values that cannot affect canonical results. Concrete coverage is owned by the
@@ -266,7 +307,7 @@ Before an explicitly bounded execution starts after `prepare`, the application
 reads the session manifest and accepts only `succeeded` or `skipped` for every
 excluded canonical prefix stage. The first other status or absent record fails
 the request before layout mutation, adapter initialization, session-manifest
-writes, or run-manifest creation. Excluded prefix records are not passed to
+writes, or run-directory creation. Excluded prefix records are not passed to
 resume validators. Records after the selected end are not prerequisites and
 may be made stale by selected work without being scheduled.
 
@@ -283,34 +324,32 @@ session manifest must agree on campaign and session identity before execution;
 the current projection is refreshed for every invocation while stage progress,
 inputs, and durable artifacts remain session history.
 
-For handled failures after an invocation record is created, the runner records
+For handled failures after the session identity transition, the runner records
-the failure on the session ledger and persists it before persisting the failed
+the failure on the session ledger and atomically attempts its save while
-run audit record. This preserves the resume authority while making a partial
+preserving the originating error. Abrupt process death remains an accepted case
-persistence disagreement visible. Abrupt process death remains an accepted case
+where the last session transition can require operator interpretation.
-where a durable running record can require operator interpretation.
 
 ## Invariants
 
 - stage resume/skip decisions are session-manifest driven.
 - semantic fingerprint comparison precedes stage-specific resume validation.
 - only successful and intentional-skipped results promote current semantic
-  evidence; invocation reuse copies evidence without replacing session state.
+  evidence.
 - running, failed, and self-skipped stages do not retain result payloads from
   an earlier success.
 - stale stages retain prior details until replacement execution starts.
 - force reruns stale succeeded stages in the fixed dependency relation.
-- run manifest does not replace session manifest as progress authority.
+- the session manifest is the only durable local progress ledger.
 - remote commitment is established by a verified current pointer and remote
   commit relationship, never by a mutable session-manifest boolean.
 
 ## Implementation And Tests
 
-- Models and transitions: `internal/manifest/manifest.go`,
+- Models and transitions: `internal/manifest/manifest.go`
-  `internal/manifest/run_manifest.go`
 - Remote commit model and readers: `internal/artifacts/remote_commit.go`,
   `internal/artifacts/current_state_commit.go`,
-  `internal/artifacts/current_state_legacy.go`
 - Persistence and validation: `internal/manifest/store.go`
+- Transcribe speaker state: `internal/manifest/transcribe_state.go`
 - Package tests: `internal/manifest/*_test.go`
 - Assembled execution behavior: `internal/app/runner_test.go`,
   `internal/app/run_stage_test.go`

docs/internal/overview.md

@@ -22,20 +22,29 @@ session safety controls, and runs commands. Pipeline commands execute the
 canonical stage sequence through adapter interfaces, while manifests record
 progress and artifact services resolve durable inputs and outputs.
 
+The pipeline runner follows one explicit invocation lifecycle: inspect bounded
+prerequisites without mutation, prepare the workspace layout, acquire one
+session lock, load or create the authoritative session manifest, compose only
+the collaborators needed by the selected range, execute each selected stage in
+order, and then return the summary. Cleanup, when selected, is the ordinary
+final stage rather than runner finalization work.
+The lock remains held from manifest initialization through finalization, and a
+failed stage or required manifest save prevents later stages from starting.
+
 ## Components
 
 | Area | Implemented owners | Responsibility |
 | --- | --- | --- |
 | Executable | `cmd/narratio` | Process entry, standard stream wiring, argument handoff, and exit status. |
-| Application orchestration | `internal/app` | Command dispatch, configuration selection, secret-file environment loading, production composition, session locking, planning, execution, restore, cleanup gates, and user-facing reporting. |
+| Application orchestration | `internal/app` | Command dispatch, configuration selection, secret-file environment loading, production composition, session locking, planning, execution, restore, and user-facing reporting. |
-| Configuration | [`internal/config`](configuration.md) | Presence-aware root/import/profile composition, canonical party and family expansion, strict YAML loading, defaults, normalization, session templating, and validation. |
+| Configuration | [`internal/config`](configuration.md) | ConfigKit-backed pipeline imports and profiles, canonical party and family expansion, strict YAML loading, defaults, normalization, session templating, and validation. |
 | Pipeline stages | `internal/stage` | Canonical stage registry, shared stage contract, execution dependencies, and implemented stage behavior. |
 | External boundaries | `internal/adapters`, `internal/audio` | WhisperX HTTP, downstream subprocesses, notification, object storage, and S3 audio materialization behind Narratio contracts. |
-| Manifests | `internal/manifest` | Durable session progress, invocation audit state, stage transitions, validation, and atomic persistence. |
+| Manifests | `internal/manifest` | Durable session progress, stage transitions, validation, and atomic persistence. |
 | Artifacts and paths | `internal/artifacts`, `internal/pathsafe` | Artifact identities and resolution, local and remote path/key models, current-state discovery, and confined relative destinations. |
 | Previous-session cache | `internal/previouscache` | Deterministic planning and materialization requirements for configured previous-session inputs. |
 | Artifact policy | `internal/artifactpolicy` | Source and destination policy, configured artifact identity validation, and publish destination safety. |
-| Shared models and file operations | `internal/artifactmodel`, `internal/contracts`, [`internal/fileops`](fileops.md) | Transcript and artifact data contracts plus durable single-file replacement helpers; unsupported directory syncing is reported explicitly. |
+| Shared models and file operations | `internal/artifactmodel`, `internal/contracts`, [`internal/fileops`](fileops.md) | Transcript and artifact data contracts plus atomic single-file replacement, bounded reads, workspace modes, and root-scoped cleanup helpers. |
 | Logging | `internal/logging` | Application logger construction and shared structured logging behavior. |
 
 The application boundary composes concrete implementations. Stages depend on
@@ -43,10 +52,20 @@ Narratio-level contracts; external transport and SDK details remain in
 adapters. The normative rules for these relationships remain in
 [Architecture](../policy/architecture.md).
 
+Command execution normalizes one private dependency value per invocation.
+Pipeline configuration-path discovery, configuration reads, secret-directory
+reads, remote downloads, object-store construction, and restore phase functions
+use that value throughout the command. Public command entry points supply the
+canonical discovery and production defaults, while focused tests may provide
+invocation-owned discovery candidates or replace only the external or sequencing
+boundary they exercise without mutating package state.
+
 Pipeline execution and `session plan` share the same inclusive contiguous-range
-model. Planning clones session state and applies selected-stage transitions and
+model and the same pure pre-execution decision over recorded status, force, and
-resume validation in memory; it does not create invocation state or initialize
+normalized resume validation. Planning applies that decision and modeled
-stage-execution adapters. Command configuration loading can still retrieve a
+selected-stage transitions to cloned session state; it does not persist, create
+invocation state, initialize stage-execution adapters, or claim to observe an
+execution result. Command configuration loading can still retrieve a
 missing session file through configured remote storage. It retains the initially
 composed pipeline and selected campaign while resolving either a local or
 downloaded remote session, so one invocation cannot mix pipeline revisions.
@@ -68,6 +87,7 @@ The implemented canonical order is:
 9. [`analyze`](stage-analyze.md)
 10. [`publish`](stage-publish.md)
 11. `notify` (no-op)
+12. [`cleanup`](stage-cleanup.md)
 
 `notify` currently has no persisted pipeline outputs and uses the explicit
 `noop` notification mode. Its versioned semantic evidence records that delivery
@@ -81,18 +101,19 @@ registry owns the flat execution sequence. The application orchestration owner
 uses a fixed, validated dependency relation to find transitive dependents in
 canonical order. In particular, `render` and `extract` are sibling consumers of
 trimmed transcript state: neither invalidates the other, while either can stale
-`analyze`, `publish`, and `notify`.
+`analyze`, `publish`, `notify`, and `cleanup`. Cleanup is a dependent of notify,
+so any earlier replacement can stale a prior cleanup result.
 
 ## Focused Documentation
 
-- [Configuration Internals](configuration.md): pipeline composition, import
+- [Configuration Internals](configuration.md): pipeline imports and profiles,
-  confinement, field ownership, decoding, and root-relative path semantics.
+  field ownership, decoding, and root-relative path semantics.
 - [Adapter Internals](adapters.md): external adapter boundaries, composition,
   failure behavior, and test surfaces.
 - [Artifact Internals](artifacts.md): source identities, runtime catalog,
   resolution, previous requirements, and current-state helpers.
-- [Manifest Internals](manifest.md): session and run records, persistence, and
+- [Manifest Internals](manifest.md): session records, persistence, and execution
-  execution transitions.
+  transitions.
 - [Storage Internals](storage.md): object-store interface and S3 behavior.
 - [Workspace Internals](workspace.md): local layout, locking, and cleanup
   guardrails.
@@ -108,6 +129,7 @@ trimmed transcript state: neither invalidates the other, while either can stale
 - [`extract`](stage-extract.md)
 - [`analyze`](stage-analyze.md)
 - [`publish`](stage-publish.md)
+- [`cleanup`](stage-cleanup.md)
 
 Use this map to find an owner, then read its focused documentation and tests
 before changing behavior.

docs/internal/stage-analyze.md

@@ -10,7 +10,9 @@ dependency order, and safely materialize validated outputs.
 - ordinary configured artifacts from `pipeline.scriptorium.artifacts`; canonical
   party artifact families have already expanded into this map during
   configuration resolution, including corresponding member dependencies and
-  rewritten member-artifact input sources
+  rewritten member-artifact input sources. Each artifact's dependency list is
+  already the normalized union of explicit ordering edges and configured-artifact
+  input edges.
 - optional selected artifact keys supplied through the stage environment
 - built-in, configured, extraction, and previous-session source references in
   artifact inputs
@@ -31,12 +33,14 @@ Supported source families:
 ## Key Behavior
 
 - when `pipeline.scriptorium` is absent or no configured artifact is
-  executable, completes successfully with no outputs and records explanatory
+  executable, completes successfully and records explanatory metadata. Before
-  metadata. This is not an explicit self-skip: both manifests record success,
+  returning, it reconciles and checkpoints the complete session collection:
-  satisfy publish's prerequisite, and an ordinary later run reuses the result
+  records removed from configuration are dropped, while valid records for
-  while the effective set remains empty. Enabling or selecting an artifact
+  configured but unselected artifacts remain authoritative. This is not an
-  later makes missing versioned evidence non-resumable and schedules it without
+  explicit self-skip: the session manifest records success, satisfies publish's
-  requiring force.
+  prerequisite, and an ordinary later run reuses the result while the effective
+  set remains empty. Enabling or selecting an artifact later makes missing
+  versioned evidence non-resumable and schedules it without requiring force.
 - builds a runtime artifact catalog containing built-ins, configured artifacts,
   and configured extraction lanes. Extraction availability is hydrated only
   from compatible successful extraction evidence.
@@ -49,6 +53,9 @@ Supported source families:
   manifest record and durable output pass the configured-artifact evidence
   contract; an incidental or stale canonical file is unavailable.
 - validates selected artifact dependency order (cycle-safe topo ordering).
+- consumes only the concrete dependency graph finalized by configuration;
+  planning, prerequisite closure, execution, reuse, invalidation, and semantic
+  evidence do not infer a second graph from inputs.
 - resolves required/optional inputs per artifact source definition into an
   ordered semantic identity. Each identity records the configured input name,
   canonical source ID, required policy, explicit presence, source contract,
@@ -75,7 +82,7 @@ Supported source families:
   validated current dependency-output identities, and sorted effective
   Scriptorium variables (including Narratio's sticky session variable).
 - provides read-only reconciliation that classifies each configured record as
-  current, stale, missing, failed, legacy, or otherwise non-resumable, and
+  current, stale, missing, failed, or otherwise non-resumable, and
   separately identifies manifest records removed from current configuration.
   A record is current only when its fingerprint version and value match and its
   configured output still passes manifest-authoritative evidence validation.
@@ -89,13 +96,14 @@ Supported source families:
   rebuilt when required, while unrelated disabled artifacts are excluded.
 - the work plan carries explicit targets, prerequisite-only work, deterministic
   execution and reuse lists, invalidated and removed records, and a cloned
-  projected record collection. Valid unrelated configured records survive the
+  reconciled starting collection. Valid unrelated configured records survive
-  projection, removed records are omitted, and legacy files never become
+  reconciliation, removed records are omitted, and incidental files never
-  current without regeneration.
+  become current without manifest-backed regeneration.
 - implements aggregate resume validation by running the same read-only catalog,
   fingerprint reconciliation, and work planner used by execution. A succeeded
   aggregate record is reusable exactly when the selected closure schedules no
-  artifact work; stale unrelated records do not block a partial selection.
+  artifact work and no removed configuration records require a checkpoint;
+  stale unrelated records do not block a partial selection.
 - exposes the typed artifact decision to `session plan`. Planning applies it to
   a cloned manifest after modeling earlier selected stage transitions, so
   aggregate run/skip and artifact execute/reuse decisions match the ordinary
@@ -104,6 +112,10 @@ Supported source families:
   prerequisites remain available through the runtime catalog without invoking
   Scriptorium; newly produced prerequisites enter that catalog with the same
   contract, checksum, and size identity used for persisted current evidence.
+- before the first adapter call, checkpoints the reconciled starting collection
+  when it removed configuration records or cleared obsolete authority. After
+  each successful artifact, checkpoints the complete collection before making
+  that artifact available to the next scheduled dependent.
 - keeps adapter output in the invocation's run-local analyze directory until
   it is a safe, non-empty, bounded regular file with a calculated checksum and
   complete output contract. Canonical replacement uses the shared atomic file
@@ -112,16 +124,15 @@ Supported source families:
 - records each successful artifact's freshly computed fingerprint, canonical
   relative output path, contract, checksum, size, producer run ID, bounded
   Scriptorium provenance, logs, and generated configuration references in the
-  analyze-owned projection.
+  analyze-owned manifest collection.
 - preserves valid unrelated current records during partial execution. If a
   rebuilt output's bytes and contract are unchanged, unselected dependents may
   remain current. If that semantic identity changes, unselected transitive
   dependents become stale without being executed; dependents included in the
   invocation are evaluated in dependency order instead.
-- reports all evaluated targets and prerequisites in invocation state. The
+- returns aggregate outputs from every current session record in normalized-key
-  runner reconstructs aggregate session outputs from every current session
+  order after the complete reconciled collection has been checkpointed;
-  record and invocation outputs from only records produced by the current run.
+  unrelated stale records do not make an otherwise successful partial
-  Unrelated stale records do not make an otherwise successful partial
   invocation fail.
 - resolves previous-session sources from local `previous/` cache only.
 - runs optional render-debug, then artifact execution.
@@ -137,25 +148,31 @@ Supported source families:
   guidance.
 - dependency cycles or unavailable required dependencies fail.
 - adapter validation failures fail stage.
-- a scheduled artifact failure returns the restricted analyze-state projection
+- a scheduled artifact failure installs a retry collection on the live manifest
   with the active artifact marked `failed`, a bounded error, and no output
   authority. Current transitive dependents become stale without execution.
-- earlier artifacts from the invocation remain current only after their
+- earlier artifacts remain current only after their run-local output passed
-  run-local output passed validation and canonical materialization. They remain
+  validation, canonical materialization, and a successful complete-collection
-  in invocation history; unattempted later artifacts do not appear there.
+  manifest checkpoint; unattempted later artifacts retain their reconciled
+  session state.
 - unrelated current records survive a partial failure. Old canonical bytes for
-  the failed artifact and newly materialized bytes whose projection cannot be
+  the failed artifact and newly materialized bytes whose checkpoint fails are
-  persisted are incidental, not current evidence.
+  incidental, not current evidence.
-- the runner persists a valid partial projection before it marks aggregate
+- a checkpoint failure stops the queue immediately. The in-memory retry
-  analyze failed and invalidates publish and notify through the application
+  collection starts from the last successful checkpoint, marks the active
-  dependency relation. Projection-persistence errors retain the last durable
+  artifact failed without output authority, and stales its current transitive
-  per-artifact authority and are joined with the original failure context.
+  dependents.
+- the runner's ordinary error path marks aggregate analyze failed, clears its
+  aggregate outputs, invalidates publish, notify, and cleanup through the
+  application dependency relation, and persists the live retry collection.
+  Terminal-save errors are joined with the original artifact or checkpoint
+  failure.
 
 ## Invariants
 
 - `analyze` performs no remote storage calls for previous-session source resolution.
 - input-identity resolution is read-only: it does not invoke adapters,
-  materialize outputs, update status, or create run records.
+  materialize outputs, update status, or create run-local files.
 - fingerprints exclude timeouts, retries, timestamps, producer and Narratio run
   IDs, executable and config paths, workspace roots, diagnostic locations, and
   executable or private transitive configuration contents. A change that is
@@ -164,6 +181,9 @@ Supported source families:
 - output provenance and metadata are deterministic per execution.
 - a canonical file without current per-artifact manifest evidence is never
   promoted to current state.
+- an ordinary retry after interruption or partial failure reuses every valid
+  earlier checkpoint and schedules only failed, stale, missing, or otherwise
+  non-current work in dependency order.
 
 ## Related Contracts And Tests
 
@@ -173,13 +193,18 @@ Supported source families:
 - [CLI](../cli.md) owns user-visible artifact selection.
 - [Scriptorium](../integrations/scriptorium.md) owns the subprocess contract.
 - Implementation and tests: `internal/stage/analyze.go`,
-  `internal/stage/analyze_input_identity.go`, `internal/stage/analyze_test.go`,
+  `internal/stage/analyze_artifact_execution.go`,
-  `internal/stage/analyze_input_identity_test.go`,
+  `internal/stage/analyze_execution_state.go`,
+  `internal/stage/analyze_input_identity.go`,
   `internal/stage/analyze_fingerprint.go`,
+  `internal/stage/analyze_reconciliation.go`, `internal/stage/analyze_plan.go`,
+  `internal/stage/analyze_resume.go`, `internal/stage/analyze_test.go`,
+  `internal/stage/analyze_checkpoint_test.go`,
+  `internal/stage/analyze_input_identity_test.go`,
   `internal/stage/analyze_fingerprint_test.go`,
-  `internal/stage/analyze_reconciliation.go`, and
   `internal/stage/analyze_reconciliation_test.go`,
-  `internal/stage/analyze_plan.go`, `internal/stage/analyze_plan_test.go`, and
+  `internal/stage/analyze_plan_test.go`,
-  `internal/stage/analyze_incremental_execution_test.go`, and
+  `internal/stage/analyze_incremental_execution_test.go`,
   `internal/stage/analyze_failure_test.go`,
-  `internal/stage/analyze_resume.go`, and `internal/stage/analyze_resume_test.go`
+  `internal/stage/analyze_resume_test.go`, and
+  `internal/app/analyze_checkpoint_recovery_test.go`.

docs/internal/stage-cleanup.md

@@ -0,0 +1,54 @@
+# Stage: cleanup
+
+## Purpose
+
+Remove explicitly configured local data for the run selected by a successful
+publish commit. Cleanup is the twelfth and final canonical stage, after
+`notify`.
+
+## Authorization And Targets
+
+Cleanup is disabled unless at least one of these policies is true:
+
+- `pipeline.spool.delete_audio_after_publish`
+- `pipeline.workspace.cleanup_after_publish`
+
+An enabled cleanup requires the manifest's publish stage to be succeeded with
+`uploaded=true`, a canonical `published_run_id`, and the exact remote commit and
+current commit-pointer keys for the manifest campaign/session prefix. The
+configured campaign and session must match the manifest. Cleanup derives its
+only possible targets through the canonical artifact path helpers:
+
+- the published run's session spool-audio directory; and
+- the published run's workspace run directory.
+
+It does not scan parent directories, infer the newest run, or read object
+storage. Every deletion uses the root-confined file operation, which rejects a
+root itself, traversal, symlink escape, and out-of-root targets.
+
+## Execution And Resume
+
+Disabled cleanup records an explicit skipped result. Enabled cleanup removes
+targets in deterministic policy order and treats an already-missing target as
+success. A successful record contains the published run ID and completed policy
+names.
+
+A previous cleanup success is reusable only while its semantic configuration
+matches and every selected target remains absent. A recreated target makes the
+stage non-resumable. A failure is recorded against cleanup; publish remains
+succeeded and remote current state is unchanged. On a later normal full run,
+valid upstream stages are skipped and cleanup is retried without republishing.
+
+`narratio publish` selects publish only and never invokes cleanup. A bounded run
+through `notify` or any earlier endpoint also excludes it. Operators may run
+`narratio run-stage cleanup <session_id>` directly; missing, failed, malformed,
+or identity-mismatched publication evidence fails before deletion.
+
+## Implementation And Tests
+
+- Planning and evidence validation: `internal/stage/cleanup_plan.go`
+- Execution and resume: `internal/stage/cleanup.go`
+- Root-confined deletion: `internal/fileops/cleanup.go`
+- Focused tests: `internal/stage/cleanup_plan_test.go`,
+  `internal/stage/cleanup_test.go`, and
+  `internal/app/cleanup_orchestration_test.go`

docs/internal/stage-extract.md

@@ -19,12 +19,17 @@ procedures belong in [Operations](../operations.md).
 1. resolves the final trimmed transcript from the shared artifact catalog;
 2. resolves every configured prepared reference through the shared
    manifest-authoritative identity resolver before creating run-local output;
-3. streams each verified reference into an invocation-local snapshot and
+3. resolves each reference to a stable content-addressed session snapshot,
-   rejects any source change observed while copying;
+   reusing verified bytes or reconstructing a missing/corrupt regular entry,
+   and rejects any source change observed while copying;
 4. fingerprints the byte- and provenance-bearing Notarius invocation evidence,
-   including sorted reference identities;
+   including sorted reference identities and the explicit downstream
+   configuration-profile selection;
 5. creates a run-local staging directory and invokes the injected
-   `notarius.Runner`;
+   `notarius.Runner`, carrying the optional already-resolved credential
+   reference without exposing its value, the optional exact configuration
+   profile, plus effective resume and chunk-cache controls after force
+   precedence;
 6. revalidates the reference snapshots, then validates the v2 successful
    receipt, confined index, management documents, configured required lane
    descriptors, validation summaries, and regular payload files;
@@ -41,13 +46,21 @@ Narratio run ID, the resolved trimmed-input identity, and invocation
 fingerprint. The input identity binds the exact transcript bytes, canonical
 source ID, producer stage/output/run identity, and resolution provenance.
 Reference metadata contains only selector, source ID, canonical session-relative
-path, checksum, and size; adapter requests receive selector and absolute
+path, checksum, and size; adapter requests receive selector and absolute stable
-invocation-local snapshot path, never payload contents. Snapshot bytes must
+snapshot path, never payload contents. Snapshot bytes must
 match the prepared identity both before and after Notarius runs, so a concurrent
 prepared-file replacement cannot make recorded provenance describe different
 bytes from those supplied to Notarius.
-Validation completes before
+Validation completes before the single durable rename, so a rejected result
-promotion, so a rejected result cannot expose a partial durable bundle.
+cannot expose a partial durable bundle.
+
+Snapshot destinations are owned by `internal/artifacts` under
+`cache/notarius/references/<sha256>/<prepared-basename>`. They contain no run ID
+or timestamp and can be reused across invocations. Identical bytes and basename
+share a verified file; changed bytes select a different digest directory.
+Unsafe path components and non-regular conflicts fail without recursive
+deletion. Snapshot paths are delivery details and never become manifest output
+evidence or prerequisites for reuse of an already complete extract bundle.
 
 Any executed extraction outcome that replaces a different effective outcome
 marks succeeded analysis and delivery dependents stale. Render is an independent
@@ -59,17 +72,28 @@ outputs is stable and does not repeatedly invalidate dependent stages.
 Before the focused validator runs, the application compares extract's versioned
 semantic fingerprint. It covers enablement, Notarius pipeline identity, sorted
 reference selector/source mappings, sorted declared output contracts, and each
-canonical `narratio.extraction.<key>` output identity. It excludes executable,
+canonical `narratio.extraction.<key>` output identity. It also covers the
+explicit Notarius configuration-profile selection; omitted, empty, and exact
+`default` selectors share the base-selection identity. It excludes executable,
 timeout, working directory, config path, and private Notarius config contents.
+Credential presence, reference name, and resolved value are likewise
+operational-only and excluded. Ordinary command loading still resolves a
+selected reference before resume evaluation, but a credential-only change does
+not rerun a valid completed extraction or stale its dependents.
 
 `internal/stage/extract_resume.go` then permits a skip only when the existing
 stage record still matches the current byte- and provenance-bearing invocation
 evidence. That evidence covers the current direct trimmed-transcript identity,
 sorted prepared-reference identities, pipeline identity, and configured output
-contracts. The same reference helper and transcript identity are resolved again
+contracts, plus the same downstream profile-selection identity. The same
-for artifact evidence, so changing current transcript bytes, reference bytes,
+reference helper and transcript identity are resolved again for artifact
-or producer identity makes the prior extraction obsolete. Operational runner
+evidence, so changing current transcript bytes, reference bytes, producer
-settings do not invalidate otherwise current durable evidence.
+identity, or a named downstream selection makes the prior extraction obsolete.
+Operational runner settings do not invalidate otherwise current durable
+evidence. This includes `resume` and `chunk_cache`; changing them alone cannot
+make a complete extraction non-resumable. If extraction executes, ordinary
+runs forward the configured values, while force disables resume and selects
+`refresh`.
 
 A valid prepared-reference change makes extraction non-resumable. Missing,
 unsafe, or checksum-inconsistent prepared evidence is a hard validation error
@@ -82,27 +106,29 @@ contracts and provenance, regular-file status, and stored checksums. Missing or
 obsolete results are non-resumable and run again; unsafe filesystem conditions
 return an error rather than silently accepting or replacing data.
 
-Neither contract can observe files imported by Notarius configuration, profile
+This sensitivity is limited to the explicit downstream selector. Narratio's
-contents, prompt/module definitions, or other transitive inputs. Operators must
+own profile label and whole effective-configuration digest are not blanket
-force extraction after changing any such private input behind a stable
+extraction invalidators. Neither contract can observe files imported by
-identifier.
+Notarius configuration, profile contents, prompt/module definitions, or other
+transitive inputs. Operators must force extraction after changing any such
+private input behind a stable identifier.
 
 ## Failure Behavior
 
 Adapter startup, timeout, nonzero exit, receipt decoding, path confinement,
 index compatibility, inconsistent warning or diagnostic envelopes,
 required-lane rejection or incomplete validation, payload inspection,
-checksum, or promotion errors fail the stage through ordinary manifest
+checksum, or installation errors fail the stage through ordinary manifest
 transition handling.
 Stdout receipt and stderr diagnostics remain separate. Downstream stages are
 not given selectable extraction sources unless the complete configured result
-has passed validation and promotion.
+has passed validation and installation.
 
 When a replacement attempt begins, the current session-stage record no longer
 advertises payload from the previous success. A failed replacement therefore
 has no current outputs, logs, generated configuration references, or metadata,
-while the earlier invocation manifest and immutable promoted bundle remain
+while the earlier immutable promoted bundle remains available for audit and
-available for audit and recovery.
+recovery.
 
 ## Implementation And Focused Tests
 
@@ -110,9 +136,13 @@ available for audit and recovery.
   `internal/stage/extract_resume.go`,
   `internal/stage/extract_test.go`,
   `internal/stage/semantic_contracts_delivery.go`
+- Stable snapshot materialization and verification:
+  `internal/stage/extract_references.go`
 - Subprocess boundary: `internal/adapters/notarius/subprocess.go`,
   `internal/adapters/notarius/subprocess_test.go`
 - Catalog hydration: `internal/artifacts/extraction_catalog.go`,
   `internal/artifacts/extraction_catalog_test.go`
 - Composition and downstream behavior: `internal/app/runner_test.go`,
   `internal/stage/analyze_test.go`, `internal/stage/publish_test.go`
+- Opt-in Notarius v0.7.0 production-boundary coverage:
+  `internal/stage/notarius_cache_integration_test.go`

docs/internal/stage-merge.md

@@ -17,7 +17,11 @@ Normalize raw transcript inputs and merge into base transcript via Seriatim.
 
 ## Key Behavior
 
-- discovers and validates raw transcript inputs.
+- discovers and validates raw transcript inputs. For versioned transcribe
+  state, discovery requires aggregate transcribe success and consumes only the
+  aggregate outputs that exactly match the current per-speaker collection.
+  Partial checkpoints and orphaned files are not merge inputs. Legacy/manual
+  workflows without versioned speaker state retain directory discovery.
 - normalizes each raw transcript (`seriatim.Normalize`) into run-local scratch output.
 - merges normalized inputs (`seriatim.Run`) into base transcript.
 - validates merged transcript and optional report JSON.
@@ -26,6 +30,8 @@ Normalize raw transcript inputs and merge into base transcript via Seriatim.
 ## Invariants
 
 - merge always consumes normalized forms of raw inputs.
+- a versioned transcribe collection is usable only through its complete,
+  successful aggregate output declaration.
 - base transcript must validate before stage success.
 - report output is config-gated.
 

docs/internal/stage-normalize.md

@@ -15,7 +15,7 @@ Normalize polished transcript into final transcript using Seriatim.
 
 ## Key Behavior
 
-- resolves polished transcript from manifest outputs/canonical fallback.
+- resolves polished transcript only from verified manifest output evidence.
 - applies `pipeline.normalize` config or default normalize config.
 - runs Seriatim normalize with configured timeout/binary.
 - validates normalized transcript and optional report.

docs/internal/stage-polish.md

@@ -16,7 +16,7 @@ Run Audita polishing on base transcript and produce polished transcript.
 
 ## Key Behavior
 
-- resolves base transcript from merge outputs/canonical fallback.
+- resolves base transcript only from verified merge manifest output evidence.
 - invokes an Audita runner configured with static model/runtime options; the
   invocation supplies paths and modules.
 - validates processed transcript structure (`segments` array required).
@@ -35,8 +35,9 @@ validation model, module set, transcript description, output schema, selected
 external configuration path, and canonical polished-transcript identity. Module
 ordering is normalized because the configured modules form a set. Audita's
 executable, timeouts, concurrency, report and debug behavior, work retention,
-and credential environment name are operational and do not invalidate a
+and resolved credential value are operational and do not invalidate a
-successful result.
+successful result. The public credential reference name may participate in
+configuration provenance but never exposes its resolved value.
 
 Narratio can fingerprint a selected model, module, or configuration identifier,
 but it cannot inspect content that Audita privately resolves behind that stable

docs/internal/stage-prepare.md

@@ -39,9 +39,8 @@ Materialize canonical current-session inputs before processing stages.
 - materializes a configured spell catalog with checksum and provenance, or
   safely removes an obsolete canonical spell catalog and its manifest record
   when the effective input is omitted.
-- in canonical party mode, copies the validated raw party bytes unchanged and
+- copies the validated canonical party bytes unchanged and deterministically
-  deterministically generates the prepared players projection; legacy mode
+  generates the prepared players projection.
-  continues to copy its opaque party and explicit players sources.
 - scans enabled configured artifact inputs for `narratio.previous_session.artifact.*` requirements.
 - clears managed `previous/` state on every invocation, then, when requirements exist:
   - resolves the pointer-selected previous source through the shared resolver;
@@ -49,8 +48,10 @@ Materialize canonical current-session inputs before processing stages.
   - records previous inputs in `manifest.inputs`.
 
 Required previous-session inputs fail when unavailable; optional missing inputs
-are typed skipped results. Committed sources use their exact source-to-destination
+are typed skipped results. Each requirement resolves only from its unique
-mapping, while the isolated legacy reader rejects ambiguous fallback matches.
+`published_output` declaration in the pointer-selected remote commit. Prepare
+strips that selected run's immutable `published/` prefix to derive the local
+cache path, then validates the committed checksum, size, and generation.
 
 ## Invariants
 

docs/internal/stage-publish.md

@@ -7,15 +7,16 @@ Upload run/session outputs to object storage and atomically advance remote curre
 ## Inputs
 
 - successful preceding stages from the [canonical stage set](overview.md#pipeline-stage-set)
-- invocation-scoped run files
+- declared run-local diagnostic files
 - resolved publish output rules
-- effective publish locks (static + remote merged lock set), revalidated at the
+- independent per-source remote lock objects read at planning start
-  remote commit boundary
+- the validated prior current commit when an initially locked source may carry
 - durable previous-session cache files when present
 
 ## Outputs
 
-- uploaded invocation record and selected publish outputs;
+- uploaded session manifest, declared run-local diagnostics, and selected
+  publish outputs;
 - uploaded durable previous-session cache files when present;
 - immutable run-scoped commit manifest; and
 - current commit pointer, written last.
@@ -25,14 +26,27 @@ Exact remote placement and the operator workflow belong in
 
 ## Key Behavior
 
+`publish.go` retains the transaction-owning `Run` path. Source discovery,
+manifest-derived archive metadata, confined opening, and checksum validation
+live in `publish_sources.go`; artifact-catalog matching, selection, required
+and optional rules, destination resolution, and collision inputs live in
+`publish_outputs.go`. Copying, hashing, immutable upload, reuse checks, and
+remote verification live in `publish_transfer.go`; committed-manifest,
+immutable-commit, and current-pointer encoding and writes live in
+`publish_commit.go`. `publish_output_plan.go` owns the deterministic, I/O-free
+five-outcome model and its secret-free metadata projection.
+
 - when publishing or run upload is disabled, completes successfully with no
   outputs and records explanatory metadata. This is not an explicit self-skip:
-  both manifests record success. Enablement and upload policy are fingerprinted,
+  the session manifest records success. Enablement and upload policy are
-  so changing either automatically makes the prior result non-resumable.
+  fingerprinted, so changing either automatically makes the prior result
+  non-resumable.
 - validates prerequisite stage success and object-store availability.
-- derives a deterministic run-archive allowlist from the validated run
+- derives a deterministic run-archive allowlist from the session manifest's
-  `manifest.json`: declared run-local outputs, logs, generated configs, and the
+  declared current-run outputs, logs, and generated configurations. References
-  manifest itself. Unlisted workspace files are not archive candidates.
+  outside the current run root, including canonical session files and prior-run
+  diagnostics, are not archive candidates. No distinguished diagnostic file is
+  read or uploaded.
 - opens each archive candidate beneath its archive root without following
   symlinked ancestors or leaf entries, verifies that it is a regular file and
   checks a declared checksum when present, then streams the opened descriptor.
@@ -41,26 +55,34 @@ Exact remote placement and the operator workflow belong in
 - resolves publish output sources through runtime artifact catalog and
   manifest-aware resolution. Configured Scriptorium outputs are publishable
   only from validated `current` per-artifact analyze evidence; an incidental
-  canonical file, legacy aggregate output, stale/failed/unselected record, or
+  canonical file, aggregate output, stale/failed/unselected record, or
-  mismatched path, size, or checksum remains unavailable. This does not change
+  mismatched path, size, or checksum remains unavailable. Built-in and
-  the explicit compatibility policies owned by built-in, extraction, or
+  extraction sources likewise require their own current manifest evidence.
-  previous-session sources.
 - publishes extraction lanes only through explicit configured output rules;
   neither run-local nor durable Notarius bundles are scanned or uploaded wholesale.
 - selected artifact filter applies to configured artifact sources only.
-- locked outputs are skipped intentionally (including required ones).
+- plans every configured output into exactly one source-sorted outcome:
-- optional missing outputs are skipped; required missing unlocked outputs fail.
+  `published`, `carried_locked`, `locked_unpublished`, `optional_missing`, or
+  `unselected`.
+- an initially locked selected output carries its exact prior
+  `published_output` declaration when one exists. The declaration may point to
+  an earlier run in the same session and retains its source, key, checksum,
+  size, and generation; the bytes are not copied.
+- verifies a carried object's exact key, size, and generation through object
+  metadata without downloading or opening the local output source.
+- an initially locked source with no prior declaration remains absent even
+  when the output rule is required. Optional missing unlocked outputs remain
+  absent; required missing unlocked outputs fail.
 - creates one complete immutable source-to-destination mapping before upload;
 - uploads and verifies every declared immutable object and the commit manifest;
-- updates `current/commit-pointer.json` exactly once, last; and
+- writes the immutable commit from the combined new and exact carried
-- does not write the legacy `current/manifest.json` or `current/run_id.txt` pair.
+  declarations.
-- rechecks remote lock state immediately before the pointer update. A newly
+- rereads every initially unlocked participating source after the immutable
-  committed lock aborts selection, leaving any uploaded immutable attempt
+  commit. A newly present lock aborts selection, leaving the uploaded immutable
-  unselected.
+  attempt unselected. Removing an initial lock may conservatively produce one
-- reads the mutable remote lock document through a direct limit-plus-one read
+  carry because the initial snapshot controls the plan.
-  capped by `MaxRemoteLockStoreBytes` (1 MiB), retaining the generation returned
+- updates `current/commit-pointer.json` exactly once, last, and writes no other
-  with the opened body for conditional updates. Oversized lock documents fail
+  mutable publication control record.
-  before YAML decoding; published artifact payloads do not use this limit.
 
 ## Metadata Signals
 
@@ -68,9 +90,8 @@ Includes counts/lists for:
 - run uploads
 - published output uploads
 - previous uploads
-- skipped optional outputs
+- one `publish_output_outcomes` collection containing all five output states,
-- skipped unselected outputs
+  exact declared keys, and key-derived origin runs where applicable
-- locked outputs
 - remote commit and current-pointer key paths
 - the run identifier selected by the commit
 
@@ -83,35 +104,44 @@ Includes counts/lists for:
   from verified descriptors; symlinks, special files, replacement races, and
   undeclared entries are rejected or ignored before uploads begin.
 - run-local diagnostics, including Notarius receipt and stderr files, are
-  archived only when recorded by the run manifest.
+  archived only when recorded by the session manifest. A run with no declared
-- publish locks are not overridden by `--force`; remote locks are revalidated
+  current-run diagnostics legitimately uploads none.
-  immediately before current-state selection.
+- independent publish locks are not overridden by `--force`; every initially
-- post-commit local cleanup is authorized by the committed publish metadata and
+  unlocked relevant source is reread immediately before current-state
-  is durably recorded by the application lifecycle before any local deletion.
+  selection.
+- successful uploaded publication records the exact run ID, remote commit key,
+  and current-pointer key consumed by the later cleanup stage.
 
 ## Resume Semantics
 
 The versioned semantic fingerprint covers enabled behavior, run-upload policy,
-normalized source/destination/required output rules, static lock policy, and
+normalized source/destination/required output rules, and the remote backend,
-the remote backend, bucket, region, endpoint, and root-prefix identity. Rule
+bucket, region, endpoint, and root-prefix identity. Rule ordering is
-and lock ordering is canonicalized. Credential environment names,
+canonicalized. Credential
-path-addressing transport mode, local workspace placement, and run identifiers
+environment names, path-addressing transport mode, local workspace placement,
-are excluded. Remote locks remain mutable state and are still revalidated at
+and run identifiers are excluded. Independent remote locks remain mutable state
-the commit boundary; semantic evidence does not replace that safety check.
+and are read directly by publish at planning and commit boundaries; semantic
+evidence does not replace those safety checks.
 
-The commit boundary and cleanup gate are normative architecture invariants; see
+The commit boundary and cleanup authorization are normative architecture invariants; see
 [Architecture](../policy/architecture.md#publish-commit-boundary).
 
+Publish never deletes local state. The standalone `publish` command selects
+only this stage; a full run selects notify and cleanup afterward.
+
 ## Related Contracts And Tests
 
-- [Configuration](../config.md#publish-configuration-summary) owns output and
+- [Configuration](../config.md#publish-configuration-summary) owns output
-  static-lock fields.
+  rules.
 - [Operations](../operations.md#publish-locks) owns remote lock lifecycle and
   physical remote state.
 - [Artifact Internals](artifacts.md) explains source resolution and current-state
   helpers.
 - Implementation and tests: `internal/stage/publish.go`,
+  `internal/stage/publish_sources.go`, `internal/stage/publish_outputs.go`,
+  `internal/stage/publish_output_plan.go`,
+  `internal/stage/publish_transfer.go`, `internal/stage/publish_commit.go`,
   `internal/stage/publish_test.go`,
   `internal/stage/semantic_contracts_delivery.go`,
   `internal/app/operator_helpers_test.go`, and
-  `internal/app/post_publish_cleanup_test.go`
+  `internal/app/cleanup_orchestration_test.go`

docs/internal/stage-render.md

@@ -21,7 +21,7 @@ records that may consume rendered transcripts.
 ## Key Behavior
 
 - uses `pipeline.render` settings (enabled/format/title/booleans).
-- resolves inputs manifest-first, then canonical fallback.
+- resolves inputs only from verified manifest output evidence.
 - writes run-local outputs first, then materializes canonical session outputs.
 - records input provenance, output paths, adapter metadata, logs, and generated config refs.
 - when `pipeline.render.enabled=false`, completes successfully with no outputs

docs/internal/stage-transcribe.md

@@ -14,12 +14,19 @@ Generate raw per-speaker transcripts from prepared audio using WhisperX.
 
 ## Key Behavior
 
-- discovers prepared audio from manifest inputs or canonical audio directory.
+- discovers prepared audio from manifest inputs or the canonical audio
-- derives the transcript identity from the prepared `.flac` filename.
+  directory and derives each speaker identity from its `.flac` filename.
-- dispatches WhisperX requests through a bounded worker pool.
+- hashes the current audio and reconciles it with the versioned per-speaker
-- validates each output as JSON.
+  completion collection before starting HTTP work.
-- writes run-local outputs then materializes canonical transcript outputs only
+- reuses only records whose audio identity, observable WhisperX semantics,
-  after every planned request succeeds.
+  canonical output location, JSON, checksum, and size remain current.
+- durably checkpoints the reconciled collection before dispatch, then sends
+  only missing or obsolete speakers through the bounded worker group.
+- validates each successful run-local response, rechecks the audio bytes,
+  materializes the canonical raw transcript, and checkpoints that speaker
+  before accepting another completion.
+- rebuilds aggregate stage outputs in stable speaker order from the complete
+  current collection, including records produced by earlier invocations.
 
 ## Invariants
 
@@ -28,21 +35,48 @@ Generate raw per-speaker transcripts from prepared audio using WhisperX.
 - output path returned by adapter must match requested output path.
 - an empty adapter result path means the requested path; adapters cannot select
   an alternate destination.
-- each successful output is validated before stage success, and cancellation or
+- candidate collection construction, manifest save, and installation into live
-  incomplete dispatch cannot be reported as a successful result.
+  state are serialized. Expensive hashing and HTTP work remain outside that
+  critical section.
+- a canonical file has no completion authority until its manifest checkpoint
+  succeeds. A save failure preserves the last successfully committed
+  collection for ordinary retry.
+- each successful output is validated before its checkpoint, and cancellation,
+  incomplete dispatch, or a failed aggregate save cannot be reported as
+  aggregate stage success.
+- removed audio inputs are omitted from the current collection and aggregate
+  output list; reconciliation does not delete arbitrary transcript files.
+- merge cannot consume a partial versioned collection: it requires transcribe
+  aggregate success and an exact match between current speaker records and
+  declared aggregate outputs.
 
 ## Resume Evidence
 
-Transcribe records a versioned semantic-configuration fingerprint containing
+Each completed speaker record binds the speaker ID, session-relative audio
-the Narratio-visible WhisperX service URL and recognition language. Changes to
+path, streamed audio checksum and size, the Narratio-visible WhisperX service
-either rerun transcription and stale its fixed descendants while leaving
+URL and recognition language, and the canonical transcript path, checksum,
-prepare reusable. Retry count/delay, concurrency, timeout, credentials, and
+size, and producer run. Changed audio bytes select that speaker for replacement;
-diagnostic locations are operational and do not change this evidence.
+a changed URL or language selects every speaker. Either case stales fixed
+descendants through the normal runner decision. Retry count/delay, concurrency,
+timeout, credentials, and diagnostic locations are operational and do not
+change this evidence.
+
+An ordinary retry reuses every valid speaker checkpoint even when aggregate
+transcribe status is failed, running, or stale. A succeeded stage is skipped only
+when its stored speaker set and aggregate output evidence also match the current
+collection. Otherwise, execution repairs those records without retranscribing
+reusable speakers and stales dependents through the normal runner path.
+`--force` first durably replaces
+the collection with an empty versioned collection, so a later ordinary retry
+cannot resurrect pre-force results; speakers checkpointed by that forced
+attempt remain reusable.
 
 WhisperX models or private service configuration not exposed by Narratio's
 adapter contract cannot be fingerprinted; use `--force` after changing them.
-An existing successful transcribe record without evidence reruns once when
+An older aggregate-only transcribe record has no trustworthy audio-to-output
-selected.
+binding. When transcribe is next selected, every current speaker is therefore
+generated once and recorded in the versioned collection; later invocations can
+reuse it normally.
 
 ## Related Contracts And Tests
 
@@ -51,6 +85,9 @@ selected.
 - [Configuration](../config.md#pipeline) owns concurrency and other
   operator-selected values.
 - Implementation and tests: `internal/stage/transcribe.go`,
+  `internal/stage/transcribe_reconciliation.go`,
+  `internal/stage/transcribe_execution_state.go`,
+  `internal/stage/transcribe_resume.go`,
   `internal/stage/transcribe_test.go`,
-  `internal/stage/semantic_contracts_initial.go`, and
+  `internal/stage/transcribe_reconciliation_test.go`, and
-  `internal/stage/semantic_contracts_initial_test.go`
+  `internal/stage/transcribe_checkpoint_test.go`

docs/internal/storage.md

@@ -16,6 +16,7 @@ operator-selected storage fields and credential mechanisms belong in
 - `Download(ctx, key, localPath)`
 - `Upload(ctx, localPath, key, opts)`
 - `UploadConditional(ctx, source, key, opts, condition)`
+- `Delete(ctx, key)` removes one object and succeeds when it is already absent
 - `Exists(ctx, key)`
 
 Key invariant:
@@ -47,6 +48,8 @@ not own discovery, defaults, or configuration validation.
 - `Read` binds a returned body to its S3 ETag. `UploadConditional` maps an ETag
   match or absence precondition directly to the provider request and reports a
   failed precondition without performing a local check-then-write replacement.
+- `Delete` maps one normalized key to the provider delete-object operation and
+  treats an absent object as success.
 - `Exists` maps not-found responses to `false`.
 
 ## Invariants
@@ -55,6 +58,11 @@ not own discovery, defaults, or configuration validation.
 - bounded reads never retain more than the caller's limit plus one byte and do
   not replace owner-specific size policy.
 - publish ordering semantics are owned by stage/app code, not storage adapters.
+- A caller that installs remote content into a canonical workspace path first
+  writes a sibling temporary file, validates the complete object against its
+  owner-specific contract, and only then replaces the destination. Temporary
+  files used solely for remote comparisons are removed by their callers and do
+  not represent installed workspace state.
 
 ## Implementation And Tests
 

docs/internal/workspace.md

@@ -14,8 +14,8 @@ path model, and layout creation goes through `EnsureLayoutFor`. Callers should
 consume those helpers instead of rebuilding relative paths.
 
 `internal/pathsafe` validates relative destinations. `internal/fileops` opens
-cleanup roots and their descendants through no-follow directory handles before
+explicit cleanup roots and removes only normalized proper descendants through
-removing them.
+the resulting root handle.
 
 `internal/fileops` owns the ordinary workspace mode contract. On POSIX,
 `WorkspaceDirectoryMode` is setgid `02775` and `WorkspaceFileMode` is `0664`.
@@ -33,15 +33,16 @@ previous-session cache paths remain session-durable and are never redirected
 into run-local output space.
 
 Extraction uses run-local receipt, stderr, and output-root helpers, then
-promotes the validated external bundle to the unique immutable Notarius bundle
+validates and renames the external bundle to the unique immutable Notarius
-path supplied by `internal/artifacts`. `internal/fileops.PromoteDirectory`
+bundle path supplied by `internal/artifacts`. Exact physical paths belong in
-copies only regular files and directories to a same-filesystem temporary
-sibling. Source traversal uses confined directory handles and identity checks
-so replacing an inspected root, directory, or file is rejected rather than
-followed. The completed tree is atomically renamed without replacing an
-existing destination. Exact physical paths belong in
 [Operations](../operations.md#extraction-workflow).
 
+Prepared references delivered to Notarius use session-durable paths owned by
+`internal/artifacts` beneath
+`cache/notarius/references/<sha256>/<prepared-basename>`. These verified
+content-addressed snapshots are outside run-local layout, survive ordinary
+published-run cleanup, and are excluded from publish and restore enumeration.
+
 ## Locking
 
 `artifacts.LocalStore` enforces the single-writer session lock via an
@@ -50,21 +51,25 @@ file retains owner metadata after release or process death; its existence is
 not evidence that a lock is active. Command and restore flows wait for this
 lock only while their context remains active, and report a release failure.
 
+The runner assigns every invocation a run-workspace identity but creates the
+physical directory only before an executing stage that requires it. Fully
+reused invocations and cleanup-only invocations therefore do not accumulate
+empty run directories.
+
 ## Cleanup Semantics
 
-Automatic post-publish cleanup:
+The canonical cleanup stage:
 
-- is created only after a successful publish commit with complete publish
+- runs after notify only when selected by the invocation range;
-  metadata, then is persisted before any deletion;
+- requires a succeeded publish record with `uploaded=true`, the published run
-- requires `uploaded=true`, a remote commit key, and a current commit-pointer
+  ID, remote commit key, and current commit-pointer key;
-  key in publish metadata;
 - consumes the resolved cleanup policy described in
   [Configuration](../config.md);
-- refuses unsafe deletes (root delete, out-of-root delete, and symlinked
+- refuses unsafe deletes (root delete, traversal, and out-of-root targets);
-  ancestors or entries);
+- treats missing targets as successful, idempotent deletion; and
-- retries any recorded incomplete target on later invocations even when no
+- uses its ordinary failed/succeeded manifest stage record for retry. A later
-  publish work is selected. Missing targets are a successful, idempotent
+  full run can skip reusable upstream stages and retry cleanup without
-  cleanup result only after the completion evidence is saved.
+  republishing.
 
 Manual cleanup uses the same root-confined deletion mechanism. Invocation
 syntax and exact deletion scope belong in [CLI](../cli.md#clean) and
@@ -75,6 +80,8 @@ syntax and exact deletion scope belong in [CLI](../cli.md#clean) and
 - campaign-aware session root is mandatory.
 - manifest-driven stage state is durable across runs.
 - cleanup guardrails prevent destructive root/out-of-scope deletion.
+- restore installation is confined to validated relative paths beneath the
+  selected session root.
 - ordinary workspace paths retain group-writable directory and file modes across
   nested creation, replacement, and Notarius promotion.
 
@@ -83,12 +90,14 @@ syntax and exact deletion scope belong in [CLI](../cli.md#clean) and
 - Path model and local store: `internal/artifacts/paths.go`,
   `internal/artifacts/local.go`
 - Run-local materialization: `internal/stage/run_local.go`
-- Immutable bundle promotion: `internal/fileops/directory.go`
+- Immutable bundle installation: `internal/stage/extract.go`
 - Workspace modes: `internal/fileops/modes.go`
 - Cleanup confinement: `internal/fileops/cleanup.go`,
-  `internal/app/cleanup_targets.go`, `internal/app/post_publish_cleanup.go`
+  `internal/stage/cleanup.go`, `internal/stage/cleanup_plan.go`, and
+  `internal/app/cleanup_targets.go`
 - Tests: `internal/artifacts/paths_model_test.go`,
   `internal/artifacts/local_test.go`, `internal/stage/run_local_test.go`,
-  `internal/fileops/directory_test.go`, `internal/fileops/modes_posix_test.go`,
+  `internal/fileops/modes_posix_test.go`,
   `internal/fileops/cleanup_test.go`, `internal/app/cleanup_targets_test.go`,
-  `internal/app/post_publish_cleanup_test.go`
+  `internal/stage/cleanup_test.go`, and
+  `internal/app/cleanup_orchestration_test.go`

docs/operations.md

@@ -76,12 +76,15 @@ value from the last value persisted in the session manifest, which helps
 diagnose profile switches without changing resume authority.
 
 Before switching an operational profile, compare its effective meaning with the
-current selection through `narratio config diff <left-profile> <right-profile>`.
+production base through `narratio config diff default <profile>`.
 The command is read-only and succeeds whether it finds differences or not. Its
 sorted records describe defaulted, expanded concrete configuration—not source
 file layout—so it can be used to review model, artifact, and publish changes
-without creating a session or run. Select the same campaign explicitly when
+without creating a session or run. Use `narratio config artifacts` to review the
-profiles could resolve different campaign paths; see the [CLI reference](cli.md#config-validate-config-show-config-sources-and-config-diff) for syntax and record format.
+concrete artifact identities, definitions, outputs, and dependencies. Select the
+same campaign explicitly when profiles could resolve different campaign paths;
+see the [CLI reference](cli.md#config-validate-config-show-config-sources-config-artifacts-and-config-diff)
+for syntax and record format.
 
 ## Stage Execution and Continuation Behavior
 
@@ -98,12 +101,12 @@ Canonical stage order:
 9. `analyze`
 10. `publish`
 11. `notify`
+12. `cleanup`
 
 Execution rules:
 
 - succeeded stages are skipped unless `--force` is set; stages with semantic
-  configuration contracts additionally require matching versioned evidence,
+  configuration contracts additionally require matching versioned evidence;
-  and missing legacy evidence causes a safe one-time rerun;
 - `run` continues interrupted or partially completed sessions by running non-succeeded stages;
 - forcing a stage marks succeeded transitive dependents as `stale` before the
   replacement runs; render and extract are independent siblings; and
@@ -113,8 +116,8 @@ Execution rules:
   perpetually rerun dependent work.
 
 Every aggregate stage except analyze currently provides semantic-configuration
-evidence; analyze retains its more precise per-artifact fingerprints and
+evidence; transcribe additionally reconciles per-speaker completion evidence,
-validator.
+while analyze retains its more precise per-artifact fingerprints and validator.
 Prepare additionally validates current stable/local/S3 source identity and the
 checksums of its durable prepared copies before reuse. Changed bytes, audio
 membership, S3 object identity, or missing/tampered copies rerun prepare and
@@ -155,7 +158,7 @@ narratio run 2026-04-04 --from extract --through analyze --force
 ```
 
 Omitting `--from` selects from `prepare`; omitting `--through` selects through
-`notify`. Force applies only within the selected range. Repeating `--from`,
+`cleanup`. Force applies only within the selected range. Repeating `--from`,
 `--through`, or `--force` is rejected instead of resolving by argument order.
 The plan command uses the same selection contract and prints only the selected
 range. Planning is read-only: it clones the loaded manifest, models selected
@@ -169,11 +172,11 @@ selected artifact evidence is still semantically current.
 Before a bounded run or plan whose range starts after `prepare`, every excluded
 prefix stage must already have a session-manifest status of `succeeded` or
 `skipped`. Narratio reports the first absent, pending, running, failed, stale,
-or interrupted prerequisite without creating a run record or changing session
+or interrupted prerequisite without creating a run-local directory or changing
-state. Widen `--from` to include that stage, or recover it explicitly before
+session state. Widen `--from` to include that stage, or recover it explicitly
-retrying. Excluded prefix stages are not resume-validated or repaired as part
+before retrying. Excluded prefix stages are not resume-validated or repaired as
-of the bounded invocation; selected stages still reject missing, unsafe, or
+part of the bounded invocation; selected stages still reject missing, unsafe,
-manifest-inconsistent inputs at their owning boundary.
+or manifest-inconsistent inputs at their owning boundary.
 
 Stages after `--through` are not prerequisites and are never scheduled by the
 bounded invocation. A selected forced stage can mark one of those succeeded
@@ -184,7 +187,68 @@ shared session lifecycle. In particular, render does not require Notarius or
 Scriptorium, extract does not require Scriptorium, and analyze does not require
 the transcription, Seriatim, Audita, or Notarius adapters.
 
-For the common post-transcript development loop, use:
+### Transcription continuation and replacement
+
+Transcribe checkpoints each validated speaker independently. If a later
+speaker exhausts HTTP retries, the process is canceled, or the aggregate stage
+save fails, the aggregate stage is not successful and downstream merge cannot
+consume the partial set. The last successfully saved speaker records and their
+canonical `transcripts/raw/<speaker>.json` files remain durable, however.
+
+After correcting the underlying problem, retry normally:
+
+```bash
+narratio run-stage transcribe 2026-04-04
+```
+
+Narratio hashes the current prepared audio, validates recorded transcript JSON,
+checksum, size, and location, and calls WhisperX only for missing or obsolete
+speakers. A changed audio file replaces only its speaker; a changed WhisperX
+URL or language makes every current speaker non-reusable. Retry, timeout, and
+concurrency tuning do not invalidate completed speakers.
+
+Use force only when every current transcript must be replaced, including after
+a WhisperX model or private server-side configuration changes behind the same
+URL and language:
+
+```bash
+narratio run-stage transcribe 2026-04-04 --force
+```
+
+Force durably clears the reusable collection before sending requests. If that
+forced attempt fails, a later ordinary retry reuses only replacement speakers
+checkpointed after the reset; it does not revive pre-force results. Do not edit
+speaker checksums or copy older JSON into the canonical directory to influence
+reuse.
+
+The first current Narratio invocation that selects transcribe for an older
+aggregate-only manifest retranscribes every current speaker once because the
+old record cannot prove which audio bytes produced each file. Subsequent runs
+reuse the new versioned evidence normally. The overall session-manifest schema
+identifier is unchanged, but older strict-decoding Narratio binaries may reject
+a manifest after current Narratio writes the additive speaker fields; retain a
+backup or complete the upgrade before relying on downgrade.
+
+Restoring a manifest does not implicitly publish or restore every raw speaker
+file. If a recorded canonical transcript is absent locally, ordinary
+reconciliation schedules that speaker again. Removing run directories does not
+remove canonical checkpointed transcripts and does not by itself invalidate
+them.
+
+## Artifact Development Workflows
+
+Choose the narrowest workflow that matches the intended replacement:
+
+| Intent | Command | Preserved and rebuilt state |
+| --- | --- | --- |
+| Rebuild analysis only | [`narratio analyze 2026-04-04`](./cli.md#analyze) | Preserves transcripts and the current extraction bundle; force-runs selected analysis artifacts and their required analysis prerequisites. |
+| Rebuild extraction and analysis | [`narratio regenerate-artifacts 2026-04-04`](./cli.md#regenerate-artifacts) | Preserves transcripts; force-runs the complete configured extraction bundle, then selected analysis. `--artifacts` narrows analysis only. |
+| Continue bounded work normally | [`narratio run 2026-04-04 --from extract --through analyze`](./cli.md#run) | Uses ordinary resume decisions: current extraction and analysis can be skipped, while incomplete or obsolete work runs. |
+
+All three stop before publication. Inspect the results and run
+[`narratio publish`](./cli.md#publish) separately when delivery is intended.
+
+For the extraction-plus-analysis development loop, use:
 
 ```bash
 narratio regenerate-artifacts 2026-04-04
@@ -195,8 +259,8 @@ This command is a transparent expansion to a forced bounded `run` from
 `extract` through `analyze`. Extraction always rebuilds its complete configured
 bundle. Analysis rebuilds the selected targets and their required analysis
 prerequisites, or uses the normal default selection when no artifact names are
-given. The command does not run publish or notify; delivery remains a separate
+given. The command does not run publish, notify, or cleanup; delivery and
-operator action.
+retention remain separate operator actions.
 
 Inspect current artifact evidence, then publish explicitly when the regenerated
 set is ready:
@@ -206,11 +270,41 @@ narratio session artifacts 2026-04-04
 narratio publish 2026-04-04
 ```
 
-If planning or execution reports stale, missing, failed, legacy, or tampered
+If planning or execution reports stale, missing, failed, or tampered
 analysis evidence, regenerate the affected target instead of copying an older
 canonical file into place or editing the manifest. See
 [Troubleshooting: Analysis artifact evidence is not current](./troubleshooting.md#analysis-artifact-evidence-is-not-current).
 
+Analyze checkpoints each completed configured artifact before starting later
+artifact work. After a partial failure or interruption, retry normally: valid
+earlier checkpoints are reused and execution resumes with failed, stale, or
+missing work in dependency order. `--force` is unnecessary unless the operator
+intends to replace otherwise-current artifacts.
+
+### Developing With A Notarius Configuration Profile
+
+A Narratio development overlay can explicitly select a Notarius overlay without
+requiring the names to match. For example, `profiles/testing.yml` may contain:
+
+```yaml
+notarius:
+  config_profile: downstream-economy
+```
+
+Selecting Narratio's `testing` profile then supplies `downstream-economy` only
+when extraction executes. Switching that explicit selector makes a recorded
+extraction non-current and reruns it when `extract` is in the selected stage
+range; the usual dependent invalidation then applies. Selecting a profile does
+not create a separate workspace or retain a separate current extraction result.
+
+Narratio cannot observe edits inside Notarius configuration behind an unchanged
+selector. Force extraction after such an edit when a replacement is intended.
+Use `regenerate-artifacts` when both extraction and analysis must be rebuilt;
+use `analyze` when the existing extraction bundle should be preserved. Field
+semantics belong in [Configuration](./config.md#notarius-configuration-profile),
+and the exact downstream invocation belongs in the
+[Notarius integration contract](./integrations/notarius.md#invocation).
+
 ## Artifact Selection
 
 For a configured artifact family, selecting its family key expands to every
@@ -251,9 +345,14 @@ are not selectable or published implicitly.
 Configured Notarius references resolve only from the current manifest-backed
 prepared inputs. Their canonical locations are `inputs/party.yml`,
 `inputs/players.yml`, `inputs/glossary.yml`, and, when configured,
-`inputs/spell_catalog.json`. Extraction supplies Notarius with verified copies
+`inputs/spell_catalog.json`. Extraction supplies Notarius with verified,
-under `runs/<run_id>/extract/references/` so a concurrent refresh of canonical
+content-addressed copies under
-prepared files cannot change the bytes consumed by an in-flight invocation.
+`cache/notarius/references/<sha256>/<prepared-basename>`. The full lowercase
+SHA-256 digest identifies the reference bytes, so retries and later invocations
+reuse the same path while changed bytes select a new path. Older snapshots are
+retained and are not current artifact evidence. A concurrent refresh of
+canonical prepared files therefore cannot change the bytes consumed by an
+in-flight invocation.
 For a canonical party, preparation retains the validated authored party bytes
 at `inputs/party.yml` and generates `inputs/players.yml` from that roster.
 The manifest records their checksums separately, with the players input marked
@@ -280,15 +379,35 @@ narratio run-stage prepare 2026-04-04 --force
 
 Starting a replacement clears the previous extraction payload from the current
 session-stage record. If that replacement fails or self-skips, the current
-record does not fall back to the earlier outputs. The earlier run manifest and
+record does not fall back to the earlier outputs. The earlier immutable bundle
-immutable bundle remain available for inspection, but downstream resolution
+remains available for inspection, but downstream resolution requires a new
-requires a new current successful extraction record.
+current successful extraction record.
 
-Atomic Notarius bundle promotion is supported on Linux and macOS. On Windows
+Notarius extraction validates its run-local bundle and moves it once to the
-and other operating systems, extraction fails before copying the bundle into a
+unique durable bundle path. A pre-existing durable destination is a collision;
-temporary promotion tree because Narratio has no verified atomic no-replace
+no duplicate promotion tree is created.
-directory primitive there. This is an extraction limitation, not a broader
+
-platform-support guarantee for every Narratio workflow.
+`pipeline.notarius.resume` and `pipeline.notarius.chunk_cache` control work
+inside an executed Notarius invocation. Ordinary extraction forwards their
+configured values. Forced extraction always disables checkpoint loading and
+requests chunk-plan `refresh`, even when the configured mode is `bypass`.
+Refresh regenerates and replaces only the active input's accepted chunk plan;
+it is not cache deletion.
+
+Notarius checkpoint recording must be enabled in the deployed Notarius
+configuration. Narratio neither creates a Notarius pipeline/configuration nor
+reads it to infer checkpoint availability. Configure explicit persistent
+Notarius chunk-plan and checkpoint roots when reuse must survive separate
+Narratio subprocesses. A changed reference, transcript, selected profile,
+profile content, endpoint, or other Notarius checkpoint identity input can
+produce a legitimate cold miss.
+
+There are two independent resume boundaries: Narratio may skip a completed
+`extract` stage without invoking Notarius, while `--resume` allows Notarius to
+reuse compatible accepted work only after Narratio has decided extraction must
+execute. Deleting Narratio's local reference snapshots does not invalidate a
+complete durable extraction; missing snapshots are reconstructed only when a
+later extraction actually runs.
 
 ## External Command Lifecycle
 
@@ -311,12 +430,15 @@ Run-local diagnostics are:
 - `runs/{run_id}/extract/notarius.stderr.log`
 - `runs/{run_id}/extract/notarius-output/` before durable promotion
 
-The run-record upload is an allowlist derived from the validated run manifest,
+The run upload is an allowlist derived from the session manifest's declared
-not a workspace scan. Each declared source is opened without following
+current-run outputs, logs, and generated configurations, not a workspace scan.
-symlinked ancestors or the leaf, verified as a regular file, and streamed from
+References to canonical session files and prior-run paths are excluded. Each
-that verified descriptor. Unlisted files and unsafe entries are never uploaded.
+declared source is opened without following symlinked ancestors or the leaf,
-The durable bundle is never scanned for implicit publication; only lanes named
+verified as a regular file, and streamed from that verified descriptor.
-by explicit `pipeline.publish.outputs` rules are uploaded.
+Unlisted files and unsafe entries are never uploaded; a run with no declared
+current-run diagnostics uploads no run-local files. No distinguished diagnostic
+file is uploaded. The durable bundle is never scanned for implicit publication;
+only lanes named by explicit `pipeline.publish.outputs` rules are uploaded.
 
 To intentionally replace the current extraction result, run:
 
@@ -327,22 +449,23 @@ narratio run-stage extract 2026-04-04 --force
 Narratio automatically reruns extraction when its recorded invocation contract,
 prepared Narratio reference identities, or durable output validation changes.
 The semantic portion covers Notarius enablement, pipeline identity, declared
-reference mapping, and output contracts. Executable, timeout, working directory,
+reference mapping, output contracts, and the explicit Notarius configuration
-and private config-file paths are operational and do not invalidate a current
+profile selector. Executable, timeout, working directory, and private
-result.
+config-file paths are operational and do not invalidate a current result.
-It cannot fingerprint configuration files, profiles, prompts, modules, or
+It cannot fingerprint configuration or profile contents, prompts, modules, or
 other references loaded transitively by Notarius itself. Force extraction after
-changing any of those inputs, even when the top-level Narratio and Notarius
+changing any of those inputs behind an unchanged selector, even when the
-config paths remain the same. A forced extract
+top-level Narratio and Notarius config paths remain the same. A forced extract
 marks successful downstream stages stale. Ordinary extraction failures or
 outcome changes also stale affected downstream stages, while an identical
 repeated `notarius_disabled` self-skip does not repeatedly invalidate them.
 
 Publish reuse additionally tracks enabled/run-upload behavior, normalized
-output rules, static locks, and remote backend/bucket/region/endpoint/root
+output rules, and remote backend/bucket/region/endpoint/root identity.
-identity. Credential environment names, local workspace placement, and run IDs
+Credential environment names, local workspace placement, and run IDs are
-are excluded. Regardless of semantic reuse evidence, executing publish still
+excluded. Regardless of semantic reuse evidence, executing publish reads
-revalidates mutable remote locks immediately before commit selection.
+independent remote locks directly and rechecks initially unlocked sources
+before commit selection.
 
 ## Publish Workflow
 
@@ -358,6 +481,9 @@ Equivalent:
 narratio run-stage publish 2026-04-04 --force
 ```
 
+This standalone command stops after publish and does not invoke cleanup. A
+default full `run` continues through notify and cleanup.
+
 Publish commit model:
 
 - uploads eligible run files under `{session_prefix}/runs/{run_id}/`, excluding
@@ -381,31 +507,44 @@ commit by version, checksum, size, and storage generation; committed artifacts
 are also checksum- and generation-bound. Readers accept this contract now and
 strictly reject mismatched or unknown data.
 
-Legacy reads are limited to a coherent `current/manifest.json` and
+A selected commit may carry a `published_output` declaration from an earlier
-`current/run_id.txt` pair; a torn pair is rejected. New publication does not
+run in the same campaign/session. Restore, previous-session caching, status,
-write that pair and remote commit state does not carry local
+and artifact inspection use that declaration's exact immutable key and stable
-`current_pointer_written` metadata.
+source; they do not reconstruct a current-run key or scan the session prefix.
+Other artifact types remain confined to the selected commit's own run.
+
+When the pointer is absent, remote state is not published; republish the
+session to establish a committed snapshot. Remote commit state does not carry
+local `current_pointer_written` metadata.
 
 ## Publish Locks
 
-Lock sources:
+Each mutable lock is an independent strict JSON object at
+`{session_prefix}/locks/{sha256(source)}.lock`. The lowercase digest is computed
+from the exact canonical source. Listing validates every object in that
+reserved prefix and reports locks sorted by source.
 
-- static locks in `pipeline.publish.locks`
+Upgrades do not read, migrate, or delete the retired `{session_prefix}/locks.yml`
-- mutable remote locks in `{session_prefix}/locks.yml`
+object. It has no effect on publication and may be removed manually after
+verifying lock commands and publication behavior with the upgraded Narratio
+version. No migration command is required.
 
-Effective lock rules:
+Operational rules:
 
-- static and remote locks are merged;
+- add uses conditional creation and reports an existing lock;
-- static locks win on source collisions;
+- add with `--force` observes the current generation and attempts one
-- locked outputs are intentional skips;
+  generation-matched replacement. A concurrent change is reported for an
-- lock add/remove commands mutate only remote lock state through generation-bound
+  explicit operator retry;
-  conditional writes. A command retries a bounded number of concurrent
+- remove is idempotent and reports whether a lock was present;
-  conflicts while its invocation context remains active, so it never replaces a
+- the initial publish lock snapshot controls output planning. A locked source
-  different lock-document generation; and
+  carries its exact declaration from the prior current commit when available;
-- a publish re-reads remote locks immediately before it writes the current
+  a never-published locked source remains absent even when required;
-  commit pointer. A lock committed before that recheck prevents selecting the
+- carried bytes are not copied. Publish verifies the prior object's exact key,
-  new snapshot, even though its already-uploaded immutable objects may remain
+  size, and generation without downloading it; and
-  available for a later retry.
+- after immutable uploads and the immutable commit, publish re-reads every
+  initially unlocked participating source. A newly added lock prevents the
+  current pointer from advancing, although the immutable attempt may remain.
+  A removed initial lock may conservatively cause one carry.
 
 Examples:
 
@@ -439,13 +578,22 @@ Default restore scope:
 
 - the committed session manifest and the committed transcript/artifact objects
   declared by the selected remote commit
-- `previous/**` when needed by configured previous-session artifact inputs
+- `previous/**` when needed by configured previous-session artifact inputs;
+  each is selected from its exact committed `published_output` declaration
 
 Optional:
 
 - `--include-audio` to include `audio/**`
 - `--force` to overwrite eligible conflicting regular files; it never replaces
-  directories or other non-regular local targets
+  directories, symlinks, or other non-regular local targets
+
+Restore treats a local leaf symlink or any symlinked ancestor beneath the
+selected session root as a conflict, even when the linked content matches the
+selected remote object. `--force` does not override that conflict.
+
+Before each install, restore validates the destination again beneath the
+selected session root. A target changed into a symlink, directory, or other
+non-regular entry is not overwritten.
 
 Restore writes an execution report at `reports/restore-latest.json`.
 
@@ -473,10 +621,24 @@ Durable session paths:
 - `logs/**`
 - `config/**`
 - `runs/**`
+- `cache/notarius/references/<sha256>/<prepared-basename>`
+
+`manifest.json` is the versioned session-progress document. Narratio accepts
+only `narratio.session-manifest.v1`. Do not edit unsupported session state to
+add missing markers. Move it aside and regenerate it, restore a verified
+current committed snapshot, or republish healthy regenerated state as the
+relevant workflow requires.
 
 Validated Notarius bundles live below `artifacts/notarius/{run_id}/`; receipt,
 stderr, and pre-promotion output remain in the producing run's `extract`
 directory as described in [Extraction Workflow](#extraction-workflow).
+Stable reference snapshots live in the session cache path above. Ordinary
+published-run cleanup removes its selected run-local target but retains this
+cache, and publish/restore never enumerate it as an artifact. Remove a snapshot
+or digest directory manually only while no command is active for the session;
+regular missing or corrupt content is reconstructed on the next executed
+extraction, while symlinks, directories, and other unsafe conflicts must be
+investigated and removed deliberately.
 
 Run-local layout:
 
@@ -498,6 +660,8 @@ Cache layout (durable S3 audio cache):
 Each cached audio file has an adjacent managed identity record. It binds the
 file to its remote object version and verified digest; deleting or altering the
 record simply causes Narratio to download and verify the object again.
+Remote audio and restored files are fully transferred into sibling temporary
+files and validated before they replace their canonical workspace paths.
 
 ### Workspace Permissions
 
@@ -509,18 +673,28 @@ converge on `0664`. Narratio explicitly applies these modes so a restrictive
 caller umask does not remove group write or setgid. It does not change file or
 directory ownership: the configured workspace's existing group is inherited.
 
-Windows does not implement POSIX mode bits or setgid semantics. Configure the
-workspace, spool, and cache locations with an ACL that grants the collaborating
-group read/write access, and configure credential locations with an ACL limited
-to the intended credential owner. Do not use POSIX mode displays as evidence of
-Windows access control.
-
 API keys are credentials, not ordinary workspace data. Store them outside the
 shared workspace or in a separately restricted credential location; ordinary
 workspace group access must never be treated as authorization to read keys.
-On POSIX, provision a credential directory as `0700` and credential files as
+Provision a credential directory as `0700` and credential files as `0600`;
-`0600`; Narratio rejects group- or other-readable configured credential paths.
+these permission modes are an operator requirement and are not enforced by
-On Windows, restrict the directory and files with ACLs to the credential owner.
+ConfigKit.
+
+Every pipeline-loading command bootstraps credentials once, before the first
+configuration load and before any adapter is constructed. `--secrets-dir`
+selects an explicit directory (relative paths use caller CWD). Otherwise a
+present `secrets/` beside the resolved root pipeline is used, while an absent
+conventional directory is skipped. Explicit missing/empty selections and
+invalid existing directories fail the command. ConfigKit follows a symlink used
+as the selected directory, but immediate entries must be regular non-symlink
+files. Existing environment variables—including empty values—always win.
+
+S3 `access_key_id` and `secret_access_key` are an optional pair of typed
+environment references. When both are absent, the AWS SDK retains its default
+credential provider chain. When both are configured, Narratio supplies their
+load-time resolved values through an explicit static provider. Audita's optional
+`llm_api_key` is likewise captured at pipeline load and supplied only through
+the child process's `AUDITA_LLM_API_KEY` override.
 
 External adapter results are individually bounded before Narratio validates or
 materializes them. These per-file limits do not reserve disk space: prevent hard
@@ -553,17 +727,29 @@ Rules:
 - `clean` deletes work/spool session state;
 - cache is preserved unless `--clear-cache` is set;
 - each deletion is confined beneath its configured workspace, spool, or cache
-  root and refuses symlinked paths;
+  root; root, traversal, and out-of-root targets are refused;
-- automatic post-publish cleanup is gated by successful publish commit plus:
+- the final cleanup stage is gated by successful publish evidence plus:
   - `pipeline.spool.delete_audio_after_publish=true`
   - `pipeline.workspace.cleanup_after_publish=true`
-  - Narratio first records the exact run-scoped cleanup obligation. If cleanup
+  - the manifest publish record has `uploaded=true`, the published run ID, the
-    reports incomplete, the remote committed snapshot remains current; rerun
+    exact remote commit key, and the current commit-pointer key.
-    publish to retry only the outstanding confined local cleanup.
 
-Post-publish cleanup is evaluated only when `publish` actually executes in the
+Cleanup derives only the configured run-scoped spool and workspace targets,
-current invocation. A bounded range that excludes publish does not replay a
+uses root-confined deletion, and does not read or mutate object storage. A
-cleanup obligation as an unrelated side effect.
+failure leaves publish succeeded and the remote committed snapshot current.
+Retry with a normal full run: valid upstream work, including publish and notify,
+is skipped while the failed cleanup stage runs again. `run-stage cleanup` is
+also available when the manifest has valid publish evidence. A range ending at
+`notify` or earlier and the standalone `publish` command never delete local
+state.
+
+Older local manifests that contain the retired cleanup-obligation field are
+intentionally incompatible with strict manifest decoding. If such an
+obligation is still outstanding, complete it with the pre-upgrade binary.
+Otherwise verify that the remote publication is committed, move aside or
+regenerate the incompatible local manifest (or restore the verified remote
+snapshot), and then use the cleanup stage. Manual JSON editing is not the
+normal migration path.
 
 ## Operational Caveats
 
@@ -571,4 +757,5 @@ cleanup obligation as an unrelated side effect.
 - Publish requires prerequisite stages through `render` and `analyze` to be succeeded.
 - Markdown publish defaults require render outputs (`transcripts/final.md` and `transcripts/final.trimmed.md`).
 - Restore requires configured object storage and committed remote current state.
-- Storage-backed commands load filesystem secrets before object-store initialization.
+- Pipeline-loading commands bootstrap filesystem secrets once before configuration
+  resolution and object-store initialization.

docs/policy/architecture.md

@@ -116,8 +116,11 @@ adapter and publish metadata.
 
 Resume and skip decisions are manifest-driven. Filesystem state may be
 inspected and validated, but file presence alone does not replace recorded
-stage state. Invocation-scoped run records provide an audit of one execution;
+stage state. Each invocation retains an in-memory command summary and a
-they do not replace the session manifest as progress authority.
+run-local workspace identity, but neither is a durable progress ledger. The
+physical run directory is created on demand only when an executing stage
+requires it; a fully reused invocation or a cleanup-only invocation does not
+leave an empty run directory.
 
 Restore treats committed remote current state as its authority. It must plan
 deterministically, confine remote-to-local paths, protect local conflicts, and
@@ -143,7 +146,7 @@ Configuration is strict, explicit, centralized, and operator-oriented.
 - YAML decoding rejects unknown fields.
 - Defaults are centralized and testable.
 - Empty configured values do not silently replace meaningful defaults.
-- Validation rejects invalid composition before stage execution where
+- Validation rejects invalid configuration before stage execution where
   practical.
 - Root-owned imports and one selected profile resolve deterministically through
   the configuration owner; commands do not implement their own merge rules.
@@ -185,8 +188,8 @@ availability.
 Writes, moves, replacements, and deletions must use narrow, explicit,
 root-confined destinations. Symlinks, traversal, broad roots, and ambiguous
 relative destinations must not expand the scope of an operation. Cleanup is
-permitted only through explicit operator action or configured post-publish
+permitted only through explicit operator action or the configured final cleanup
-gates, and it must preserve durable cache unless cache removal is explicitly
+stage, and it must preserve durable cache unless cache removal is explicitly
 requested.
 
 Physical layout, retention, and operational lifecycle belong in
@@ -202,16 +205,26 @@ the immutable commit manifest, and finally the current commit pointer.
 `current/commit-pointer.json` is the sole mutable selector and must be written
 exactly once, last. Failed, incomplete, skipped, or uncommitted publish attempts
 must not be presented as current remote state. Publish locks remain authoritative
-and are not bypassed by a forced run. Mutable remote locks use provider-enforced
+and are not bypassed by a forced run. Each mutable remote lock is an independent
-generation preconditions and are revalidated immediately before pointer
+per-source object; lock mutations use provider-enforced generation preconditions.
-selection; loss of that check leaves the prior committed snapshot current.
+Publish revalidates every initially unlocked participating source after writing
+the immutable commit and before pointer selection. A lock observed then leaves
+the prior committed snapshot current. A lock created after that final check may
+race the pointer write; the commit protocol does not claim a stronger distributed
+transaction boundary.
 
 Automatic local cleanup is permitted only after a successful publish commit,
 only when explicitly configured, and only through the path-safety guardrails.
-It is a durable local obligation bound to that committed run and its exact
+It is the final ordinary pipeline stage after `notify`. The stage derives its
-targets, not an inferred side effect of the current stage list. A cleanup
+exact run-scoped targets from validated publish metadata and the configured
-failure makes the invocation incomplete while leaving the committed remote
+campaign/session identity; it never infers a target from directory contents or
-snapshot authoritative; later invocations resume the recorded obligation.
+rereads remote storage. A cleanup failure makes the invocation incomplete while
+leaving publish succeeded and the committed remote snapshot authoritative. A
+normal later full run reuses valid upstream stages and retries cleanup.
+
+The standalone `publish` command selects only publish and therefore never
+invokes cleanup. Operators may select cleanup directly only when the manifest
+contains valid successful publication evidence.
 
 ## Security, Privacy, And Diagnostics
 
@@ -221,20 +234,26 @@ configuration, logs, reports, diagnostics, and Notarius artifacts—is
 intentionally shareable with the configured workspace group. API-key material
 is sensitive and is not covered by the ordinary workspace-sharing policy.
 
-On POSIX systems, Narratio-created ordinary workspace directories converge on
+Narratio supports Linux on amd64 and arm64. macOS arm64 is best effort: it has
+a cross-build gate and release asset, but no native runtime release gate.
+Windows and macOS amd64 are unsupported and have no compatibility promise or
+release asset.
+
+On supported systems, Narratio-created ordinary workspace directories converge on
 setgid `02775` and ordinary workspace files on `0664`, even when the caller's
 umask is restrictive. This preserves the existing workspace group for nested
 creation and atomic replacements without changing ownership. API-key storage
-uses a separate restrictive contract. On Windows, POSIX mode bits and setgid
+uses a separate restrictive contract.
-are not authoritative; operators must provide the equivalent shared-group and
-credential-restricted ACLs described in [Operations](../operations.md#workspace-permissions).
 
 Raw secrets must not be stored in pipeline, campaign, or session YAML or written
 to manifests, logs, generated configuration, reports, publish metadata,
-documentation, or examples. Secrets enter through configured environment
+documentation, or examples. Pipeline configuration names environment variables;
-variable names or secret-file references. Diagnostics should avoid transcript
+command-owned startup may populate previously unset variables from ConfigKit's
-and prompt content unless a deliberate, bounded inspection mechanism requires
+explicit or conventional secrets directory. This bootstrap occurs once per
-it.
+invocation before any typed configuration load or adapter construction and is
+not hidden inside library loaders. Safe bootstrap diagnostics may identify
+files and statuses but never values. Diagnostics should avoid transcript and
+prompt content unless a deliberate, bounded inspection mechanism requires it.
 
 Logs, reports, generated invocation files, generated configuration, and render
 debug files are diagnostics, not canonical pipeline products. They should be

docs/release.md

@@ -5,6 +5,13 @@ binary release. The synchronous release boundary is a successful push of one
 new tag to `origin`; Woodpecker and Gitea publication happen later and do not
 change that result.
 
+## Go prerequisite
+
+Release preparation requires the maintained Go 1.26 baseline. Tag CI currently
+uses the explicitly pinned `golang:1.26.7` image for candidate validation and
+asset construction; that asynchronous work does not change the tag-push
+completion boundary.
+
 ## Choose a version and write its note
 
 Narratio is past `v1.0.0`. Use an unused stable tag in the exact form
@@ -51,14 +58,14 @@ scripts/check-release-candidate.sh vMAJOR.MINOR.PATCH
 
 It validates the version and matching note, module hygiene, formatting,
 whitespace, uncached tests, race tests, static checks, documentation, examples,
-and six official cross-build assets. It uses `GOWORK=off`, does not contact
+and three official cross-build assets. It uses `GOWORK=off`, does not contact
 application services, CI, or Gitea, and does not create tags or modify tracked
 source. Fix any failure on `main`, commit it, push it normally, and rerun the
 checker.
 
-The checker builds Linux, macOS, and Windows assets for `amd64` and `arm64`.
+The checker builds Linux assets for `amd64` and `arm64` and a macOS asset for
-Cross-builds prove compilation; they are not native macOS or Windows runtime
+`arm64`. The macOS cross-build proves compilation; it is not native macOS
-evidence.
+runtime evidence.
 
 ## Publish the tag
 

docs/releases/README.md

@@ -22,5 +22,7 @@ See the [release procedure](../release.md) for creating a candidate and tag.
 When asynchronous publication succeeds, the corresponding Gitea release is the
 canonical source for downloadable binaries and checksums.
 
+- [v2.0.0](v2.0.0.md)
+- [v1.7.0](v1.7.0.md)
 - [v1.6.0](v1.6.0.md)
 - [v1.5.0](v1.5.0.md)

docs/releases/v1.7.0.md

@@ -0,0 +1,100 @@
+# Narratio v1.7.0
+
+Narratio v1.7.0 completes the maintainability refactor begun after v1.6.0. It
+simplifies filesystem and process handling, makes resumable work more precise,
+replaces legacy local and remote state with explicit authorities, and turns
+cleanup into the final ordinary pipeline stage.
+
+## Summary
+
+The session manifest is now the sole durable local progress ledger. Analyze
+checkpoints successful artifacts individually, allowing an interrupted or
+partially failed analysis to resume at the first incomplete artifact without
+rerunning successful siblings unless forced.
+
+Remote publication now commits immutable run state before advancing one
+verified current pointer. Publish locks are independent per-source objects;
+locked outputs retain their previously committed declarations rather than
+being overwritten or silently omitted. Local cleanup runs after notification
+as a separately recorded, safely retryable stage.
+
+The release also adopts Go 1.26, simplifies ordinary file operations around
+Narratio's trusted group-owned workspace model, improves subprocess lifecycle
+handling, decomposes several large orchestration and validation components,
+and removes mutable global test seams.
+
+## Compatibility
+
+This is a minor release with deliberate removal of compatibility paths that
+were deprecated and documented as subject to removal:
+
+- Campaigns must use one versioned `narratio.party.v1` `party.yml`. Separate
+  players files, session-level party overrides, and unversioned party documents
+  are rejected.
+- Session manifests must declare `schema_version:
+  narratio.session-manifest.v1` and satisfy the current strict schema. Older
+  unversioned state and manifests containing the retired
+  `post_publish_cleanup` obligation are rejected rather than migrated.
+- Remote current state must be selected by
+  `current/commit-pointer.json` and its referenced immutable run commit. Legacy
+  current-state layouts are no longer inferred.
+- The retired aggregate `locks.yml` object is ignored. Publish locks now live
+  as independent strict objects under the session lock prefix.
+- Run manifests are no longer produced or consumed. The session manifest owns
+  durable stage progress; invocation summaries and run workspaces remain
+  operational data rather than a second progress authority.
+
+The CLI and current pipeline/configuration formats otherwise remain compatible.
+Building and maintaining Narratio now requires Go 1.26. Cross-built Linux,
+macOS, and Windows assets remain available for `amd64` and `arm64`.
+
+## Upgrade
+
+Before upgrading a campaign that still uses compatibility data:
+
+1. Replace legacy party/player inputs with the documented
+   [`narratio.party.v1`](../integrations/party.md) campaign roster.
+2. If a local manifest contains an outstanding `post_publish_cleanup`
+   obligation, complete it with the pre-upgrade binary. Otherwise verify the
+   remote commit, move aside or regenerate the incompatible manifest, or
+   restore a verified committed snapshot. Do not migrate state by manually
+   editing JSON.
+3. Republish sessions that do not yet have a valid immutable run commit and
+   `current/commit-pointer.json`; legacy remote layouts are not current state in
+   v1.7.0.
+4. Recreate any required publication locks with `narratio session locks add`.
+   After verifying the new locks and publication behavior, the ignored legacy
+   `locks.yml` object may be removed manually.
+
+Existing compatible session manifests and already committed remote snapshots
+need no special action. Consult [Operations](../operations.md) and
+[Troubleshooting](../troubleshooting.md) for recovery procedures.
+
+## Changes
+
+- Upgraded the maintained development and release baseline to Go 1.26.
+- Reworked subprocess execution and Windows process-tree supervision while
+  preserving contextual wait and cleanup failures.
+- Replaced custom confined ordinary I/O and directory-promotion machinery with
+  simpler standard-library operations while retaining bounded reads, atomic
+  single-file visibility, workspace modes, validated downloads, and
+  root-confined destructive operations.
+- Removed legacy party, manifest, artifact-discovery, remote-current, and lock
+  compatibility paths.
+- Consolidated durable progress in the session manifest and removed the run
+  manifest domain model.
+- Unified runner planning and execution decisions, separated invocation
+  lifecycle responsibilities, and avoided creating empty run workspaces for
+  fully reused or cleanup-only invocations.
+- Added per-artifact analyze execution, checkpointing, reconciliation, and
+  recovery.
+- Decomposed configuration validation by domain and made validation diagnostics
+  deterministic without treating incidental ordering or wording as a contract.
+- Introduced immutable remote publish commits, one final current pointer,
+  explicit output planning, safe cross-run carried outputs, and independent
+  per-artifact remote locks.
+- Added `cleanup` as the final canonical stage after `notify`, with validated
+  publish evidence, safe retry, explicit selection, and no separate cleanup
+  obligation.
+- Replaced mutable command test hooks with invocation-scoped dependencies and
+  improved structural test diagnostics.

docs/releases/v2.0.0.md

@@ -0,0 +1,94 @@
+# Narratio v2.0.0
+
+Narratio v2.0.0 replaces its configuration infrastructure with ConfigKit,
+simplifies the supported release platforms, and makes artifact definitions and
+their effective dependency graph directly inspectable.
+
+## Summary
+
+Pipeline roots now use ConfigKit v0.1.2 for strict YAML loading, explicit
+imports, fixed-path named profiles, provenance, defaults, and typed credential
+references. The recommended production layout keeps production values in the
+base, defines each ordinary artifact and artifact family in its own imported
+file, and keeps testing changes in `profiles/testing.yml`.
+
+Configured-artifact inputs now create their dependency edges automatically.
+`narratio config artifacts` reports the finalized concrete catalog, including
+generated party-family members, without loading a session or initializing
+runtime adapters.
+
+## Compatibility
+
+This is a breaking configuration and platform release. There is no runtime
+fallback for the retired Narratio composition/profile catalog, `pipeline.secrets`
+or `secrets.env_dir`, credential `_env` fields, or the previous YAML
+presence/default behavior. Defaults apply only to absent values; explicit empty,
+zero, false, and null values retain their typed meaning and may fail Narratio
+domain validation.
+
+The maintained binaries are Linux amd64, Linux arm64, and macOS arm64. Windows
+and macOS amd64 builds are no longer produced or supported. Existing v1 release
+notes remain historical descriptions of their releases.
+
+Inspection commands still avoid session, manifest, workspace, adapter, and
+remote-service activity, but they are not credential-free loaders: every
+explicit typed credential reference in the selected effective configuration must
+resolve; references replaced or removed by that profile need not resolve.
+Pipelines containing artifact families also require a campaign so the canonical
+party can be expanded.
+`config show` is a secret-safe review projection, not a promised reloadable
+authoring file.
+
+## Upgrade
+
+Migrate each operational bundle before running v2:
+
+1. Replace old composition syntax with root-owned
+   [`configkit.imports`](../config.md#pipeline-imports-and-profiles). Imports are
+   additive base sources; they do not override one another.
+2. Fold the old production overlay into the base pipeline. Move testing changes
+   to `profiles/testing.yml`; compare the two independent resolutions with
+   `narratio config diff default testing`.
+3. Follow the maintained
+   [`examples/production-testing`](../../examples/production-testing/pipeline.yml)
+   layout: one explicit artifact entry per `artifacts/<key>.yml` and one family
+   entry per `artifact-families/<key>.yml`. The map key, not the filename, is the
+   stable artifact identity.
+4. Move credential files to an untracked `secrets/` directory beside
+   `pipeline.yml`, or pass `--secrets-dir`. Rename S3 fields to `access_key_id`
+   and `secret_access_key`, and rename the Audita field to `llm_api_key`.
+5. For static S3 credentials, configure both S3 references (commonly
+   `OBJECT_STORAGE_KEY_ID` and `OBJECT_STORAGE_KEY`) and make them resolvable.
+   To use instance roles, workload identity, shared SDK files, or another AWS SDK
+   discovery provider, omit both references. A partial pair is invalid.
+6. Remove a `depends_on` entry when it only duplicates a
+   `narratio.artifact.<key>` input. Keep explicit entries that express ordering
+   without supplying the upstream output as an input.
+7. Review ConfigKit's stricter explicit-value/null semantics and correct values
+   that previously acquired defaults after being explicitly empty. Do not use
+   `config show` output as a source migration mechanism.
+8. Run `config validate`, `config show`, `config sources`, and `config artifacts`
+   with the same profile, campaign, and secrets selection intended for runtime.
+   Resolve the campaign-owned party whenever families are present.
+9. Update deployment and download automation to one of the three maintained
+   platform assets before removing the v1 binary.
+
+The [configuration reference](../config.md), [CLI reference](../cli.md), and
+[maintained examples](../../examples/README.md) are the canonical detailed
+guides.
+
+## Changes
+
+- Adopted ConfigKit v0.1.2 across pipeline, campaign, session, and party loading
+  while keeping Narratio domain validation and post-campaign expansion explicit.
+- Added ConfigKit imports and fixed `profiles/<name>.yml` pipeline overlays,
+  unified safe effective show/diff/digest projection, and source provenance.
+- Added invocation-scoped ConfigKit secrets-directory bootstrap and typed S3 and
+  Audita credential references without exposing resolved values.
+- Inferred, normalized, and validated configured-artifact input dependencies as
+  part of the one concrete graph consumed by analyze planning and resume logic.
+- Added `config artifacts` with stable concrete identities, definition sources,
+  effective dependencies, and generated family/party provenance.
+- Added the maintained production-base/testing-overlay artifact authoring layout.
+- Reduced release assets and active platform support to Linux amd64/arm64 and
+  macOS arm64.

docs/releases/v2.1.0.md

@@ -0,0 +1,70 @@
+# Narratio v2.1.0
+
+This release adds explicit Notarius credential forwarding and cache reuse,
+with predictable forced regeneration and clearer artifact-development guidance.
+
+## Summary
+
+Narratio can pass one typed API-key reference to Notarius and opt into reuse of
+its accepted execution checkpoints. Stable, verified reference snapshots allow
+unchanged references to retain their cache identity across Narratio invocations.
+Configuration loading now uses ConfigKit v0.3.0.
+
+## Compatibility
+
+The new `notarius.llm_api_key`, `notarius.resume`, and `notarius.chunk_cache`
+fields are optional. Existing v2 configuration needs no schema migration.
+Configured credential references must resolve during configuration loading,
+even when extraction will not execute; omitted references add no requirement.
+
+Forced extraction now omits Notarius checkpoint loading and requests chunk-plan
+refresh. This includes `regenerate-artifacts`, which remains an alias for a
+forced run from extract through analyze. `analyze` continues to force analysis
+only, preserving extraction results and transcripts.
+
+Notarius v0.7.0 is the tested cache integration baseline, not an unconditional
+minimum-version gate. The executable must support the flags selected for the
+invocation, including `--chunk_cache refresh` for forced extraction. Reference
+bindings retain the contract introduced in Notarius v0.6.0. Receipt and artifact
+schemas are unchanged; reused producers may report zero production attempts.
+
+Supported release platforms remain Linux amd64/arm64 and best-effort macOS
+arm64. macOS validation is cross-compilation, not native runtime testing.
+
+## Upgrade
+
+No special migration is required for existing v2 configuration. To use the new
+features:
+
+- Declare `notarius.llm_api_key` with the environment-variable name expected by
+  the deployed Notarius profile, supplying its value through the existing
+  environment or secrets-directory mechanism.
+- Enable checkpoint recording in Notarius configuration before setting
+  Narratio's `notarius.resume: true`. Prefer explicit persistent Notarius cache
+  roots for service deployments. An optional `notarius.chunk_cache` overrides
+  its ordinary chunk-plan policy.
+- Expect an initial checkpoint miss when moving from old run-local reference
+  paths to stable session snapshots. No migration or deletion of old caches is
+  required. Force refreshes the current input's plan, not the whole cache.
+- Use `narratio analyze SESSION` for analysis-only iteration, and
+  `narratio regenerate-artifacts SESSION` when extraction should also be rebuilt.
+
+See [Configuration](../config.md), the
+[Notarius integration contract](../integrations/notarius.md), and
+[Artifact Development Workflows](../operations.md#artifact-development-workflows)
+for current details.
+
+## Changes
+
+- Upgraded ConfigKit from v0.1.2 to v0.3.0 and isolated pipeline-discovery tests
+  from machine configuration.
+- Added optional typed Notarius credentials, exact-name environment delivery,
+  safe configuration projections, and captured-stream redaction for custom names.
+- Added Notarius resume and chunk-cache controls with force taking precedence.
+- Added verified, content-addressed session reference snapshots that survive
+  ordinary run cleanup without becoming published artifacts.
+- Preserved relative-workspace support for snapshot delivery and confinement.
+- Kept credentials and cache-policy controls outside extract's semantic resume
+  identity, so operational changes alone do not invalidate completed results.
+- Added an opt-in real Notarius integration test using a local provider to
+  verify reuse, partial-failure recovery, and forced refresh without paid calls.

docs/releases/v2.1.1.md

@@ -0,0 +1,33 @@
+# Narratio v2.1.1
+
+This release fixes configuration loading for publish rules that explicitly set
+requiredness while relying on a derived destination.
+
+## Summary
+
+Publish rules with a derivable source and omitted `dest` now load correctly
+whether `required` is omitted, explicitly true, or explicitly false. Previously,
+explicit requiredness could trigger ConfigKit's `resolve.unrecorded_change`
+diagnostic instead of accepting valid configuration.
+
+## Compatibility
+
+This is a backward-compatible bookkeeping correction. Publish schemas,
+destination derivation, requiredness defaults, and YAML syntax support are
+unchanged. Explicit destinations and explicit false values remain intact.
+No ConfigKit upgrade or publication/storage behavior change is involved.
+
+## Upgrade
+
+No special upgrade action or configuration migration is required. Operators
+need not add a derivable `dest`, remove explicit `required` values, or change
+flow-style YAML to block-style YAML. See the
+[configuration reference](../config.md#publish-configuration-summary).
+
+## Changes
+
+- Recorded derived publish destinations at ConfigKit's atomic sequence boundary
+  while retaining field-level resolution records.
+- Added real load/validation and safe effective-configuration regression coverage
+  for built-in and configured-artifact sources, explicit requiredness, and an
+  explicit destination that must remain unchanged.

docs/releases/v2.1.2.md

@@ -0,0 +1,26 @@
+# Narratio v2.1.2
+
+This release allows secrets directories to contain hidden metadata without
+preventing credential loading.
+
+## Summary
+
+ConfigKit v0.3.1 ignores dot-prefixed entries such as `.git`, `.gitignore`, and
+`.DS_Store` during secrets bootstrap. These entries no longer block otherwise
+valid credentials from loading and are omitted from bootstrap reports.
+
+## Compatibility
+
+This is a backward-compatible dependency bug fix. Configuration schemas,
+existing-environment precedence, and validation of non-hidden entries are
+unchanged. Hidden entries still count toward the directory entry limit.
+
+## Upgrade
+
+No special upgrade action or configuration migration is required. See the
+[credential configuration reference](../config.md) for secrets-directory usage.
+
+## Changes
+
+- Updated ConfigKit from v0.3.0 to v0.3.1.
+- Clarified hidden-entry handling in the secrets-directory documentation.

docs/releases/v2.2.0.md

@@ -0,0 +1,53 @@
+# Narratio v2.2.0
+
+This release adds resumable per-speaker transcription, replayable WhisperX
+uploads, and explicit Notarius configuration profile selection.
+
+## Summary
+
+Successfully checkpointed speaker transcripts survive later transcription
+failures. Ordinary retry reuses valid completed speakers and processes only the
+remaining work. WhisperX uploads can now be replayed by the HTTP transport
+without buffering whole audio files.
+
+## Compatibility
+
+Existing CLI commands, speaker identities, transcript formats, and pipeline
+ordering are unchanged. Session manifests gain optional per-speaker completion
+evidence. Older manifests remain readable, but older Narratio binaries may
+reject manifests containing the new fields; downgrading after transcription
+is not transparent.
+
+The optional `pipeline.notarius.config_profile` field selects a downstream
+Notarius configuration profile independently of Narratio's own profile. Using
+it requires a Notarius version supporting `--config-profile`; v0.7.0 is a known
+supporting version. Omitting the field retains downstream default selection.
+
+## Upgrade
+
+No configuration migration is required. When transcribe is next selected for
+an older aggregate-only manifest, all current speakers are transcribed once
+to establish reusable evidence. Previously unrecorded results from failed runs
+are not automatically imported.
+
+Use ordinary run/resume to retain valid checkpoints. Forcing transcribe resets
+reuse of prior results; an ordinary retry after a failed forced invocation can
+retain replacements already checkpointed by that attempt. See the
+[operations guide](../operations.md) and
+[Notarius profile configuration](../config.md#notarius-configuration-profile).
+
+## Changes
+
+- Persist validated speaker completions independently of aggregate stage status,
+  with serialized saves for concurrent work and input/output integrity checks.
+- Rebuild aggregate transcript references when speakers are removed or those
+  references need repair, without retranscribing otherwise-current speakers.
+- Keep removed and incomplete speaker results out of versioned merge inputs.
+- Permit checkpoint evidence for zero-byte audio when WhisperX accepts it;
+  service rejection still fails normally.
+- Provide fresh streaming multipart bodies for HTTP transport replay while
+  preserving application retries, cancellation, and resource cleanup. This
+  addresses the missing-body-replay limitation, not persistent upstream
+  protocol failures or exactly-once service execution.
+- Pass explicit Notarius configuration profiles through the CLI and include
+  their selection in extraction reuse decisions.

docs/roadmap/feature-request-notarius-config-profile.yml

@@ -0,0 +1,102 @@
+---
+title: Select a Notarius configuration profile through Narratio
+status: proposed
+consumer: rakestraw.dnd Ansible collection and its pipeline role
+
+use_case: >-
+  The Ansible pipeline role deploys Narratio and Notarius as independently
+  configured applications. Operators select Narratio's testing profile to
+  develop prompts and other pipeline assets with inexpensive models while
+  retaining production settings in the base configuration. That selection
+  currently cannot select a Notarius configuration profile, leaving extraction
+  on its production settings.
+
+desired_outcome: >-
+  Narratio's testing overlay can explicitly select Notarius's testing overlay.
+  Notarius owns the resulting model choices and any other testing-specific
+  settings. The operator selects testing once when invoking Narratio, without
+  editing production files or separately invoking Notarius.
+
+rationale: >-
+  Both applications already use ConfigKit configuration overlays. Selecting
+  the downstream overlay preserves each application's configuration ownership
+  and supports future Notarius testing adjustments without adding a Narratio
+  configuration field for every downstream setting. This is explicit profile
+  composition, not a requirement for arbitrary recursive orchestration.
+
+requested_interface:
+  field: pipeline.notarius.config_profile
+  value: Optional Notarius ConfigKit configuration profile name
+  supplied: >-
+    Invoke Notarius with --config-profile and the selected name when extraction
+    executes. The setting is available in Narratio's base configuration and
+    its ordinary configuration overlays.
+  omitted_or_null: >-
+    Do not supply a profile-selection flag. Preserve existing Notarius
+    configuration selection behavior.
+  downstream_semantics: >-
+    Retain Notarius's existing selector semantics, including the default
+    selector for its base configuration. Notarius resolves profiles relative
+    to its own configuration root, not Narratio's root.
+  explicit_selection: >-
+    Do not automatically forward Narratio's profile name. The selected
+    Notarius profile may have the same name or a different name.
+  distinction: >-
+    This selects a ConfigKit configuration overlay through --config-profile,
+    not a PromptKit LLM profile through --llm-profile. The Notarius overlay
+    can itself select the desired LLM profiles using native Notarius settings.
+
+illustrative_consumer_configuration:
+  note: >-
+    These are separate files managed by the Ansible suite, not a proposed
+    combined configuration document. The Notarius pipeline and LLM profile
+    names are illustrative; the referenced LLM profile must be available to
+    Notarius.
+  narratio_profiles_testing_yml: |
+    notarius:
+      config_profile: testing
+  notarius_profiles_testing_yml: |
+    pipelines:
+      dnd-session:
+        llm_profile: gemini-flash-latest
+  operator_workflow: >-
+    Select --profile testing on the ordinary Narratio invocation. Narratio
+    selects Notarius's testing configuration overlay when it runs extraction;
+    other Narratio testing overrides continue to apply normally.
+
+observable_behavior:
+  - >-
+    Notarius owns overlay discovery, composition, validation, and the effective
+    downstream settings. A missing or invalid selected overlay must not silently
+    fall back to production or be reported as successful extraction.
+  - >-
+    Changing the selected Notarius configuration profile must not silently
+    reuse Narratio's completed extraction from the previous selection.
+    Existing downstream invalidation behavior applies when extraction changes.
+  - >-
+    Editing Notarius overlay contents behind an unchanged selector retains
+    the existing operator responsibility to force extraction when needed.
+    This request does not require inspecting or tracking private Notarius files.
+  - >-
+    Existing credential forwarding remains independently configurable through
+    notarius.llm_api_key, including in the Narratio testing overlay when the
+    selected Notarius settings require a different credential reference.
+
+consumer_follow_up: >-
+  After Narratio releases this interface, the standalone Narratio Ansible role
+  can expose an optional variable and omit the field when unset. The suite role
+  can deploy Notarius's testing overlay and select it from Narratio's testing
+  overlay, then distribute the changes through the collection. These consumer
+  updates are separate from this upstream feature request.
+
+scope_boundaries:
+  - No Notarius change is currently required; v0.7.0 supports --config-profile.
+  - No separate Narratio-side Notarius LLM-profile selector is requested.
+  - No automatic profile-name propagation or downstream configuration rewriting.
+  - No new credential discovery, copying, or provider-name remapping.
+  - This request specifies consumer behavior, not internal Narratio code changes.
+
+references:
+  narratio_configuration: ../config.md
+  narratio_notarius_integration: ../integrations/notarius.md
+  notarius_cli: https://gitea.maximumdirect.net/eric/notarius/src/tag/v0.7.0/docs/cli.md

docs/roadmap/fix-publish-output-destination-resolution.md

@@ -0,0 +1,171 @@
+# Bugfix Request: Preserve Publish Destination Derivation with Explicit Rules
+
+Status: Implemented; release target v2.1.1
+
+## Summary
+
+Narratio should successfully derive a publish destination when a valid
+`publish.outputs` rule omits `dest`, regardless of whether that rule explicitly
+sets `required`. Narratio v2.1.0 currently rejects this valid configuration in
+some cases with a ConfigKit `resolve.unrecorded_change` diagnostic.
+
+This is a configuration-resolution bookkeeping defect, not a YAML formatting
+restriction. The fix should preserve the existing public configuration
+contract: operators may omit a derivable destination and may independently set
+`required` to either `true` or `false`.
+
+## Observed Behavior
+
+A pipeline containing a rule equivalent to the following fails during native
+configuration loading:
+
+```yaml
+publish:
+  outputs: [{"source": "narratio.transcript.final", "required": true}]
+```
+
+Narratio reports:
+
+```text
+error [resolve.unrecorded_change] /publish/outputs: application resolution changed a value without a covering record
+```
+
+The same failure is possible with ordinary block-style YAML when `required` is
+explicit and `dest` is omitted. Flow-style YAML is valid input and is not the
+cause of the defect.
+
+The existing destination-derivation tests do not expose this case because
+their rules also omit `required`. Narratio then supplies the default value for
+`required` and records a change covering the complete `publish.outputs`
+sequence. That broader record happens to cover the destination mutation as
+well, masking the missing resolution record that occurs when `required` is
+already explicit.
+
+## Expected Contract
+
+The current configuration contract remains unchanged:
+
+- `publish.outputs[].source` is required;
+- `publish.outputs[].dest` may be omitted when Narratio can derive it from the
+  selected built-in or configured artifact source;
+- `publish.outputs[].required` defaults to `true`, but an operator may set it
+  explicitly to either `true` or `false`; and
+- equivalent valid YAML representations have equivalent configuration
+  semantics.
+
+For example, both rules below should load and validate, retain their explicit
+requiredness, and receive the canonical derived destination:
+
+```yaml
+publish:
+  outputs:
+    - source: narratio.transcript.final
+      required: true
+    - source: narratio.transcript.final_markdown
+      required: false
+```
+
+## Root Cause
+
+ConfigKit intentionally treats sequences as atomic when comparing values
+before and after application resolution. Changing a field within one
+`publish.outputs` entry is therefore observed as a change to the
+`/publish/outputs` sequence.
+
+Narratio currently derives each missing destination by recording a
+normalization at the individual `publish.outputs[index].dest` path. That child
+record does not cover the atomic parent-sequence change. A separate
+container-level record is currently created only when Narratio also defaults
+an omitted `required` value, so configurations with explicit `required` values
+lose that incidental coverage.
+
+## Recommended Fix
+
+Make Narratio's publish-output resolution bookkeeping conform to ConfigKit's
+atomic-sequence contract. Whenever resolution derives one or more destinations
+inside `publish.outputs`, the recorded resolution changes must cover the
+`/publish/outputs` container. Destination derivation must not depend on an
+unrelated `required` default being applied in the same list.
+
+The implementation may retain more specific provenance where useful, but it
+must provide valid container-level coverage for the sequence mutation. Do not
+change ConfigKit's sequence comparison semantics merely to accommodate this
+call site.
+
+## Scope
+
+- Correct destination derivation for publish rules that explicitly set
+  `required` and omit a derivable `dest`.
+- Preserve explicit `required: true` and `required: false` values.
+- Preserve canonical destination derivation for built-in transcript sources
+  and configured artifact sources.
+- Add focused regression coverage through Narratio's real configuration load
+  and validation boundary.
+- Prepare a backward-compatible patch release after the fix passes the normal
+  repository validation.
+
+## Non-goals
+
+- Changing the `publish.outputs` schema, defaults, source vocabulary, or
+  destination derivation rules.
+- Requiring operators or configuration generators to supply `dest` when it is
+  derivable.
+- Requiring operators to omit explicit `required: true` values.
+- Treating flow-style YAML or JSON-compatible YAML values as unsupported.
+- Changing ConfigKit's documented atomic treatment of sequences.
+- Adding storage, publication, or network integration behavior.
+
+## Regression Coverage
+
+Add one focused package-level behavioral test that loads and validates a
+pipeline with omitted destinations and explicit requiredness. A representative
+list should include both `required: true` and `required: false`, then assert
+that:
+
+- loading and validation succeed;
+- each destination is derived from its source; and
+- each explicit `required` value is preserved.
+
+Using the compact flow-style representation that exposed the integration
+failure is reasonable, while the existing block-style cases continue to cover
+omitted `required`. The test should exercise the public configuration behavior,
+not ConfigKit call choreography or Narratio's private recording mechanism.
+
+No network service, object store, credential, or test double is required.
+
+## Target End State
+
+All valid publish rules with derivable sources behave consistently across
+these cases:
+
+| `dest` | `required` | Result |
+| --- | --- | --- |
+| omitted | omitted | destination derived; required defaults to `true` |
+| omitted | `true` | destination derived; explicit `true` retained |
+| omitted | `false` | destination derived; explicit `false` retained |
+| explicit | either value | explicit destination and requiredness retained |
+
+`narratio config validate`, `narratio config show`, and ordinary pipeline
+loading accept the corrected cases without an unrecorded-change diagnostic.
+No operator configuration migration or workaround is required.
+
+## Release
+
+This is a backward-compatible correction to documented configuration behavior
+and should be released as Narratio `v2.1.1`. The release note should identify
+the corrected explicit-required/derived-destination interaction and state that
+no special upgrade action is required.
+
+## Acceptance Criteria
+
+- The focused regression test fails against v2.1.0 behavior and passes after
+  the fix.
+- Publish rules with explicit `required: true` or `required: false` and omitted
+  derivable destinations load and validate successfully.
+- Derived destinations and explicit requiredness appear correctly in the safe
+  effective configuration.
+- Existing destination derivation, publish validation, imports, profiles, and
+  configuration inspection tests remain green.
+- The complete repository validation and release-candidate checks pass.
+- The correction is published as the immutable `v2.1.1` tag with the required
+  release note.

docs/roadmap/implementation.md

@@ -0,0 +1,431 @@
+# Implementation Plan: Speaker Checkpoints and Replayable Uploads
+
+Status: Implementation complete; Stages 1-7 complete.
+
+## Scope and Instructions
+
+Implement [the feature roadmap](transcribe-checkpoints-and-replay.md). This plan
+replaces the completed Notarius-profile plan; that work remains in Git history.
+Complete the numbered stages in order, one bounded implementation prompt per
+stage for gpt-5.6-terra. Record completion and actual verification results here.
+Intermediate stages are not a finished feature and must not be released alone.
+
+Before working, read `AGENTS.md`, `docs/development.md`, all three policies in
+`docs/policy/`, the roadmap, and the stage-specific preparation below. Use the
+maintained Go 1.26 toolchain. Preserve operator choices, speaker filenames,
+bounded concurrency, streaming, and existing credentials/settings. Do not add
+new commands, dependencies, service requirements, arbitrary speaker limits,
+strict transcript schemas, or a generic checkpoint framework. No release or
+upstream service work is part of this plan.
+
+## Settled Design Decisions
+
+### Progress Representation and Recovery
+
+Add optional `transcribe_state_version` (version 1) and `transcribe_speakers`
+to `manifest.StageRecord`. The latter is a map keyed by the existing derived
+speaker ID, containing completed results only. Absence means that work must be
+done; a separate pending/running/failed state machine is unnecessary. Aggregate
+stage status/error remains authoritative for invocation failure. A versioned
+empty collection is meaningful: no earlier completions are reusable.
+
+Use a typed `TranscribeSpeakerRecord` with fingerprint version/digest,
+session-relative audio path, audio SHA-256 and size, one `ArtifactRecord` output
+of kind `transcript_raw`, output size, and producer run ID. The output carries
+its session-relative canonical path and SHA-256. No new transcript source IDs,
+synthetic schema contracts, timestamps, or historical attempt collections are
+needed. Existing aggregate diagnostics can retain attempt/status/duration data
+for work performed in the current invocation; do not fabricate that information
+for reused outputs.
+
+Keep the session-manifest schema identifier unchanged for this optional additive
+state; validate supported per-stage versions and retain strict decoding. New
+Narratio reads old manifests. Old binaries may reject the added fields under
+their existing strict decoding; document this downgrade limitation.
+
+For pre-feature manifests with no speaker evidence, regenerate transcription
+once when transcribe is next selected, even if the aggregate status succeeded.
+Old records do not bind individual transcript results to the audio bytes they
+used. Do not guess that association from current files, old `per_file` metadata,
+or abandoned run directories. This is the conservative upgrade path, not a new
+general migration system. No other stage or unselected range is forcibly run.
+
+### Reuse and Persistence
+
+Compute a versioned per-speaker fingerprint from speaker ID, session-relative
+audio identity, actual audio SHA-256/size, and the existing transcribe semantic
+configuration fingerprint (URL and language). Do not include invocation ID,
+absolute workspace root, concurrency, retry settings, timeout, or credentials.
+Stream input hashing with cancellation; do not trust stored input checksums
+alone. Recheck input bytes before accepting a new result to catch changes
+observed during transcription; fail that job without checkpointing mismatched
+evidence. This is not a hostile-filesystem snapshot or concurrent-editor
+transaction guarantee and requires no extra audio copies or file locks.
+
+Validate reusable output identity, canonical location, JSON, checksum, and size.
+Use existing confinement/file primitives. Missing or corrupt regular outputs
+become work; unsafe paths/non-regular conflicts follow existing safety errors,
+not recursive deletion. Missing required audio is an input error, not permission
+to reuse an unverifiable transcript. Removed inputs are removed from the current
+collection and aggregate outputs without deleting arbitrary files.
+
+One stage-owned mutex serializes candidate collection construction, manifest
+save, and installation of committed state. Save a detached candidate and only
+then update live authority, following analyze's pattern. Workers must not mutate
+the live manifest outside that serialized section. Keep expensive HTTP work
+outside the mutex. Stop dispatch/acceptance on checkpoint failure, join workers
+before return, and preserve the last successfully committed collection.
+
+Before launching requests, durably checkpoint the reconciled collection. For
+force, that collection is empty: this prevents later ordinary retries from
+resurrecting pre-force outputs. If this initial save fails, do not start requests
+or overwrite outputs. Force again clears the current collection again; ordinary
+retry retains replacements already checkpointed after the previous force.
+
+Save each validated result at its canonical raw-transcript location before
+checkpointing it. Run-local files remain staging/diagnostics, not reuse
+dependencies. Bytes left by an unsuccessful checkpoint have no completion
+authority. A worker may checkpoint a completed result using the parent stage
+context even if a sibling canceled the work context; parent cancellation or a
+checkpoint failure may prevent that save. Guarantee retention of committed
+work, not all responses that happened to finish during cancellation.
+
+### Multipart Replay Mechanism
+
+Replace the pipe/producer upload mechanism with a small pull-based multipart
+body backed by the audio reader. Use `mime/multipart.Writer` to encode a bounded
+metadata prefix and suffix around the file payload: capture the prefix after
+`CreateFormFile`, then encode the language field and closing boundary as the
+suffix. Generate this framing once per application attempt. Each body combines
+fresh readers for prefix, freshly opened audio, and suffix using `io.MultiReader`.
+No full-audio buffer or temporary multipart spool is needed.
+
+The initial body and `Request.GetBody` use the same factory/framing. Each owns
+an independent audio handle. Give bodies idempotent Close and context-triggered
+closure (for example `context.AfterFunc`), without holding a mutex across a
+blocking Read. Read/Close coordination must release a blocked source and retain
+genuine source-read/close errors while not treating deliberate early closure as
+source failure. The attempt owns cleanup of all created bodies, including
+replays; reject new body creation after its context expires.
+
+Remove the old producer `Wait` dependency. Track body-local source errors with
+synchronization so a swallowed upload read error cannot become a successful
+transcript, and a deliberately closed original body cannot poison a successful
+replay. Keep the outer retry policy and per-attempt context. Do not add replay
+counters to the user retry budget or depend on Go's exact internal retry count.
+Preserve ordinary HTTP redirect policy; body replay can also enable Go's
+body-preserving redirects. Do not introduce a new redirect restriction.
+
+## Stage 1: Add Durable Speaker Completion State
+
+Status: Complete.
+
+Completion marker (2026-09-21): Added optional versioned transcribe-owned
+per-speaker completion records, strict load/save validation, isolated collection
+cloning, legacy aggregate-only compatibility, and lifecycle-preservation
+coverage. Verified with `go test ./internal/manifest` and `git diff --check`.
+
+Preparation: `internal/manifest/manifest.go`, `store.go`, `analyze_state.go`,
+their tests, and `docs/internal/manifest.md`.
+
+1. Implement the additive fields, record type, version constants, validation,
+   and deep-copy helper described above in a focused transcribe state file.
+2. Validate ownership by the transcribe stage, supported versions, required
+   completed-record evidence, output kind, safe relative paths, SHA-256 values,
+   sizes, and producer identity consistency. Use existing speaker IDs as-is;
+   do not apply analysis-artifact key regexes or copy unrelated analysis limits.
+3. Integrate validation with both load and save. Preserve speaker state through
+   running/failed/stale transitions and aggregate result clearing. Generic stage
+   invalidation does not discard checkpoints; reconciliation checks their data.
+4. Test round-trip populated and versioned-empty collections, old manifests
+   without fields, malformed evidence, deep-copy isolation, and lifecycle
+   preservation. Avoid imposing JSON formatting or map order as contracts.
+
+Verify: `go test ./internal/manifest`.
+
+Done when durable storage can represent partial completion independently of
+aggregate status, without activating new stage behavior yet.
+
+## Stage 2: Implement Read-Only Speaker Reconciliation
+
+Status: Complete.
+
+Completion marker (2026-09-21): Extracted shared deterministic transcribe job
+construction and added read-only, cancellation-aware per-speaker input and
+output reconciliation. Covered unchanged reuse, semantic and operational
+settings, changed/added/removed inputs, legacy and forced work, missing/corrupt
+outputs, unsafe evidence, and missing audio without connecting runtime skip or
+checkpoint behavior. Verified with
+`go test ./internal/stage ./internal/manifest` and `git diff --check`.
+
+Preparation: `internal/stage/transcribe.go`, `semantic_contracts_initial.go`,
+`analyze_reconciliation.go`, `extract_resume.go`, their tests, and the artifact
+path/integrity helpers they use.
+
+1. Extract the existing audio discovery/job construction into a shared,
+   transcribe-local reconciliation path. Preserve current input selection and
+   speaker identity behavior. Build sorted jobs and classify completed results
+   against the current input set and evidence defined above.
+2. Add pure fingerprint construction and streamed input/output inspection.
+   Resolve manifest input paths with existing session-relative/absolute reading
+   helpers; do not change input precedence or assume the current working directory
+   is the session root. Newly written checkpoint paths are always relative.
+3. Return current records and work-to-do without writing files or manifest state.
+   Force selects all work and an empty starting collection; old unversioned
+   transcribe state supplies no reusable records. Do not connect aggregate skip
+   validation until execution can produce the new evidence in Stage 3.
+4. Test unchanged reuse; one changed audio among unchanged speakers; URL/language
+   changes; operational-setting exclusions; added/removed speakers; missing and
+   corrupt regular outputs; unsafe record paths; missing audio; and force.
+   Prefer compact tables and real temporary files over mocked hashing.
+
+Verify: `go test ./internal/stage ./internal/manifest`.
+
+Done when one read-only decision path describes exactly which speakers need
+work, without altering existing runtime behavior prematurely.
+
+## Stage 3: Execute and Checkpoint Speakers Incrementally
+
+Status: Complete.
+
+Completion marker (2026-09-21): Connected reconciliation to bounded execution,
+durably checkpointed the starting and per-speaker collections through detached
+manifest candidates, preserved committed work across failures and cancellation,
+rebuilt deterministic aggregate outputs from reusable evidence, and added
+speaker-aware resume validation. Updated production-style fixtures and covered
+serialized concurrent saves, force/retry replacement, input rechecks, save
+failures, partial recovery, and aggregate reconstruction. Verified with
+`go test ./internal/stage ./internal/app`, `go test -race ./internal/stage`, and
+`git diff --check`.
+
+Preparation: `internal/stage/transcribe.go`, `analyze_execution_state.go`,
+`run_local.go`,
+`internal/app/runner.go`, and transcribe/checkpoint test fixtures.
+
+1. Wire reconciliation into Run, require the existing ManifestStore collaborator,
+   and checkpoint the starting collection before launching HTTP work. Implement
+   detached candidate saves and serialized live-state updates. Audit Store.Save's
+   normalization side effects so a failed candidate save cannot mutate shared
+   authority; use detached records/collections wherever it may write.
+2. Retain bounded errgroup execution and fail-fast cancellation. Check the work
+   context before beginning each queued request. After success, validate adapter
+   output path/JSON, recheck audio evidence, materialize canonical bytes, compute
+   output evidence, and checkpoint the new record. Check a checkpoint-failure
+   latch under the serialization lock before accepting further completions.
+3. Preserve committed successes on all error/cancellation exits, including
+   starting-save, canonical-write, per-speaker-save, and aggregate-save failures.
+   Return meaningful speaker/error context and let the runner retain its existing
+   terminal failure behavior. Do not report partial aggregate outputs as success.
+4. Assemble final outputs in deterministic speaker order from the reconciled
+   completed records, including reused speakers, with original producer IDs and
+   checksums. If all records are current but aggregate status is failed/running,
+   finish without WhisperX calls and rebuild the aggregate success.
+5. Implement `transcribeStage.ValidateResume` using the same read-only
+   reconciliation, so a succeeded stage cannot skip changed audio or invalid
+   transcripts. Retain the existing URL/language semantic check and runner
+   invalidation behavior; no whole-config digest invalidation.
+6. Adapt existing tests that require no canonical partial files: the new contract
+   permits checkpointed files but forbids aggregate success. Update fixtures to
+   provide real manifest stores and records rather than a production fallback
+   that silently disables checkpointing.
+7. Test A/B completion followed by C failure and D unattempted (concurrency 1),
+   reload and retry only C/D; concurrently finishing speakers without lost saves;
+   failure and cancellation after a checkpoint; save failure retaining earlier
+   authority; and force failure followed by ordinary retry without old-result reuse.
+
+Verify: `go test ./internal/stage ./internal/app` and
+`go test -race ./internal/stage`.
+
+Done when incremental progress and aggregate success work together safely under
+ordinary, concurrent, canceled, and forced execution.
+
+## Stage 4: Verify Runner, Merge, and Durable Lifecycle Integration
+
+Status: Complete.
+
+Completion marker (2026-09-22): Added real-persistence application coverage for
+cross-invocation retry, interrupted running state, failed aggregate saves,
+force/reset persistence, downstream invalidation, and one-time legacy upgrade.
+Versioned merge input discovery now requires aggregate success and an exact
+match between current speaker evidence and declared aggregate outputs, while
+legacy/manual directory fallback remains available. Restore coverage preserves
+session-relative speaker paths, leaves unpublished raw transcripts absent, and
+canonical transcripts remain reusable after run-directory cleanup. Verified
+with `go test ./internal/app ./internal/stage ./internal/manifest`,
+`go test -race ./internal/app`, and `git diff --check`.
+
+Preparation: `internal/app/runner.go`, `runner_runtime.go`,
+`analyze_checkpoint_recovery_test.go`, `runner_terminal_failure_test.go`,
+`restore_manifest_paths.go`, `internal/stage/merge.go`, and existing publish,
+restore, and cleanup tests routed from `docs/internal/overview.md`.
+
+1. Add application-level tests with real manifest persistence and fake WhisperX:
+   failed stage across separate invocations; all speakers committed before a
+   failed aggregate save; interrupted running state; force/reset persistence;
+   and unchanged upstream/succeeded downstream invalidation behavior. Inject
+   errors at durable boundaries, not every private method call.
+2. Close `discoverRawTranscripts`' directory-scan fallback for versioned
+   transcribe state: incomplete aggregate state is not usable, and completed
+   state consumes only declared current aggregate outputs. Keep legacy/manual
+   fallback behavior for manifests without the new state rather than banning
+   previously allowed standalone workflows. Test that partial files and removed
+   speaker files cannot leak into a versioned merge, including direct stage calls.
+3. Audit manifest copies/projections, remote manifest publication, restored path
+   handling, and cleanup. Preserve relative speaker paths on restore and verify
+   they resolve under the destination session, not the producer machine. Reuse
+   existing confinement validation, not a second path-policy layer.
+4. Add focused round-trip coverage through existing publish/restore fixtures for
+   speaker records. Restored outputs absent from the published selection are
+   non-reusable locally; do not silently expand remote publication to every raw
+   speaker or upload audio as a new default. Removing run directories must not
+   remove canonical checkpointed transcripts or make reuse require run files.
+5. Verify old succeeded manifests are readable and rerun once when transcribe is
+   selected, then reuse normally. Keep old manifests usable for other commands
+   and bounded ranges. Do not automatically import old unrecorded JSON files.
+
+Verify: `go test ./internal/app ./internal/stage ./internal/manifest` and
+`go test -race ./internal/app`.
+
+Done when the new authority survives runner and storage lifecycles and cannot
+expose incomplete transcripts downstream.
+
+## Stage 5: Implement Replayable Streaming Multipart Bodies
+
+Status: Complete.
+
+Completion marker (2026-09-22): Replaced the producer-goroutine/pipe upload
+with bounded standard-library multipart framing and independently opened,
+pull-based streaming bodies shared by the initial request and `GetBody`.
+Bodies are tracked, idempotently closed, canceled with their attempt context,
+and retain genuine read/close failures without treating an intentionally
+abandoned original as a replay failure. Added focused coverage for identical
+framing, partial-read restart, simultaneous handles, source failures, and
+blocked-reader release. Verified with `go test ./internal/adapters/whisperx`,
+`go test -race ./internal/adapters/whisperx` (including ten repeated runs), and
+`git diff --check`.
+
+Preparation: `internal/adapters/whisperx/http.go`, `http_test.go`,
+`docs/integrations/whisperx.md`, and the installed Go net/http Request/GetBody
+contract. Keep this change inside the WhisperX adapter.
+
+1. Implement the pull-based framing/body factory described above, using standard
+   multipart encoding rather than hand-escaped MIME headers. Add no new payload
+   cap or upload-size setting. Preserve file-before-language wire semantics and
+   the existing form names.
+2. Install the initial body and GetBody with the same per-attempt boundary.
+   Track and close all opened bodies even if a custom transport returns early;
+   make cleanup/context callbacks safe against overlapping reads and closes.
+   A GetBody failure must be returned, not converted into an empty upload.
+3. Replace pipe-specific Wait/error paths in doTranscribeAttempt with source-error
+   handling appropriate to pull bodies. Genuine source errors remain visible;
+   intentional closure during transport retry is not itself a source error.
+   Keep response bounds, JSON validation, output installation, retry
+   classification/delay, cancellation, and application attempt reporting intact.
+4. Remove obsolete producer-only helpers rather than retaining two upload paths.
+   Adapt tests tied to producer goroutines to assert streaming and resource
+   release, which are the actual contracts.
+5. Add focused body tests: original/replayed bytes parse identically; replay
+   after partial consumption starts from the beginning; simultaneous bodies use
+   independent handles; source failures propagate; Close/cancel releases a
+   blocked reader without deadlock. Use synthetic audio and gated readers.
+
+Verify: `go test ./internal/adapters/whisperx` and
+`go test -race ./internal/adapters/whisperx`.
+
+Done when replay is available without whole-file buffering or producer-lifetime
+ambiguity, and all existing HTTP behavior tests still pass.
+
+## Stage 6: Verify Replay Through the HTTP Adapter Boundary
+
+Status: Complete.
+
+Completion marker (2026-09-22): Added a deterministic transport regression
+that abandons a partially consumed original request, obtains and parses a fresh
+multipart replay from `GetBody`, and installs the returned transcript within one
+application attempt. Covered replay open/read failures, early responses,
+parent cancellation during replay, and disposal of every opened source, while
+proving an abandoned original does not poison a valid replay. Added a real HTTP
+retry-exhaustion integration that durably retains an earlier speaker and resumes
+only the uncommitted speakers. This controlled boundary verifies Narratio's
+replay contract, not recovery from every real proxy protocol failure. Verified
+with `go test ./internal/adapters/whisperx ./internal/app ./internal/stage`,
+`go test -race ./internal/adapters/whisperx ./internal/app ./internal/stage`,
+ten repeated race-enabled adapter runs, and `git diff --check`.
+
+Preparation: Existing injectable HTTPClient/RoundTripper seam and local-server
+tests in `internal/adapters/whisperx/http_test.go`; app WhisperX wiring tests.
+
+1. Add a deterministic transport-boundary regression: consume part of the
+   original request, close it as a transport would, call GetBody, consume and
+   parse the replay using the request Content-Type, then return valid JSON.
+   Verify exact audio/language, installed output, and one application attempt.
+   This must fail before replay support; merely checking non-nil GetBody is
+   insufficient. Do not assert Go's internal retry count or require a flaky
+   handcrafted HTTP/2 peer in the default suite.
+2. Cover replay-open/read failure, early response, parent cancellation during
+   replay, and disposal of all body handles. Reuse Stage 5 body coverage rather
+   than repeat every low-level case. Prove an intentionally abandoned original
+   body does not turn a valid replay response into failure.
+3. Retain local HTTP tests for application retries, non-retryable responses,
+   response limits, invalid JSON, and bounded streaming; exercise that an
+   exhausted HTTP failure still leaves earlier speaker checkpoints reusable
+   using the existing stage/app seams.
+4. Record the verification boundary honestly: controlled transport replay proves
+   Narratio's body contract, not recovery from every real proxy PROTOCOL_ERROR.
+
+Verify: `go test ./internal/adapters/whisperx ./internal/app ./internal/stage`
+and race tests for changed packages.
+
+Done when the replay defect is protected at the adapter boundary and the
+application retry/checkpoint behavior remains intact.
+
+## Stage 7: Document and Validate the Completed Feature
+
+Status: Complete.
+
+Completion marker (2026-09-22): Updated the canonical transcribe, merge,
+manifest, WhisperX, operations, and troubleshooting documentation for durable
+speaker reconciliation, serialized checkpoints, aggregate gating, durable
+force reset, legacy upgrade and downgrade limits, replayable streaming bodies,
+resource ownership, and the absence of an exactly-once remote guarantee.
+Reviewed the completed implementation against the feature roadmap and
+implementation-discipline guidance. Verified with `go test ./...`,
+`go test -race ./...`, `go vet ./...`, `go build ./...`,
+`go test ./internal/doccheck`,
+`go test ./internal/config -run '^TestExamplesLoadAndValidate$'`, and
+`git diff --check`.
+
+1. Update canonical transcribe and manifest internals for the collection,
+   reconciliation, durable-reset force behavior, serialized saves, and aggregate
+   gate. Remove obsolete all-speakers-before-canonical-write descriptions.
+2. Update the WhisperX contract for replayable bounded streaming, application
+   attempts versus transport replay, resource ownership, and the lack of an
+   exactly-once server guarantee. Do not claim the remote protocol issue is fixed.
+3. Update operations/troubleshooting with ordinary resume versus force, durable
+   partial success, old-manifest one-time retranscription on selecting transcribe,
+   downgrade limitations, and private-model-change guidance. No new CLI flags
+   or promises of automatic historical-file recovery. Keep detailed wire and
+   storage contracts in their canonical owners.
+4. Review changes against the feature roadmap and the implementation-discipline
+   guidance. Do not elevate selected storage encoding, internal ordering, test
+   synchronization, or exact error wording into new product guarantees.
+5. Run the full repository gates below; document results and update completion
+   statuses only after they pass. Leave release preparation for a separate task.
+
+```sh
+go test ./...
+go test -race ./...
+go vet ./...
+go build ./...
+go test ./internal/doccheck
+go test ./internal/config -run '^TestExamplesLoadAndValidate$'
+git diff --check
+```
+
+## Open Questions
+
+None requiring user input. The agreed roadmap supplies the product decisions;
+this plan selects a completed-only manifest collection, conservative old-record
+handling, and pull-based replay as implementation choices. No stronger remote
+reliability guarantee, new operator restriction, or generic orchestration
+surface is proposed.

docs/roadmap/notarius-config-profile.md

@@ -0,0 +1,151 @@
+# Notarius Configuration Profile Selection
+
+Status: Implemented.
+
+## Goal and Authority
+
+Implement the consumer behavior in the
+[feature request](feature-request-notarius-config-profile.yml): let a Narratio
+configuration overlay explicitly select a Notarius configuration overlay.
+Operators can select testing once in Narratio while each application retains
+ownership of its own model choices and other settings.
+
+This roadmap describes the target state and required change areas, not an
+implementation sequence; the [implementation plan](implementation.md) owns
+the ordered work and implementation decisions. The feature request is the scope
+authority; the mechanisms suggested below do not establish additional
+user-facing contracts.
+
+## Underlying Issue
+
+Narratio can select its own testing profile, but extraction currently invokes
+Notarius without a configuration-profile selector. Consequently, a Narratio
+testing invocation cannot explicitly select the independently deployed Notarius
+testing overlay through pipeline configuration.
+
+Adding individual downstream model or tuning fields would duplicate Notarius's
+configuration responsibilities. Passing its existing profile selector provides
+the requested flexibility without that duplication.
+
+## Targeted End State
+
+### Explicit Configuration Composition
+
+`pipeline.notarius.config_profile` is an optional Notarius configuration-profile
+name, usable in Narratio's base configuration and ordinary overlays. For
+example, Narratio's `profiles/testing.yml` may contain:
+
+```yaml
+notarius:
+  config_profile: testing
+```
+
+Selecting Narratio's `--profile testing` then selects that explicit downstream
+profile whenever extraction executes. The two profile names need not match.
+Narratio does not automatically forward its own profile name.
+
+When supplied, the selection is passed through Notarius's `--config-profile`
+argument. Omitted or null values add no profile-selection flag and preserve
+existing downstream selection behavior. Notarius's selector semantics remain
+authoritative, including empty and `default` selections for the composed base
+configuration. Narratio does not introduce a stricter profile-name policy.
+
+This field selects a ConfigKit configuration overlay, not a PromptKit profile
+through `--llm-profile`. The Notarius overlay owns any resulting LLM-profile
+selection. Notarius resolves overlays relative to its own configuration root;
+Narratio neither resolves those files against its own root nor rewrites them.
+
+### Execution, Reuse, and Failure
+
+A changed downstream profile selection must not silently reuse Narratio's
+completed extraction from the previous selection. Existing dependent-stage
+invalidation applies when extraction changes; the feature does not change the
+pipeline order or transcript-generation behavior.
+
+Notarius continues to own downstream overlay discovery, composition, validation,
+and cache compatibility. A missing or invalid selected overlay must fail
+extraction through the existing subprocess error path, without silently falling
+back to production or reporting successful extraction.
+
+Existing resume and chunk-cache controls, including force precedence, remain
+in place. Profile selection does not itself mean forced regeneration or cache
+deletion. Narratio's decision to execute extraction and Notarius's decision to
+reuse compatible work remain separate responsibilities.
+
+Changes to private Notarius configuration behind an unchanged selector retain
+the existing operator responsibility to force extraction when needed. There is
+no new downstream file inspection or content-tracking requirement.
+
+Credential forwarding remains independently configured through
+`notarius.llm_api_key`. A Narratio testing overlay may select a different
+credential reference alongside the downstream profile; no new credential
+discovery or provider-name mapping is introduced.
+
+## Required Change Areas
+
+### Configuration and Inspection
+
+Extend the typed Notarius configuration and its existing safe effective-YAML
+projection to carry the optional field. Preserve the ordinary configuration
+composition behavior for base values, overlay overrides, and null values.
+Configuration inspection must reflect the selected field without exposing
+credential values. No replacement configuration loader or downstream preflight
+loader is needed.
+
+### Stage and Adapter Boundary
+
+Carry the selection from resolved configuration through the extraction request
+to the subprocess adapter. Construct the optional CLI flag using the existing
+argument-vector mechanism, alongside the current configuration path, references,
+credential forwarding, and cache controls. Retain normal Notarius error handling.
+
+### Extraction Reuse Evidence
+
+Account for the downstream selection in the existing extraction reuse evidence.
+The current semantic-configuration and invocation fingerprints are the relevant
+change points. Keep their treatment of selection consistent so that a profile
+change cannot bypass the required rerun and downstream invalidation.
+
+This roadmap does not require a new manifest schema, separate profile-indexed
+workspaces, historical per-profile result retention, or an upgrade-wide forced
+rerun.
+
+### Verification and Documentation
+
+Use focused behavioral coverage for configuration composition and inspection,
+explicit argument forwarding and omission, changed-selection reuse decisions,
+and ordinary subprocess failure handling. Reuse existing test coverage where
+it already protects unchanged credentials, force behavior, and invalidation;
+do not duplicate Notarius's complete profile-validation or cache test suites.
+
+When implemented, update the canonical
+[configuration reference](../config.md),
+[Notarius integration contract](../integrations/notarius.md),
+[extract internals](../internal/stage-extract.md), and relevant maintained
+examples. The development workflow in [Operations](../operations.md) should
+explain explicit profile composition and the existing need to force extraction
+after private downstream edits. Clarify that profile selection does not create
+an isolated testing workspace or separate current extraction state.
+
+## Compatibility and Scope Boundaries
+
+Notarius v0.7.0 already documents `--config-profile`; no upstream change is
+required. Document the capability needed when the option is used, without adding
+an unconditional runtime version gate for existing unconfigured deployments.
+
+This feature adds no Narratio CLI selector, automatic profile propagation,
+Notarius LLM-profile selector, downstream configuration rewriting, new
+credential mechanism, or generalized recursive orchestration. Existing output
+and receipt contracts remain unchanged.
+
+The consumer's Ansible role and suite updates are follow-up work outside this
+repository. They can deploy downstream overlays and populate this field after
+the interface is released.
+
+## Scope Review
+
+No stronger requirements or new invariants are proposed beyond the feature
+request and existing repository contracts. In particular, minimizing one-time
+upgrade invalidation and avoiding reruns between equivalent base selectors are
+implementation considerations, not additional acceptance guarantees imposed
+by this roadmap. No user decisions are currently required before planning.

docs/roadmap/remediation.md

@@ -0,0 +1,104 @@
+# Transcription Checkpoint Review Remediation
+
+Status: Complete.
+
+Reviewed commits `7b0bf99` through `96b4e73` against the speaker-checkpoint and
+upload-replay roadmap and the original user request. No production changes were
+made during review. The two failures below were reproduced with temporary tests
+using the existing stage fixtures; those temporary tests were removed afterward.
+
+## 1. Reconcile Aggregate State Even When No Speaker Needs HTTP Work
+
+Priority: P2. Status: Complete.
+
+Implemented read-only aggregate identity/evidence checks before resume skips.
+Regression tests cover removed inputs, missing/duplicate/mismatched outputs,
+repair without HTTP requests, downstream invalidation, and exclusion of removed
+speakers from merge without deleting their files. Equivalent path forms and
+output ordering remain reusable. Focused app/stage tests and documentation
+checks pass.
+
+Location: `internal/stage/transcribe_resume.go:21` and the reconciliation result
+consumed there.
+
+`ValidateResume` returns reusable whenever `reconciliation.work` is empty. That
+only establishes that the remaining speakers need no transcription; it does
+not establish that the stored speaker collection and aggregate outputs match
+the current input set. Removing a speaker while the remaining inputs are
+unchanged produces an empty work list, so a succeeded record can be skipped
+without pruning the removed speaker. Its old aggregate output can then remain
+available to merge. Likewise, missing/inconsistent aggregate output references
+are not repaired when all individual checkpoints remain valid. A prepare rerun
+may already stale transcribe in some workflows, but the resume validator cannot
+rely on that to satisfy its own current-input/output checks.
+
+Reproduction: create complete alice/bob checkpoint records and aggregate
+outputs with the existing fixture, mark the stage succeeded, remove bob from
+the manifest's audio inputs, and call `ValidateResume`. It returns reusable
+despite still retaining bob in the stored collection and aggregate outputs.
+
+Required change:
+
+- Distinguish no HTTP work from no reconciliation work. Validate that the
+  committed speaker set and aggregate transcript references represent the
+  reconciled current set, including relevant identity/evidence, before skipping.
+- When only collection/output repair is needed, return non-resumable and let
+  normal execution checkpoint the pruned collection and rebuild aggregate
+  outputs without retranscribing reusable speakers. Preserve existing dependent
+  invalidation; do not force otherwise-current speaker jobs.
+- Keep resume inspection read-only. Do not fix this by deleting incidental
+  transcript files or changing stage selection rules.
+
+Verification: add removed-only input coverage (not just removal combined with
+an added speaker), plus incomplete/inconsistent aggregate output coverage.
+Exercise the runner skip path and assert zero additional WhisperX requests,
+correct rebuilt outputs, and no removed speaker reaching merge.
+
+## 2. Remove the Unrequested Positive Audio-Size Requirement
+
+Priority: P2. Status: Complete.
+
+Audio evidence now permits zero size while still rejecting negative values.
+Regression tests verify zero-size manifest round trips, successful service
+results being checkpointed and reused, and service rejection being propagated
+without creating a completion record. Nonempty valid-JSON output validation is
+unchanged. Both new success regressions failed before the fix.
+
+Location: `internal/manifest/transcribe_state.go:95`.
+
+The new record validator requires `AudioSize > 0`. Previously Narratio accepted
+a regular `.flac` file and delegated transcription acceptance to WhisperX;
+neither the user request nor the roadmap introduced minimum audio size or a new
+audio-content validation policy. Zero-byte content still has a valid checksum
+and can be represented by the checkpoint mechanism. The current check runs
+after a successful response and output materialization, causing the entire
+stage to fail instead of retaining the completed work.
+
+Reproduction: prepare one speaker, replace its prepared audio with an empty
+regular file, and use the existing successful fake WhisperX client. Stage
+execution fails at checkpoint validation with `audio_size must be positive`.
+This demonstrates the extra Narratio restriction; it does not assert that a
+particular production WhisperX deployment accepts empty audio.
+
+Required change:
+
+- Permit zero audio size and reject negative sizes. Keep byte hashing and
+  size equality checks. Let the downstream service decide whether the input
+  can be transcribed, as before.
+- Do not remove the nonempty valid-JSON output requirement, which has a separate
+  existing justification. Do not add a new early empty-audio rejection instead.
+- Update any new tests/documentation that promote strictly positive input size
+  to a contract.
+
+Verification: cover a zero-size record round-trip and successful adapter result
+being checkpointed/reused for that input; retain negative-size rejection and
+ordinary downstream-failure tests.
+
+## Verification After Remediation
+
+Completed on 2026-09-22 with Go 1.26.8: focused stage/application tests,
+changed-package race tests, `go test ./...`, `go test -race ./...`,
+`go vet ./...`, `go build ./...`, `go test ./internal/doccheck`,
+`go test ./internal/config -run '^TestExamplesLoadAndValidate$'`, and
+`git diff --check` all passed. Existing replay and cancellation regression
+tests remain intact; the HTTP adapter required no changes.

docs/roadmap/transcribe-checkpoints-and-replay.md

@@ -0,0 +1,183 @@
+# Resumable Speaker Transcription and Replayable Uploads
+
+Status: Implemented.
+
+## Goal and User Intent
+
+Preserve completed speaker transcriptions when another speaker fails, so an
+ordinary run or resume does not repeat expensive successful work. Separately,
+allow Go's HTTP transport to replay WhisperX uploads when it encounters a
+recoverable connection or stream failure.
+
+The motivating incident exhausted four transcription attempts for one speaker
+after three other speakers had completed. The final error reported an HTTP/2
+peer `PROTOCOL_ERROR` and an inability to replay the request body without
+`Request.GetBody`.
+
+These are related reliability improvements, not a promise to eliminate all
+service failures. This document owns the target state and scope; the
+[implementation plan](implementation.md) owns ordered changes and concrete
+storage and replay decisions.
+
+## Pre-implementation Limitations
+
+The [transcribe stage](../internal/stage-transcribe.md) writes successful
+responses into the invocation's run directory. It installs canonical transcripts
+and returns its per-file metadata only after every speaker succeeds. A failure
+therefore leaves no durable per-speaker progress that ordinary resume can use,
+even though some completed JSON files may remain on disk.
+
+The [WhisperX adapter](../integrations/whisperx.md) already retries requests at
+the application level, reopening audio for each attempt. Its multipart request
+body is streamed through a pipe and has no replay callback. Transport-level
+recovery within an attempt cannot reconstruct a consumed body. Providing that
+capability addresses the replay limitation, not the underlying peer protocol
+error; persistent failures can still require service or proxy investigation.
+
+## Targeted End State
+
+### Durable Per-Speaker Completion
+
+Each successfully transcribed speaker file is validated, saved, and recorded
+durably before that work is treated as reusable. The session manifest remains
+the progress authority. Follow the existing analyze-stage checkpoint pattern
+without turning speakers into pipeline stages or introducing a general task
+engine.
+
+Ordinary execution reconciles the current prepared audio set with recorded
+speaker results. It reuses valid completed work and schedules failed,
+unattempted, missing, or obsolete results. A result is not reusable merely
+because a matching filename exists. Evidence binds it to the speaker/audio
+identity, audio content, Narratio-visible recognition settings, and the saved
+transcript's integrity.
+
+The relevant recognition settings remain the WhisperX service URL and language.
+Retry count/delay, concurrency, timeout, credentials, and diagnostic locations
+remain operational settings rather than reasons to discard completed results.
+Changes to private service models or configuration remain an operator reason
+to force transcription; Narratio does not inspect the service's private state.
+
+Changes to individual audio inputs require replacement of affected results,
+not automatic retranscription of unrelated unchanged speakers. New speakers
+need transcription; removed speakers do not remain in the current aggregate
+output set. Records must describe the inputs actually used, not grant reuse to
+old bytes under a changed identity.
+
+### Aggregate Completion, Failure, and Force
+
+The overall stage succeeds only when every currently required speaker has a
+validated, durably recorded result. Incomplete per-speaker progress does not
+make the aggregate stage successful or authorize downstream consumption of an
+incomplete collection. Existing dependent-stage invalidation remains in effect.
+
+Retain bounded concurrency and the current fail-fast behavior: a speaker
+failure cancels outstanding work. Results already checkpointed remain reusable
+after failure, cancellation, or process interruption. A later ordinary run or
+resume retries the remaining work; it does not need a new CLI command.
+
+Checkpoint writes must be serialized so concurrently finishing speakers cannot
+overwrite each other's progress. A failed persistence operation must not
+advertise an unrecorded result as complete. Previously committed progress remains
+the recovery authority, and checkpoint errors must be surfaced rather than
+reported as successful transcription.
+
+`--force` on transcribe regenerates all currently required speakers instead of
+reusing their prior completions. Successful replacements are checkpointed too,
+so an ordinary retry after a failed forced invocation can retain the newly
+completed work. Old results for speakers not successfully regenerated must not
+satisfy that unfinished forced replacement on the next ordinary retry.
+
+This deliberately relaxes the current all-speakers-before-canonical-installation
+implementation: individual durable results may exist while the stage remains
+failed or incomplete. The all-speakers-before-stage-success boundary remains.
+
+### Replayable WhisperX Requests
+
+Provide `Request.GetBody` so the HTTP transport can obtain a fresh body for the
+same request. Replay reopens the audio and reproduces the multipart boundary,
+filename, fields, and framing used by the original request. Independent bodies
+must not share a consumed reader or mutable read position.
+
+Preserve bounded-memory streaming rather than buffering entire audio files in
+memory. Correctly close every original and replay body and its audio resources;
+preserve source-read errors and cancellation, including partial uploads and
+early server responses. Successful response handling must not accidentally wait
+on, or report errors solely from, an abandoned body instead of the replay that
+produced the response.
+
+Keep the existing application-level retry count, delay, classification, and
+per-attempt timeout semantics. Transport replays occur within an application
+attempt and its context; they are not additional Narratio-configured attempts.
+Do not add another custom retry loop, silently increase retries, or disable
+HTTP/2 globally as a substitute for replay support.
+
+The adapter still validates and installs only successful JSON responses.
+Request replay does not provide exactly-once server execution: a lost response
+can cause repeated transcription work, as can existing application retries.
+No new service-side idempotency protocol is required by this feature.
+
+## Required Change Areas
+
+- **Manifest and persistence:** Add transcribe-owned per-speaker progress and
+  evidence, with safe checkpoint updates and preservation through runner failure
+  handling. Review manifest validation, copying, serialization, and existing
+  publish/restore handling wherever the new records pass through those paths.
+- **Stage execution and reuse:** Reconcile speaker records before scheduling,
+  checkpoint validated completions, rebuild aggregate outputs from the current
+  set, and account for partial progress in ordinary and forced execution.
+  Review successful-stage reuse as well as failed-stage retries so stale audio
+  or missing/corrupt transcript evidence is not bypassed by an aggregate skip.
+- **Artifact lifecycle:** Use existing path, installation, integrity, and manifest
+  mechanisms. Ensure checkpointed results do not depend on an ephemeral run
+  directory surviving cleanup. Do not introduce a second authoritative ledger
+  or infer completion by scanning abandoned runs.
+- **HTTP adapter:** Make multipart bodies reproducible and correctly manage
+  every replay's lifetime, without changing the WhisperX HTTP schema or requiring
+  an upstream service change.
+- **Tests:** Cover per-speaker recovery after a later failure, interruption and
+  checkpoint failure; changed/missing inputs or outputs; force followed by an
+  ordinary retry; concurrent successful completions; and aggregate downstream
+  gating. Extend existing fixtures rather than duplicating analyze's complete
+  test suite. Exercise actual body reconstruction after partial consumption,
+  multipart equivalence, cancellation/resource cleanup, and existing bounded
+  streaming and retry behavior—not just the presence of a callback. Default
+  tests remain offline and use local/fake HTTP boundaries, never paid services.
+- **Documentation:** Update the canonical
+  [stage guide](../internal/stage-transcribe.md),
+  [manifest guide](../internal/manifest.md),
+  [WhisperX contract](../integrations/whisperx.md),
+  [operations guide](../operations.md), and
+  [troubleshooting guide](../troubleshooting.md) when implemented. Explain
+  partial completion, ordinary retry versus force, and the distinction between
+  request replay limitations and persistent peer protocol failures.
+
+## Compatibility and Scope Boundaries
+
+Existing speaker identities, raw transcript formats, and the fixed pipeline
+order remain unchanged. Existing manifests remain readable, but must not be
+treated as containing per-speaker evidence they never recorded. Recovery relies
+on verified recorded evidence, not guessed associations between old transcripts
+and current audio. No general legacy-migration framework is part of the feature.
+
+Automatic adoption of unrecorded outputs from previously failed runs is outside
+this feature. Such files may be recoverable after verifying their inputs, but
+their presence alone does not prove completion. No recovery/import command,
+speaker-selection CLI, historical invocation ledger, new service configuration
+discovery, or broader orchestration refactor is included.
+
+The feature does not diagnose the remote service/proxy's underlying HTTP/2
+problem, guarantee recovery from persistent protocol errors, or change the
+policy for continuing unrelated work after failure.
+
+## Completion Criteria
+
+For speakers A, B, C, and D: if A and B are checkpointed, C fails, and D has not
+completed, ordinary resume retains A and B, retries C, and completes D. No
+downstream stage receives an incomplete successful transcription collection.
+Force replaces all four, while an ordinary retry can reuse checkpoints from
+that forced replacement attempt.
+
+The HTTP transport can reconstruct the upload after consuming the original
+body, without whole-file memory buffering, lost source errors, or stranded
+upload resources. A persistent peer error still fails normally after the
+configured application attempts, without discarding checkpointed speakers.

docs/troubleshooting.md

@@ -56,7 +56,8 @@ Symptom:
 
 Likely cause:
 
-- stale field name, typo, invalid enum, or invalid duration/path format.
+- stale field name, typo, invalid enum, invalid duration/path format, or an
+  unsupported local manifest schema/state marker.
 
 Diagnostics:
 
@@ -68,9 +69,39 @@ Safe fix:
 
 - align config with [Configuration](./config.md) and the
   [maintained examples](../examples/README.md).
+- for a manifest schema error, move the unsupported local state aside and
+  regenerate it; restore a verified current committed snapshot or republish
+  healthy regenerated state when appropriate. Do not add version or stage
+  markers by hand.
+- if an older manifest is rejected for its retired cleanup-obligation field,
+  finish outstanding cleanup with the pre-upgrade binary. Otherwise verify the
+  remote commit, move aside or regenerate the incompatible manifest, and use
+  the cleanup stage. Do not treat manual JSON editing as the normal migration
+  path.
 
 Relevant reference: [Configuration](./config.md).
 
+## Party schema version rejected
+
+Symptom:
+
+- configuration loading reports that `party.schema_version` is required or
+  unsupported.
+
+Likely cause:
+
+- the campaign `party.yml` is an unversioned legacy document, or names a
+  schema other than `narratio.party.v1`.
+
+Safe fix:
+
+- replace the file with one exact, single-document
+  [`narratio.party.v1` party](./integrations/party.md); do not add a version
+  marker to an otherwise incompatible legacy structure. The versioned party is
+  the only configured roster authority.
+
+Relevant reference: [Canonical Party Input](./integrations/party.md).
+
 ## Unexpected imported or profile value
 
 Symptom:
@@ -171,8 +202,9 @@ Safe fix:
 
 - widen the bounded range to include the first reported stage, or recover that
   stage explicitly with `run-stage` before retrying. The failed check does not
-  create a run record or modify the manifest. Narratio does not resume-validate
+  create a run-local directory or modify the manifest. Narratio does not
-  excluded prefix stages, and stages after `--through` are not prerequisites.
+  resume-validate excluded prefix stages, and stages after `--through` are not
+  prerequisites.
 
 If prerequisite statuses are terminal but a selected stage reports a missing,
 unsafe, or checksum-inconsistent artifact, repair the artifact at the stage
@@ -181,6 +213,134 @@ input validation.
 
 Relevant reference: [Operations: Stage Execution and Continuation Behavior](./operations.md#stage-execution-and-continuation-behavior).
 
+## Transcription retries speakers that previously finished
+
+Symptom:
+
+- an ordinary transcribe retry calls WhisperX for a speaker that appeared to
+  finish earlier; or
+- an older successful session retranscribes all speakers once after upgrade.
+
+Likely causes:
+
+- the earlier response or canonical file was not followed by a successful
+  speaker manifest checkpoint;
+- the prepared audio bytes/path, WhisperX URL, or language changed;
+- the canonical transcript is missing, unsafe, non-regular, invalid JSON, or
+  no longer matches its recorded size/checksum;
+- the previous invocation used `--force`, which durably cleared pre-force
+  speaker authority; or
+- the manifest has only legacy aggregate transcribe state and cannot bind old
+  outputs to their source audio.
+
+Safe fix:
+
+- inspect the reported speaker and filesystem error, repair the prepared input
+  or canonical-output path when needed, then retry transcribe normally;
+- allow a legacy aggregate-only record to regenerate once; and
+- do not edit the manifest or copy an old JSON file into place as a substitute
+  for a successful checkpoint.
+
+```bash
+narratio run-stage transcribe 2026-04-04
+```
+
+Valid earlier speaker checkpoints remain reusable even while aggregate
+transcribe status is failed or interrupted. Merge remains gated until
+transcribe reaches aggregate success with the complete current set.
+
+Relevant reference: [Operations: Transcription continuation and replacement](./operations.md#transcription-continuation-and-replacement).
+
+## WhisperX model changed but transcription was reused
+
+Symptom:
+
+- WhisperX model weights, prompts, or private server configuration changed at
+  the same configured URL, but Narratio still considers speaker transcripts
+  current.
+
+Likely cause:
+
+- Narratio can fingerprint its configured URL and language, not private state
+  loaded behind the service boundary.
+
+Safe fix:
+
+- explicitly force transcribe to establish a durable reset, then use ordinary
+  retry if the replacement attempt is interrupted.
+
+```bash
+narratio run-stage transcribe 2026-04-04 --force
+```
+
+Transport replay and application retries do not provide an exactly-once remote
+execution guarantee. Investigate server-side failures or duplicate work using
+WhisperX diagnostics; do not infer completion from a server log without the
+local speaker checkpoint.
+
+Relevant references: [WhisperX integration](./integrations/whisperx.md) and
+[Transcribe internals](./internal/stage-transcribe.md).
+
+## Transcription manifest checkpoint failed
+
+Symptom:
+
+- transcribe reports that a speaker completed but the session manifest could
+  not be saved.
+
+Likely causes:
+
+- the workspace is unwritable or out of space;
+- the manifest path or parent directory is unsafe or the wrong type; or
+- a transient filesystem error interrupted atomic replacement.
+
+Safe fix:
+
+- correct the reported filesystem or permission problem and retry transcribe
+  normally. Narratio reconciles from the last readable collection and reuses
+  only successfully checkpointed speakers.
+
+A canonical raw transcript left by the failed checkpoint is incidental and
+does not become reusable authority. Preserve the manifest and diagnostics if
+the failure repeats; do not advertise the file by editing manifest evidence.
+
+Relevant references: [Manifest persistence](./internal/manifest.md#persistence-semantics)
+and [Transcribe internals](./internal/stage-transcribe.md).
+
+## Notarius credential reference unavailable
+
+Symptom:
+
+- configuration loading reports `secret.unset`, `secret.empty`, or an invalid
+  name for `pipeline.notarius.llm_api_key`, including during inspection or a
+  run that would not execute extraction.
+
+Likely causes:
+
+- the named environment variable is absent or empty;
+- the selected secrets directory lacks a file with that exact name;
+- a process environment variable exists but is empty, so bootstrap correctly
+  leaves it in place; or
+- the Narratio reference name does not match `api_key_env` in the selected
+  Notarius/PromptKit profile.
+
+Safe fix:
+
+- set the exact nonempty environment variable or provide an equivalently named
+  file through the conventional or explicit secrets directory;
+- remove or set the field to `null` when that deployment needs no credential;
+  and
+- use `config show`, `config sources`, and `config diff` to inspect the public
+  reference name and ownership without exposing its value.
+
+Narratio resolves selected credential references during ordinary pipeline
+loading, before stage selection or resume. Credential changes alone do not
+invalidate a completed extraction, but the current reference must still resolve
+for the command to proceed.
+
+Relevant references: [Secrets](./config.md#secrets-handling) and
+[Notarius credential](./config.md#notarius-credential).
+
 ## Notarius executable missing
 
 Symptom:
@@ -257,6 +417,59 @@ Do not point Notarius directly at the original source path or edit the manifest
 checksum. Relevant references: [Notarius reference configuration](./config.md#notarius-reference-bindings)
 and [Operations: Extraction Workflow](./operations.md#extraction-workflow).
 
+## Notarius reference snapshot cache conflict
+
+Symptom:
+
+- extraction reports that a path beneath
+  `cache/notarius/references/<sha256>/` is a symlink, directory, or another
+  non-regular or unsafe entry.
+
+Safe fix:
+
+- stop commands for the affected session and inspect the exact reported path;
+- investigate unexpected links or special files as possible tampering; and
+- remove only the reported conflicting leaf or digest directory after
+  confirming it belongs to the reconstructible reference cache.
+
+A missing snapshot or corrupt regular snapshot is reconstructed automatically
+from currently verified prepared input. Do not recursively delete the session
+root, edit the manifest checksum, or remove cache entries while extraction is
+active.
+
+Relevant reference: [Operations: Extraction Workflow](./operations.md#extraction-workflow).
+
+## Notarius resume executes work instead of reusing it
+
+Symptom:
+
+- Narratio executes `extract`, and Notarius makes provider calls despite
+  `pipeline.notarius.resume: true`.
+
+Likely causes:
+
+- checkpoint recording is disabled in the Notarius configuration;
+- Notarius is using a nonpersistent or different checkpoint root;
+- transcript/reference bytes, reference paths, pipeline/profile content,
+  local-backend endpoint, selected lanes, or another checkpoint identity input
+  changed; or
+- Narratio extraction was forced, which deliberately omits `--resume`.
+
+Safe fix:
+
+- validate the deployed Notarius configuration and configure explicit
+  persistent chunk-plan and checkpoint roots;
+- inspect Notarius checkpoint decision reason codes for a legitimate cold miss;
+  and
+- retry without force only when reuse is intended.
+
+Remember that Narratio skipping a completed `extract` stage is separate from
+Notarius reusing work inside an extraction that does execute. Narratio does not
+read the Notarius configuration to infer checkpoint enablement.
+
+Relevant references: [Notarius cache controls](./config.md#notarius-cache-controls)
+and [Notarius integration](./integrations/notarius.md).
+
 ## Notarius reference selector or generated-handoff collision
 
 Symptom:
@@ -282,29 +495,6 @@ Narratio validates selector structure and prepared evidence, while Notarius
 owns slot declarations, media compatibility, and generated-handoff conflicts.
 Relevant reference: [Notarius integration](./integrations/notarius.md).
 
-## Atomic Notarius promotion unsupported
-
-Symptom:
-
-- extraction fails with `atomic no-replace directory promotion is unsupported`
-  before a durable bundle or temporary promotion tree is created.
-
-Likely cause:
-
-- Narratio is running on an operating system other than Linux, macOS, or
-  Windows, where the required atomic no-replace directory primitive has not
-  been implemented and verified.
-
-Safe fix:
-
-- run extraction on Linux, macOS, or Windows. Do not replace the atomic commit
-  with a manual copy or move; the session manifest must never observe a partial
-  or overwritten bundle.
-
-This is an extraction-specific platform boundary, not a support statement for
-unrelated Narratio workflows. See
-[Operations: Extraction Workflow](./operations.md#extraction-workflow).
-
 ## Notarius receipt or index incompatible
 
 Symptom:
@@ -352,18 +542,23 @@ Symptom:
 
 Likely causes:
 
-- the executable/config path, pipeline ID, timeout, working directory, or
+- the pipeline ID or configured output contracts changed;
-  configured output contracts changed;
 - a configured prepared reference selector, source, path, checksum, or byte
   size changed;
+- the direct trimmed transcript identity or bytes changed; or
 - the durable bundle, index, lane set, provenance, regular-file status, or
   checksum no longer validates.
 
 Safe fix:
 
 - allow the automatic rerun after verifying the current configuration. Treat
-  an unsafe path or symlink error as filesystem corruption or tampering and
+  an unsafe path or bundle symlink error as filesystem corruption or tampering
-  investigate it rather than replacing files manually.
+  and investigate it rather than replacing files manually.
+
+Executable, config path, timeout, working directory, and Notarius credential
+changes are operational and do not by themselves invalidate reusable extraction
+evidence. A newly selected credential must nevertheless resolve during command
+configuration loading.
 
 ## Notarius transitive configuration changed
 
@@ -395,8 +590,7 @@ Symptom:
 
 Likely causes:
 
-- the per-artifact record is stale, missing, failed, unselected, malformed, or
+- the per-artifact record is stale, missing, failed, unselected, or malformed;
-  from the legacy aggregate-only manifest contract;
 - a configured prompt/profile, dependency, input identity, output path, or
   effective variable changed; or
 - the recorded output is missing, unsafe, empty, or has a size/checksum that no
@@ -425,36 +619,89 @@ narratio analyze 2026-04-04 --artifacts session_recap
 Relevant references: [Operations: Artifact Selection](./operations.md#artifact-selection)
 and [Artifact Internals](./internal/artifacts.md#resolution-rules).
 
-## Legacy aggregate analysis requires regeneration
+## Analysis manifest checkpoint failed
 
 Symptom:
 
-- a manifest from an older Narratio version reports aggregate analyze success
+- analyze reports that it could not save the session manifest after completing
-  and the old files are present, but configured artifact sources remain
+  an artifact.
-  unavailable.
 
-Likely cause:
+Likely causes:
 
-- the manifest has no supported per-artifact analyze state. Aggregate output
+- the workspace is unwritable or out of space;
-  lists do not establish current configured-artifact authority.
+- the manifest path or its parent directory has unsafe permissions or type; or
+- a transient local filesystem error interrupted atomic manifest replacement.
 
 Safe fix:
 
-- regenerate the required artifacts. A partial selection makes only its
+- correct the reported filesystem or permission problem;
-  targets and prerequisites eligible for current state; unselected legacy
+- inspect `session artifacts` to identify the last durable current records; and
-  files intentionally remain unavailable. Run full analysis later when every
+- rerun analyze normally. Narratio reuses earlier successful checkpoints and
-  enabled configured artifact must become current.
+  retries the failed or non-current artifact and its required dependents.
 
-```bash
+Do not treat a newly created canonical artifact file as current when its
-narratio analyze 2026-04-04 --artifacts session_recap
+checkpoint failed, and do not edit the manifest to advertise it. If the error
-narratio analyze 2026-04-04
+recurs, preserve the manifest and run-local diagnostics before repairing the
-```
+workspace.
-
-After current records exist, inspect them and publish explicitly. Do not delete
-the legacy files merely to influence selection; availability is manifest-owned.
 
 Relevant references: [Operations: Stage Execution and Continuation Behavior](./operations.md#stage-execution-and-continuation-behavior)
-and [Manifest Internals](./internal/manifest.md#analyze-owned-artifact-state).
+and [Manifest Internals](./internal/manifest.md#persistence-semantics).
+
+## Built-in transcript or bounds source is unavailable
+
+Symptom:
+
+- analyze, extract, render, or publish says a built-in source is unavailable
+  even though a file exists at its usual session path.
+
+Likely cause:
+
+- the producing stage has not succeeded, its output lacks the exact registered
+  source, run ID, or checksum, its evidence is duplicated or stale, or the
+  durable bytes no longer match the manifest checksum.
+
+Safe fix:
+
+- do not copy or rename the file to make it discoverable. Rerun the producing
+  stage to regenerate both the bytes and current manifest evidence, then retry
+  the consumer.
+
+## An older manifest schema is rejected
+
+Symptom:
+
+- Narratio rejects a local session manifest because `schema_version` is missing
+  or unsupported; or
+- restore rejects a committed remote session manifest for the same reason.
+
+Likely causes:
+
+- the manifest predates the supported canonical state contract. Narratio
+  rejects it at the schema boundary rather than interpreting aggregate output
+  lists or incidental files as current authority; or
+- an older strict-decoding Narratio binary is reading a current manifest that
+  contains additive transcribe speaker state it does not recognize.
+
+Safe fix:
+
+- for unsupported local state, move the old manifest aside and regenerate, or
+  restore a verified current committed snapshot when one is authoritative;
+- for unsupported committed remote state, regenerate healthy local state as
+  needed and republish the session; and
+- for a downgrade rejection, return to the current binary or restore a verified
+  pre-upgrade backup. Do not delete only the speaker fields by hand.
+
+```bash
+narratio run 2026-04-04
+narratio publish 2026-04-04
+```
+
+After current records exist, inspect them and publish explicitly. Do not copy
+old files into canonical paths merely to influence selection; availability is
+manifest-owned.
+
+Relevant references: [Operations: Stage Execution and Continuation Behavior](./operations.md#stage-execution-and-continuation-behavior)
+and [Manifest Internals](./internal/manifest.md#transcribe-owned-speaker-state).
 
 ## Scriptorium private input changed without a rerun
 
@@ -491,7 +738,9 @@ Symptom:
 Likely causes:
 
 - missing `session.previous_session_id`;
-- previous artifact not restored/published for source session.
+- the selected previous-session commit does not declare the required canonical
+  source as a `published_output`; or
+- the declared published object is unavailable or fails identity validation.
 
 Diagnostics:
 
@@ -549,6 +798,8 @@ Symptom:
 Likely cause:
 
 - local durable files differ from remote restore sources.
+- a restore target or one of its ancestors is a symlink or another non-regular
+  filesystem entry.
 
 Diagnostics:
 
@@ -559,7 +810,11 @@ narratio session restore 2026-04-04 --dry-run
 Safe fix:
 
 - review conflicts;
-- rerun with `--force` only when remote state should overwrite local.
+- rerun with `--force` only when remote state should overwrite an eligible
+  regular local file; repair a symlink or other non-regular conflict first.
+
+If the filesystem changed after the plan was printed, rerun restore after
+repairing the destination; installation performs the same safety check again.
 
 Relevant reference: [Operations: Restore Workflow](./operations.md#restore-workflow).
 
@@ -652,7 +907,8 @@ Symptom:
 
 Likely causes:
 
-- invalid `pipeline.secrets.env_dir`;
+- missing, unreadable, or invalid explicit `--secrets-dir`;
+- invalid entries in the conventional `secrets/` directory beside the root pipeline;
 - missing credential environment variables;
 - invalid S3 endpoint/bucket settings.
 
@@ -660,12 +916,12 @@ Diagnostics:
 
 ```bash
 ls -la /path/to/secrets_dir
-env | sed 's/=.*//' | grep -E 'OBJECT_STORAGE|AWS|AUDITA|SCRIPTORIUM'
+env | sed 's/=.*//' | grep -E 'OBJECT_STORAGE|AWS|AUDITA|NOTARIUS|SCRIPTORIUM'
 ```
 
 Safe fix:
 
-- correct secret-file path and permissions;
+- correct `--secrets-dir`, secret-file type/name, and operator-managed permissions;
 - provide required env vars;
 - keep secret values out of YAML.
 

examples/README.md

@@ -11,18 +11,26 @@ in the [configuration reference](../docs/config.md).
   WhisperX URL.
 - [Production-shaped pipeline](pipeline.production.yml): S3 storage, publish,
   external tools, and configured Scriptorium artifacts.
-- [Production/testing split bundle](production-testing/pipeline.yml): explicit
+- [Production-base/testing-overlay bundle](production-testing/pipeline.yml):
-  `conf.d` imports, a production default, and selectable production/testing
+  explicit `conf.d` imports, one ordinary artifact per `artifacts/<key>.yml`,
-  overlays. It also demonstrates canonical-party artifact families and a
+  one family per `artifact-families/<key>.yml`, and the fixed
-  testing-only disabled artifact. Validate it with:
+  `profiles/testing.yml` overlay. Production values live in the base. The
+  bundle demonstrates an input-only artifact dependency, two character families
+  sharing the campaign party, and a testing-only disabled artifact. Inspect it
+  offline with:
 
   ```sh
   narratio config validate --config examples/production-testing/pipeline.yml --campaign-file examples/campaigns/sample-campaign/campaign.yml
-  narratio config diff production testing --config examples/production-testing/pipeline.yml --campaign-file examples/campaigns/sample-campaign/campaign.yml
+  narratio config show --config examples/production-testing/pipeline.yml --campaign-file examples/campaigns/sample-campaign/campaign.yml --profile testing
+  narratio config sources --config examples/production-testing/pipeline.yml --campaign-file examples/campaigns/sample-campaign/campaign.yml --profile testing
+  narratio config artifacts --config examples/production-testing/pipeline.yml --campaign-file examples/campaigns/sample-campaign/campaign.yml --profile testing
+  narratio config diff default testing --config examples/production-testing/pipeline.yml --campaign-file examples/campaigns/sample-campaign/campaign.yml
   ```
 
-  `config show` and `config sources` accept the same selectors and remain
+  These inspection commands load configuration and resolve explicit credential
-  read-only.
+  references, but do not load a session or create runtime state. This bundle is
+  intentionally local-storage and credential-reference free so the walkthrough
+  needs no tools or cloud services.
 - [Full annotated pipeline](pipeline.full.annotated.yml): every implemented
   pipeline section with explanatory comments.
 - [Extraction subset pipeline](pipeline.extraction-subset.yml): a focused
@@ -52,11 +60,20 @@ with the sample campaign and a compatible local- or S3-audio session.
   [glossary](campaigns/sample-campaign/glossary.yml), and canonical
   [party](campaigns/sample-campaign/party.yml) fixture, plus an optional
   [spell-catalog overlay](campaigns/sample-campaign/spell_catalog.json) that
-  follows the Notarius v0.6 contract. Narratio derives the players projection
+  follows the Notarius v0.7 contract. Narratio derives the players projection
-  from this party source; the campaign deliberately has no `players_file`.
+  from this party source.
 - [Sample speaker audio](audio/sample-speaker.flac) is a text placeholder that
   reserves the expected filename and directory shape. Replace it with a real
   FLAC file before running transcription.
 
-The examples contain environment-variable names but no credential values. They
+The examples contain environment-variable names but no credential values. Set
-use fictional campaign content and reserved example domains.
+`OBJECT_STORAGE_KEY_ID`, `OBJECT_STORAGE_KEY`, `AUDITA_LLM_API_KEY`, and
+`NOTARIUS_LLM_API_KEY` to nonempty synthetic or real values before loading the
+examples that reference them. They use fictional campaign content and reserved
+example domains.
+
+For a deployed bundle, keep credential files out of source control in a
+`secrets/` directory beside the root `pipeline.yml`, one file per referenced
+environment-variable name, or select another directory with `--secrets-dir`.
+Static S3 credentials are optional: omit both S3 credential references to use
+the AWS SDK discovery chain.

examples/pipeline.full.annotated.yml

@@ -7,9 +7,8 @@ workspace:
   # Optional: remove run-scoped workdir after successful publish commit.
   cleanup_after_publish: false
 
-# Optional: local secret file loader (directory of ENV_VAR_NAME files).
+# Credential files, when used, belong in a sibling secrets/ directory or in the
-# secrets:
+# directory selected with --secrets-dir; they are not pipeline configuration.
-#   env_dir: ./secrets
 
 storage:
   # Defaults to "local". Use "s3" explicitly for publish + S3 audio workflows.
@@ -24,8 +23,8 @@ storage:
     endpoint: ""
     force_path_style: false
     # Optional; defaults shown explicitly.
-    access_key_id_env: OBJECT_STORAGE_KEY_ID
+    access_key_id: OBJECT_STORAGE_KEY_ID
-    secret_access_key_env: OBJECT_STORAGE_KEY
+    secret_access_key: OBJECT_STORAGE_KEY
 
 campaigns:
   # Optional; defaults to /usr/local/share/narratio/campaigns.
@@ -94,7 +93,7 @@ audita:
   # Optional overrides; defaults shown explicitly where applicable.
   binary: audita
   timeout: 3h
-  llm_api_key_env: AUDITA_LLM_API_KEY
+  llm_api_key: AUDITA_LLM_API_KEY
   modules: [glossary, homophones, spoken_word, grammar]
   base_url: ""
   model: ""
@@ -129,13 +128,26 @@ trim:
     report: false
 
 notarius:
-  # Optional structured extraction between trim and render.
+  # Optional structured extraction after render and before analyze.
   enabled: true
   binary: notarius
+  # Optional environment-variable reference. Match this name to api_key_env in
+  # the selected Notarius/PromptKit profile; never put the credential here.
+  llm_api_key: NOTARIUS_LLM_API_KEY
+  # Optionally select a Notarius ConfigKit overlay when extraction runs. This
+  # is independent of Narratio's selected profile and may use a different name.
+  # See docs/config.md and docs/integrations/notarius.md for the contracts.
+  # config_profile: downstream-economy
   config_path: /usr/local/etc/notarius/config.yml
   pipeline_id: dnd-session
   timeout: 3h
   working_directory: /usr/local/etc/notarius
+  # Reuse compatible accepted Notarius checkpoints during executed extraction.
+  # The Notarius config must enable recording and provide persistent cache roots.
+  resume: true
+  # auto uses or creates a source-addressed chunk plan. Forced Narratio
+  # extraction overrides this with refresh and disables checkpoint loading.
+  chunk_cache: auto
   # External campaign references use prepared Narratio source IDs. Omit an
   # optional binding when the selected Notarius pipeline does not need it.
   references:

examples/pipeline.production.yml

@@ -8,8 +8,8 @@ storage:
     bucket: my-dnd-archive
     root_prefix: dnd
     region: us-east-1
-    access_key_id_env: OBJECT_STORAGE_KEY_ID
+    access_key_id: OBJECT_STORAGE_KEY_ID
-    secret_access_key_env: OBJECT_STORAGE_KEY
+    secret_access_key: OBJECT_STORAGE_KEY
 
 campaigns:
   root: /usr/local/share/narratio/campaigns
@@ -57,7 +57,7 @@ seriatim:
 audita:
   binary: audita
   timeout: 3h
-  llm_api_key_env: AUDITA_LLM_API_KEY
+  llm_api_key: AUDITA_LLM_API_KEY
   modules: [glossary, homophones, spoken_word, grammar]
   output_schema: audita-v1
   work_dir_retention: auto

examples/production-testing/artifact-families/character_items.yml

@@ -1,23 +1,11 @@
 scriptorium:
-  binary: scriptorium
-  config_path: ./scriptorium/config.yml
   artifact_families:
-    character_meta:
-      enabled: true
-      for_each: party.characters
-      prompt_id: dnd.character_meta
-      output_path_pattern: artifacts/characters/{character_id}/meta.md
-      member_vars:
-        character_id: character_id
-        character_name: character.name
-        player_name: player.name
-        class_summary: character.class_summary
     character_items:
       enabled: true
       for_each: party.characters
       prompt_id: dnd.character_items
+      profile_id: production-placeholder
       output_path_pattern: artifacts/characters/{character_id}/items.md
-      member_dependencies: [character_meta]
       inputs:
         character_meta:
           source: narratio.member_artifact.character_meta

examples/production-testing/artifact-families/character_meta.yml

@@ -0,0 +1,13 @@
+scriptorium:
+  artifact_families:
+    character_meta:
+      enabled: true
+      for_each: party.characters
+      prompt_id: dnd.character_meta
+      profile_id: production-placeholder
+      output_path_pattern: artifacts/characters/{character_id}/meta.md
+      member_vars:
+        character_id: character_id
+        character_name: character.name
+        player_name: player.name
+        class_summary: character.class_summary

examples/production-testing/artifacts/player_handout.yml

@@ -0,0 +1,11 @@
+scriptorium:
+  artifacts:
+    player_handout:
+      enabled: true
+      prompt_id: dnd.player_handout
+      profile_id: production-placeholder
+      output_path: artifacts/player-handout.md
+      inputs:
+        recap:
+          source: narratio.artifact.session_recap
+          required: true

examples/production-testing/artifacts/session_recap.yml

@@ -0,0 +1,11 @@
+scriptorium:
+  artifacts:
+    session_recap:
+      enabled: true
+      prompt_id: dnd.session_recap
+      profile_id: production-placeholder
+      output_path: artifacts/session-recap.md
+      inputs:
+        transcript:
+          source: narratio.transcript.final_trimmed
+          required: true

examples/production-testing/conf.d/transcript.yml

@@ -10,6 +10,8 @@ audita:
   binary: audita
   modules: [glossary, grammar]
   output_schema: audita-v1
+  model: narratio-production-model-placeholder
+  validation_model: narratio-production-validator-placeholder
 
 normalize:
   output_path: transcripts/final.json

examples/production-testing/pipeline.yml

@@ -1,15 +1,14 @@
-# Copyable production/testing pipeline entry point. Every fragment is named
+# Copyable production-base/testing-overlay entry point. ConfigKit expands only
-# explicitly; Narratio never scans conf.d automatically.
+# these explicit path and glob imports; Narratio never scans directories itself.
-composition:
+configkit:
   imports:
-    - conf.d/platform.yml
+    - path: conf.d/platform.yml
-    - conf.d/storage.yml
+    - path: conf.d/storage.yml
-    - conf.d/transcript.yml
+    - path: conf.d/transcript.yml
-    - conf.d/artifacts.yml
+    - path: conf.d/publish.yml
-    - conf.d/publish.yml
+    - glob: artifacts/*.yml
-  default_profile: production
+    - glob: artifact-families/*.yml
-  profiles:
+
-    production:
+scriptorium:
-      overlay: profiles/production.yml
+  binary: scriptorium
-    testing:
+  config_path: ./scriptorium/config.yml
-      overlay: profiles/testing.yml

examples/production-testing/profiles/production.yml

@@ -1,10 +0,0 @@
-audita:
-  model: narratio-production-model-placeholder
-  validation_model: narratio-production-validator-placeholder
-
-scriptorium:
-  artifact_families:
-    character_meta:
-      profile_id: production-placeholder
-    character_items:
-      profile_id: production-placeholder

examples/production-testing/profiles/testing.yml

@@ -4,6 +4,10 @@ audita:
 
 scriptorium:
   artifacts:
+    session_recap:
+      profile_id: testing-placeholder
+    player_handout:
+      profile_id: testing-placeholder
     testing_notes:
       enabled: false
       output_path: artifacts/testing-notes.md

go.mod

@@ -1,12 +1,15 @@
 module gitea.maximumdirect.net/eric/narratio
 
-go 1.25.0
+go 1.26.0
 
 require (
+	gitea.maximumdirect.net/eric/ConfigKit v0.3.1
 	github.com/aws/aws-sdk-go-v2/config v1.32.17
 	github.com/aws/aws-sdk-go-v2/credentials v1.19.16
 	github.com/aws/aws-sdk-go-v2/service/s3 v1.101.0
 	github.com/aws/smithy-go v1.25.1
+	github.com/google/go-cmp v0.7.0
+	golang.org/x/sync v0.22.0
 	golang.org/x/sys v0.47.0
 	gopkg.in/yaml.v3 v3.0.1
 )
@@ -26,4 +29,5 @@ require (
 	github.com/aws/aws-sdk-go-v2/service/sso v1.30.17 // indirect
 	github.com/aws/aws-sdk-go-v2/service/ssooidc v1.35.21 // indirect
 	github.com/aws/aws-sdk-go-v2/service/sts v1.42.1 // indirect
+	go.yaml.in/yaml/v4 v4.0.0-rc.6 // indirect
 )

go.sum

@@ -1,3 +1,5 @@
+gitea.maximumdirect.net/eric/ConfigKit v0.3.1 h1:etfTI/dxO3NnS18La0V6IeUh/Iv+3PjVfEyaN8hTuOM=
+gitea.maximumdirect.net/eric/ConfigKit v0.3.1/go.mod h1:tYWpsrbLyu05r8+1MapWp5fN6ZrX9I8hU2xEPedYmYk=
 github.com/aws/aws-sdk-go-v2 v1.41.7 h1:DWpAJt66FmnnaRIOT/8ASTucrvuDPZASqhhLey6tLY8=
 github.com/aws/aws-sdk-go-v2 v1.41.7/go.mod h1:4LAfZOPHNVNQEckOACQx60Y8pSRjIkNZQz1w92xpMJc=
 github.com/aws/aws-sdk-go-v2/aws/protocol/eventstream v1.7.10 h1:gx1AwW1Iyk9Z9dD9F4akX5gnN3QZwUB20GGKH/I+Rho=
@@ -34,6 +36,12 @@ github.com/aws/aws-sdk-go-v2/service/sts v1.42.1 h1:F/M5Y9I3nwr2IEpshZgh1GeHpOIt
 github.com/aws/aws-sdk-go-v2/service/sts v1.42.1/go.mod h1:mTNxImtovCOEEuD65mKW7DCsL+2gjEH+RPEAexAzAio=
 github.com/aws/smithy-go v1.25.1 h1:J8ERsGSU7d+aCmdQur5Txg6bVoYelvQJgtZehD12GkI=
 github.com/aws/smithy-go v1.25.1/go.mod h1:YE2RhdIuDbA5E5bTdciG9KrW3+TiEONeUWCqxX9i1Fc=
+github.com/google/go-cmp v0.7.0 h1:wk8382ETsv4JYUZwIsn6YpYiWiBsYLSJiTsyBybVuN8=
+github.com/google/go-cmp v0.7.0/go.mod h1:pXiqmnSA92OHEEa9HXL2W4E7lf9JzCmGVUdgjX3N/iU=
+go.yaml.in/yaml/v4 v4.0.0-rc.6 h1:1h7H1ohdUh93/FyE4YaDa1Zh64K6VVbjF4K6WUxMtH4=
+go.yaml.in/yaml/v4 v4.0.0-rc.6/go.mod h1:aZqd9kCMsGL7AuUv/m/PvWLdg5sjJsZ4oHDEnfPPfY0=
+golang.org/x/sync v0.22.0 h1:SZjpbeLmrCk4xhRSZFNZW5gFUeCeFgjekvI/+gfScek=
+golang.org/x/sync v0.22.0/go.mod h1:9xrNwdLfx4jkKbNva9FpL6vEN7evnE43NNNJQ2LF3+0=
 golang.org/x/sys v0.47.0 h1:o7XGOvZQCADBQQ4Y7VNq2dRWQR7JmOUW8Kxx4ZsNgWs=
 golang.org/x/sys v0.47.0/go.mod h1:4GL1E5IUh+htKOUEOaiffhrAeqysfVGipDYzABqnCmw=
 gopkg.in/check.v1 v0.0.0-20161208181325-20d25e280405 h1:yhCVgyC4o1eVCa2tZl7eS0r+SDo693bJlVdllGtEeKM=

internal/adapters/audita/fake.go

@@ -6,7 +6,6 @@ import (
 	"fmt"
 	"path/filepath"
 
-	"gitea.maximumdirect.net/eric/narratio/internal/adapters/subprocess"
 	"gitea.maximumdirect.net/eric/narratio/internal/fileops"
 )
 
@@ -84,17 +83,17 @@ func materializePlaceholders(req PolishRequest) error {
 			"merged_transcript_path": req.MergedTranscriptPath,
 			"output_path":            req.OutputProcessedPath,
 		}
-		if err := subprocess.WriteYAMLAtomic(req.GeneratedConfigPath, payload, fileops.WorkspaceFileMode); err != nil {
+		if err := writeGeneratedYAML(req.GeneratedConfigPath, payload); err != nil {
 			return fmt.Errorf("write generated config %q: %w", req.GeneratedConfigPath, err)
 		}
 	}
 	if req.StdoutLogPath != "" {
-		if err := subprocess.WriteFileAtomic(req.StdoutLogPath, []byte("audita noop/fake stdout placeholder\n"), fileops.WorkspaceFileMode); err != nil {
+		if err := fileops.WriteFileAtomic(req.StdoutLogPath, []byte("audita noop/fake stdout placeholder\n"), fileops.WorkspaceFileMode); err != nil {
 			return fmt.Errorf("write stdout log %q: %w", req.StdoutLogPath, err)
 		}
 	}
 	if req.StderrLogPath != "" {
-		if err := subprocess.WriteFileAtomic(req.StderrLogPath, []byte("audita noop/fake stderr placeholder\n"), fileops.WorkspaceFileMode); err != nil {
+		if err := fileops.WriteFileAtomic(req.StderrLogPath, []byte("audita noop/fake stderr placeholder\n"), fileops.WorkspaceFileMode); err != nil {
 			return fmt.Errorf("write stderr log %q: %w", req.StderrLogPath, err)
 		}
 	}
@@ -124,7 +123,7 @@ func writeJSONIfRequested(path string, payload any) error {
 	if err != nil {
 		return fmt.Errorf("marshal placeholder json for %q: %w", path, err)
 	}
-	if err := subprocess.WriteFileAtomic(path, data, fileops.WorkspaceFileMode); err != nil {
+	if err := fileops.WriteFileAtomic(path, data, fileops.WorkspaceFileMode); err != nil {
 		return fmt.Errorf("write placeholder json %q: %w", path, err)
 	}
 	return nil

internal/adapters/audita/generated_config.go

@@ -0,0 +1,19 @@
+package audita
+
+import (
+	"fmt"
+
+	"gitea.maximumdirect.net/eric/narratio/internal/fileops"
+	"gopkg.in/yaml.v3"
+)
+
+func writeGeneratedYAML(path string, payload any) error {
+	data, err := yaml.Marshal(payload)
+	if err != nil {
+		return fmt.Errorf("marshal generated YAML for %q: %w", path, err)
+	}
+	if err := fileops.WriteFileAtomic(path, data, fileops.WorkspaceFileMode); err != nil {
+		return fmt.Errorf("write generated YAML %q: %w", path, err)
+	}
+	return nil
+}

internal/adapters/audita/subprocess.go

@@ -5,7 +5,6 @@ import (
 	"encoding/json"
 	"fmt"
 	"net/url"
-	"os"
 	"strconv"
 	"strings"
 	"time"
@@ -24,7 +23,8 @@ const MaxReportOutputBytes int64 = 16 * 1024 * 1024
 type SubprocessRunnerConfig struct {
 	Binary                   string
 	Timeout                  time.Duration
-	LLMAPIKeyEnv             string
+	LLMAPIKey                string
+	LLMAPIKeyReference       string
 	Modules                  []string
 	BaseURL                  string
 	Model                    string
@@ -43,7 +43,8 @@ type SubprocessRunnerConfig struct {
 type SubprocessRunner struct {
 	binary                   string
 	timeout                  time.Duration
-	llmAPIKeyEnv             string
+	llmAPIKey                string
+	llmAPIKeyReference       string
 	modules                  []string
 	baseURL                  string
 	model                    string
@@ -62,7 +63,8 @@ type SubprocessRunner struct {
 func NewSubprocessRunnerFromConfigValues(
 	binary string,
 	timeout string,
-	llmAPIKeyEnv string,
+	llmAPIKey string,
+	llmAPIKeyReference string,
 	modules []string,
 	baseURL string,
 	model string,
@@ -86,7 +88,8 @@ func NewSubprocessRunnerFromConfigValues(
 	return NewSubprocessRunner(SubprocessRunnerConfig{
 		Binary:                   binary,
 		Timeout:                  parsedTimeout,
-		LLMAPIKeyEnv:             llmAPIKeyEnv,
+		LLMAPIKey:                llmAPIKey,
+		LLMAPIKeyReference:       llmAPIKeyReference,
 		Modules:                  modules,
 		BaseURL:                  baseURL,
 		Model:                    model,
@@ -110,6 +113,9 @@ func NewSubprocessRunner(cfg SubprocessRunnerConfig) (*SubprocessRunner, error)
 	if cfg.Timeout <= 0 {
 		return nil, fmt.Errorf("audita timeout must be > 0")
 	}
+	if (cfg.LLMAPIKey == "") != (strings.TrimSpace(cfg.LLMAPIKeyReference) == "") {
+		return nil, fmt.Errorf("audita credential value and reference must be provided together")
+	}
 	for i, module := range cfg.Modules {
 		if strings.TrimSpace(module) == "" {
 			return nil, fmt.Errorf("audita module at index %d is empty", i)
@@ -152,7 +158,8 @@ func NewSubprocessRunner(cfg SubprocessRunnerConfig) (*SubprocessRunner, error)
 	return &SubprocessRunner{
 		binary:                   strings.TrimSpace(cfg.Binary),
 		timeout:                  cfg.Timeout,
-		llmAPIKeyEnv:             strings.TrimSpace(cfg.LLMAPIKeyEnv),
+		llmAPIKey:                cfg.LLMAPIKey,
+		llmAPIKeyReference:       strings.TrimSpace(cfg.LLMAPIKeyReference),
 		modules:                  modules,
 		baseURL:                  strings.TrimSpace(cfg.BaseURL),
 		model:                    strings.TrimSpace(cfg.Model),
@@ -195,16 +202,10 @@ func (r *SubprocessRunner) Run(ctx context.Context, req PolishRequest) (PolishRe
 	}
 	args := r.buildArgs(req, reqModules)
 
-	credentialEnvVar := strings.TrimSpace(r.llmAPIKeyEnv)
+	credentialPresent := r.llmAPIKeyReference != ""
-	credentialPresent := false
 	env := map[string]string{}
-	if credentialEnvVar != "" {
+	if credentialPresent {
-		credential, ok := os.LookupEnv(credentialEnvVar)
+		env["AUDITA_LLM_API_KEY"] = r.llmAPIKey
-		if !ok || strings.TrimSpace(credential) == "" {
-			return PolishResult{}, fmt.Errorf("audita: required credential environment variable %s is not set", credentialEnvVar)
-		}
-		env["AUDITA_LLM_API_KEY"] = credential
-		credentialPresent = true
 	}
 
 	if req.GeneratedConfigPath != "" {
@@ -269,7 +270,7 @@ func (r *SubprocessRunner) Run(ctx context.Context, req PolishRequest) (PolishRe
 			"total_llm_concurrency":      r.totalLLMConcurrency,
 			"proposal_llm_concurrency":   r.proposalLLMConcurrency,
 			"validation_llm_concurrency": r.validationLLMConcurrency,
-			"credential_env_var":         r.llmAPIKeyEnv,
+			"credential_env_var":         r.llmAPIKeyReference,
 			"credential_present":         credentialPresent,
 		},
 	}, nil
@@ -299,7 +300,7 @@ func (r *SubprocessRunner) failureResult(req PolishRequest, modules []string, ru
 			"total_llm_concurrency":      r.totalLLMConcurrency,
 			"proposal_llm_concurrency":   r.proposalLLMConcurrency,
 			"validation_llm_concurrency": r.validationLLMConcurrency,
-			"credential_env_var":         r.llmAPIKeyEnv,
+			"credential_env_var":         r.llmAPIKeyReference,
 			"credential_present":         credentialPresent,
 		},
 	}
@@ -375,10 +376,10 @@ func (r *SubprocessRunner) writeInvocationConfig(req PolishRequest, args []strin
 		"output_path":                req.OutputProcessedPath,
 		"report_path":                req.ReportPath,
 		"work_dir":                   req.WorkDir,
-		"credential_env_var":         r.llmAPIKeyEnv,
+		"credential_env_var":         r.llmAPIKeyReference,
 		"credential_present":         credentialPresent,
 	}
-	return subprocess.WriteYAMLAtomic(req.GeneratedConfigPath, payload, fileops.WorkspaceFileMode)
+	return writeGeneratedYAML(req.GeneratedConfigPath, payload)
 }
 
 func validateProcessedOutput(path string) error {
@@ -425,7 +426,7 @@ func validateJSONFile(path string) error {
 }
 
 func readAuditaResult(path string, limit int64, category string) ([]byte, error) {
-	data, err := fileops.ReadRegularFile(path, limit)
+	data, err := fileops.ReadFileBounded(path, limit)
 	if err != nil {
 		return nil, fmt.Errorf("audita %s result exceeds or cannot be read within %d-byte limit: %w", category, limit, err)
 	}

internal/adapters/audita/subprocess_test.go

@@ -6,16 +6,12 @@ import (
 	"fmt"
 	"os"
 	"path/filepath"
-	"runtime"
 	"strings"
 	"testing"
 	"time"
 )
 
 func TestSubprocessRunnerSuccessArgsEnvAndValidation(t *testing.T) {
-	if runtime.GOOS == "windows" {
-		t.Skip("helper wrapper script uses /bin/sh")
-	}
 
 	t.Setenv("GO_WANT_AUDITA_HELPER", "1")
 	t.Setenv("AUDITA_HELPER_MODE", "success")
@@ -31,7 +27,8 @@ func TestSubprocessRunnerSuccessArgsEnvAndValidation(t *testing.T) {
 	runner, err := NewSubprocessRunner(SubprocessRunnerConfig{
 		Binary:                   wrapper,
 		Timeout:                  mustParseAuditaDuration(t, "2s"),
-		LLMAPIKeyEnv:             "OPENAI_KEY_SOURCE",
+		LLMAPIKey:                "super-secret",
+		LLMAPIKeyReference:       "OPENAI_KEY_SOURCE",
 		Modules:                  []string{"glossary", "homophones", "glossary"},
 		BaseURL:                  "https://openrouter.ai/api/v1",
 		Model:                    "openrouter/google/gemma-4-31b-it",
@@ -48,6 +45,7 @@ func TestSubprocessRunnerSuccessArgsEnvAndValidation(t *testing.T) {
 	if err != nil {
 		t.Fatalf("NewSubprocessRunner() error = %v", err)
 	}
+	t.Setenv("OPENAI_KEY_SOURCE", "changed-after-resolution")
 
 	dir := t.TempDir()
 	req := PolishRequest{
@@ -132,47 +130,31 @@ func TestSubprocessRunnerSuccessArgsEnvAndValidation(t *testing.T) {
 	}
 }
 
-func TestSubprocessRunnerMissingConfiguredCredentialFails(t *testing.T) {
+func TestSubprocessRunnerRejectsIncompleteResolvedCredential(t *testing.T) {
-	if runtime.GOOS == "windows" {
+	_, err := NewSubprocessRunner(SubprocessRunnerConfig{
-		t.Skip("helper wrapper script uses /bin/sh")
+		Binary:             writeAuditaHelperWrapper(t),
-	}
+		Timeout:            mustParseAuditaDuration(t, "2s"),
-	runner := mustAuditaRunner(t, SubprocessRunnerConfig{
+		LLMAPIKeyReference: "MISSING_AUDITA_KEY",
-		Binary:       writeAuditaHelperWrapper(t),
-		Timeout:      mustParseAuditaDuration(t, "2s"),
-		LLMAPIKeyEnv: "MISSING_AUDITA_KEY",
-		Modules:      []string{"glossary"},
-		BaseURL:      "https://openrouter.ai/api/v1",
-		Model:        "openrouter/google/gemma-4-31b-it",
-		Report:       false,
 	})
-	req := auditaReqForTest(t, false)
+	if err == nil || !strings.Contains(err.Error(), "credential value and reference must be provided together") {
-
+		t.Fatalf("NewSubprocessRunner() error = %v, want credential pair error", err)
-	_, err := runner.Run(context.Background(), req)
-	if err == nil {
-		t.Fatal("expected error, got nil")
-	}
-	if !strings.Contains(err.Error(), "required credential environment variable MISSING_AUDITA_KEY is not set") {
-		t.Fatalf("error = %q, want missing credential guidance", err.Error())
 	}
 }
 
 func TestSubprocessRunnerUnconfiguredCredentialEnvOmitsCredential(t *testing.T) {
-	if runtime.GOOS == "windows" {
-		t.Skip("helper wrapper script uses /bin/sh")
-	}
 	t.Setenv("GO_WANT_AUDITA_HELPER", "1")
 	t.Setenv("AUDITA_HELPER_MODE", "success")
 	recordPath := filepath.Join(t.TempDir(), "record.json")
 	t.Setenv("AUDITA_HELPER_RECORD_PATH", recordPath)
 
 	runner := mustAuditaRunner(t, SubprocessRunnerConfig{
-		Binary:       writeAuditaHelperWrapper(t),
+		Binary:    writeAuditaHelperWrapper(t),
-		Timeout:      mustParseAuditaDuration(t, "2s"),
+		Timeout:   mustParseAuditaDuration(t, "2s"),
-		LLMAPIKeyEnv: "",
+		LLMAPIKey: "",
-		Modules:      []string{"glossary"},
+		Modules:   []string{"glossary"},
-		BaseURL:      "https://openrouter.ai/api/v1",
+		BaseURL:   "https://openrouter.ai/api/v1",
-		Model:        "openrouter/google/gemma-4-31b-it",
+		Model:     "openrouter/google/gemma-4-31b-it",
-		Report:       false,
+		Report:    false,
 	})
 	req := auditaReqForTest(t, false)
 
@@ -190,9 +172,6 @@ func TestSubprocessRunnerUnconfiguredCredentialEnvOmitsCredential(t *testing.T)
 }
 
 func TestSubprocessRunnerOmitsUnspecifiedParentEnvironment(t *testing.T) {
-	if runtime.GOOS == "windows" {
-		t.Skip("helper wrapper script uses /bin/sh")
-	}
 	t.Setenv("GO_WANT_AUDITA_HELPER", "1")
 	t.Setenv("AUDITA_HELPER_MODE", "success")
 	t.Setenv("AUDITA_INHERITED_MARKER", "inherited-from-parent")
@@ -200,13 +179,13 @@ func TestSubprocessRunnerOmitsUnspecifiedParentEnvironment(t *testing.T) {
 	t.Setenv("AUDITA_HELPER_RECORD_PATH", recordPath)
 
 	runner := mustAuditaRunner(t, SubprocessRunnerConfig{
-		Binary:       writeAuditaHelperWrapper(t),
+		Binary:    writeAuditaHelperWrapper(t),
-		Timeout:      mustParseAuditaDuration(t, "2s"),
+		Timeout:   mustParseAuditaDuration(t, "2s"),
-		Modules:      []string{"glossary"},
+		Modules:   []string{"glossary"},
-		BaseURL:      "https://openrouter.ai/api/v1",
+		BaseURL:   "https://openrouter.ai/api/v1",
-		Model:        "openrouter/google/gemma-4-31b-it",
+		Model:     "openrouter/google/gemma-4-31b-it",
-		Report:       false,
+		Report:    false,
-		LLMAPIKeyEnv: "",
+		LLMAPIKey: "",
 	})
 	req := auditaReqForTest(t, false)
 	if _, err := runner.Run(context.Background(), req); err != nil {
@@ -220,21 +199,18 @@ func TestSubprocessRunnerOmitsUnspecifiedParentEnvironment(t *testing.T) {
 }
 
 func TestSubprocessRunnerOmitsModulesFlagWhenNotConfigured(t *testing.T) {
-	if runtime.GOOS == "windows" {
-		t.Skip("helper wrapper script uses /bin/sh")
-	}
 	t.Setenv("GO_WANT_AUDITA_HELPER", "1")
 	t.Setenv("AUDITA_HELPER_MODE", "success")
 	recordPath := filepath.Join(t.TempDir(), "record.json")
 	t.Setenv("AUDITA_HELPER_RECORD_PATH", recordPath)
 
 	runner := mustAuditaRunner(t, SubprocessRunnerConfig{
-		Binary:       writeAuditaHelperWrapper(t),
+		Binary:    writeAuditaHelperWrapper(t),
-		Timeout:      mustParseAuditaDuration(t, "2s"),
+		Timeout:   mustParseAuditaDuration(t, "2s"),
-		LLMAPIKeyEnv: "",
+		LLMAPIKey: "",
-		BaseURL:      "https://openrouter.ai/api/v1",
+		BaseURL:   "https://openrouter.ai/api/v1",
-		Model:        "openrouter/google/gemma-4-31b-it",
+		Model:     "openrouter/google/gemma-4-31b-it",
-		Report:       false,
+		Report:    false,
 	})
 	req := auditaReqForTest(t, false)
 	if _, err := runner.Run(context.Background(), req); err != nil {
@@ -249,19 +225,16 @@ func TestSubprocessRunnerOmitsModulesFlagWhenNotConfigured(t *testing.T) {
 }
 
 func TestSubprocessRunnerOmitsBaseURLAndModelFlagsWhenNotConfigured(t *testing.T) {
-	if runtime.GOOS == "windows" {
-		t.Skip("helper wrapper script uses /bin/sh")
-	}
 	t.Setenv("GO_WANT_AUDITA_HELPER", "1")
 	t.Setenv("AUDITA_HELPER_MODE", "success")
 	recordPath := filepath.Join(t.TempDir(), "record.json")
 	t.Setenv("AUDITA_HELPER_RECORD_PATH", recordPath)
 
 	runner := mustAuditaRunner(t, SubprocessRunnerConfig{
-		Binary:       writeAuditaHelperWrapper(t),
+		Binary:    writeAuditaHelperWrapper(t),
-		Timeout:      mustParseAuditaDuration(t, "2s"),
+		Timeout:   mustParseAuditaDuration(t, "2s"),
-		LLMAPIKeyEnv: "",
+		LLMAPIKey: "",
-		Report:       false,
+		Report:    false,
 	})
 	req := auditaReqForTest(t, false)
 	if _, err := runner.Run(context.Background(), req); err != nil {
@@ -279,22 +252,20 @@ func TestSubprocessRunnerOmitsBaseURLAndModelFlagsWhenNotConfigured(t *testing.T
 }
 
 func TestSubprocessRunnerSubprocessFailure(t *testing.T) {
-	if runtime.GOOS == "windows" {
-		t.Skip("helper wrapper script uses /bin/sh")
-	}
 	t.Setenv("GO_WANT_AUDITA_HELPER", "1")
 	t.Setenv("AUDITA_HELPER_MODE", "fail")
 	t.Setenv("OPENAI_KEY_SOURCE", "super-secret")
 	t.Setenv("AUDITA_HELPER_RECORD_PATH", filepath.Join(t.TempDir(), "record.json"))
 
 	runner := mustAuditaRunner(t, SubprocessRunnerConfig{
-		Binary:       writeAuditaHelperWrapper(t),
+		Binary:             writeAuditaHelperWrapper(t),
-		Timeout:      mustParseAuditaDuration(t, "2s"),
+		Timeout:            mustParseAuditaDuration(t, "2s"),
-		LLMAPIKeyEnv: "OPENAI_KEY_SOURCE",
+		LLMAPIKey:          "super-secret",
-		Modules:      []string{"glossary"},
+		LLMAPIKeyReference: "OPENAI_KEY_SOURCE",
-		BaseURL:      "https://openrouter.ai/api/v1",
+		Modules:            []string{"glossary"},
-		Model:        "openrouter/google/gemma-4-31b-it",
+		BaseURL:            "https://openrouter.ai/api/v1",
-		Report:       true,
+		Model:              "openrouter/google/gemma-4-31b-it",
+		Report:             true,
 	})
 	req := auditaReqForTest(t, true)
 	_, err := runner.Run(context.Background(), req)
@@ -310,25 +281,26 @@ func TestSubprocessRunnerSubprocessFailure(t *testing.T) {
 	if !strings.Contains(err.Error(), req.StdoutLogPath) || !strings.Contains(err.Error(), req.StderrLogPath) {
 		t.Fatalf("error = %q, want stdout/stderr log paths", err.Error())
 	}
+	if strings.Contains(err.Error(), "super-secret") {
+		t.Fatalf("failure diagnostic exposed credential value: %q", err.Error())
+	}
 }
 
 func TestSubprocessRunnerSubprocessFailureAddsStderrDescriptorHint(t *testing.T) {
-	if runtime.GOOS == "windows" {
-		t.Skip("helper wrapper script uses /bin/sh")
-	}
 	t.Setenv("GO_WANT_AUDITA_HELPER", "1")
 	t.Setenv("AUDITA_HELPER_MODE", "fail_bad_fd")
 	t.Setenv("OPENAI_KEY_SOURCE", "super-secret")
 	t.Setenv("AUDITA_HELPER_RECORD_PATH", filepath.Join(t.TempDir(), "record.json"))
 
 	runner := mustAuditaRunner(t, SubprocessRunnerConfig{
-		Binary:       writeAuditaHelperWrapper(t),
+		Binary:             writeAuditaHelperWrapper(t),
-		Timeout:      mustParseAuditaDuration(t, "2s"),
+		Timeout:            mustParseAuditaDuration(t, "2s"),
-		LLMAPIKeyEnv: "OPENAI_KEY_SOURCE",
+		LLMAPIKey:          "super-secret",
-		Modules:      []string{"glossary"},
+		LLMAPIKeyReference: "OPENAI_KEY_SOURCE",
-		BaseURL:      "https://openrouter.ai/api/v1",
+		Modules:            []string{"glossary"},
-		Model:        "openrouter/google/gemma-4-31b-it",
+		BaseURL:            "https://openrouter.ai/api/v1",
-		Report:       true,
+		Model:              "openrouter/google/gemma-4-31b-it",
+		Report:             true,
 	})
 	req := auditaReqForTest(t, true)
 	_, err := runner.Run(context.Background(), req)
@@ -341,22 +313,20 @@ func TestSubprocessRunnerSubprocessFailureAddsStderrDescriptorHint(t *testing.T)
 }
 
 func TestSubprocessRunnerMissingOutputFails(t *testing.T) {
-	if runtime.GOOS == "windows" {
-		t.Skip("helper wrapper script uses /bin/sh")
-	}
 	t.Setenv("GO_WANT_AUDITA_HELPER", "1")
 	t.Setenv("AUDITA_HELPER_MODE", "missing_output")
 	t.Setenv("OPENAI_KEY_SOURCE", "super-secret")
 	t.Setenv("AUDITA_HELPER_RECORD_PATH", filepath.Join(t.TempDir(), "record.json"))
 
 	runner := mustAuditaRunner(t, SubprocessRunnerConfig{
-		Binary:       writeAuditaHelperWrapper(t),
+		Binary:             writeAuditaHelperWrapper(t),
-		Timeout:      mustParseAuditaDuration(t, "2s"),
+		Timeout:            mustParseAuditaDuration(t, "2s"),
-		LLMAPIKeyEnv: "OPENAI_KEY_SOURCE",
+		LLMAPIKey:          "super-secret",
-		Modules:      []string{"glossary"},
+		LLMAPIKeyReference: "OPENAI_KEY_SOURCE",
-		BaseURL:      "https://openrouter.ai/api/v1",
+		Modules:            []string{"glossary"},
-		Model:        "openrouter/google/gemma-4-31b-it",
+		BaseURL:            "https://openrouter.ai/api/v1",
-		Report:       false,
+		Model:              "openrouter/google/gemma-4-31b-it",
+		Report:             false,
 	})
 	req := auditaReqForTest(t, false)
 	_, err := runner.Run(context.Background(), req)
@@ -369,22 +339,20 @@ func TestSubprocessRunnerMissingOutputFails(t *testing.T) {
 }
 
 func TestSubprocessRunnerInvalidOutputJSONFails(t *testing.T) {
-	if runtime.GOOS == "windows" {
-		t.Skip("helper wrapper script uses /bin/sh")
-	}
 	t.Setenv("GO_WANT_AUDITA_HELPER", "1")
 	t.Setenv("AUDITA_HELPER_MODE", "invalid_output")
 	t.Setenv("OPENAI_KEY_SOURCE", "super-secret")
 	t.Setenv("AUDITA_HELPER_RECORD_PATH", filepath.Join(t.TempDir(), "record.json"))
 
 	runner := mustAuditaRunner(t, SubprocessRunnerConfig{
-		Binary:       writeAuditaHelperWrapper(t),
+		Binary:             writeAuditaHelperWrapper(t),
-		Timeout:      mustParseAuditaDuration(t, "2s"),
+		Timeout:            mustParseAuditaDuration(t, "2s"),
-		LLMAPIKeyEnv: "OPENAI_KEY_SOURCE",
+		LLMAPIKey:          "super-secret",
-		Modules:      []string{"glossary"},
+		LLMAPIKeyReference: "OPENAI_KEY_SOURCE",
-		BaseURL:      "https://openrouter.ai/api/v1",
+		Modules:            []string{"glossary"},
-		Model:        "openrouter/google/gemma-4-31b-it",
+		BaseURL:            "https://openrouter.ai/api/v1",
-		Report:       false,
+		Model:              "openrouter/google/gemma-4-31b-it",
+		Report:             false,
 	})
 	req := auditaReqForTest(t, false)
 	_, err := runner.Run(context.Background(), req)
@@ -397,22 +365,20 @@ func TestSubprocessRunnerInvalidOutputJSONFails(t *testing.T) {
 }
 
 func TestSubprocessRunnerSegmentsMissingFails(t *testing.T) {
-	if runtime.GOOS == "windows" {
-		t.Skip("helper wrapper script uses /bin/sh")
-	}
 	t.Setenv("GO_WANT_AUDITA_HELPER", "1")
 	t.Setenv("AUDITA_HELPER_MODE", "segments_missing")
 	t.Setenv("OPENAI_KEY_SOURCE", "super-secret")
 	t.Setenv("AUDITA_HELPER_RECORD_PATH", filepath.Join(t.TempDir(), "record.json"))
 
 	runner := mustAuditaRunner(t, SubprocessRunnerConfig{
-		Binary:       writeAuditaHelperWrapper(t),
+		Binary:             writeAuditaHelperWrapper(t),
-		Timeout:      mustParseAuditaDuration(t, "2s"),
+		Timeout:            mustParseAuditaDuration(t, "2s"),
-		LLMAPIKeyEnv: "OPENAI_KEY_SOURCE",
+		LLMAPIKey:          "super-secret",
-		Modules:      []string{"glossary"},
+		LLMAPIKeyReference: "OPENAI_KEY_SOURCE",
-		BaseURL:      "https://openrouter.ai/api/v1",
+		Modules:            []string{"glossary"},
-		Model:        "openrouter/google/gemma-4-31b-it",
+		BaseURL:            "https://openrouter.ai/api/v1",
-		Report:       false,
+		Model:              "openrouter/google/gemma-4-31b-it",
+		Report:             false,
 	})
 	req := auditaReqForTest(t, false)
 	_, err := runner.Run(context.Background(), req)
@@ -425,22 +391,20 @@ func TestSubprocessRunnerSegmentsMissingFails(t *testing.T) {
 }
 
 func TestSubprocessRunnerInvalidReportJSONFails(t *testing.T) {
-	if runtime.GOOS == "windows" {
-		t.Skip("helper wrapper script uses /bin/sh")
-	}
 	t.Setenv("GO_WANT_AUDITA_HELPER", "1")
 	t.Setenv("AUDITA_HELPER_MODE", "invalid_report")
 	t.Setenv("OPENAI_KEY_SOURCE", "super-secret")
 	t.Setenv("AUDITA_HELPER_RECORD_PATH", filepath.Join(t.TempDir(), "record.json"))
 
 	runner := mustAuditaRunner(t, SubprocessRunnerConfig{
-		Binary:       writeAuditaHelperWrapper(t),
+		Binary:             writeAuditaHelperWrapper(t),
-		Timeout:      mustParseAuditaDuration(t, "2s"),
+		Timeout:            mustParseAuditaDuration(t, "2s"),
-		LLMAPIKeyEnv: "OPENAI_KEY_SOURCE",
+		LLMAPIKey:          "super-secret",
-		Modules:      []string{"glossary"},
+		LLMAPIKeyReference: "OPENAI_KEY_SOURCE",
-		BaseURL:      "https://openrouter.ai/api/v1",
+		Modules:            []string{"glossary"},
-		Model:        "openrouter/google/gemma-4-31b-it",
+		BaseURL:            "https://openrouter.ai/api/v1",
-		Report:       true,
+		Model:              "openrouter/google/gemma-4-31b-it",
+		Report:             true,
 	})
 	req := auditaReqForTest(t, true)
 	_, err := runner.Run(context.Background(), req)
@@ -453,21 +417,22 @@ func TestSubprocessRunnerInvalidReportJSONFails(t *testing.T) {
 }
 
 func TestSubprocessRunnerConstructorValidation(t *testing.T) {
-	_, err := NewSubprocessRunnerFromConfigValues("", "3h", "AUDITA_LLM_API_KEY", []string{"glossary"}, "https://openrouter.ai/api/v1", "openrouter/google/gemma-4-31b-it", "", "", "", "", nil, nil, "", nil, true)
+	_, err := NewSubprocessRunnerFromConfigValues("", "3h", "synthetic-secret", "AUDITA_LLM_API_KEY", []string{"glossary"}, "https://openrouter.ai/api/v1", "openrouter/google/gemma-4-31b-it", "", "", "", "", nil, nil, "", nil, true)
 	if err == nil {
 		t.Fatal("expected binary validation error")
 	}
-	_, err = NewSubprocessRunnerFromConfigValues("audita", "bad", "AUDITA_LLM_API_KEY", []string{"glossary"}, "https://openrouter.ai/api/v1", "openrouter/google/gemma-4-31b-it", "", "", "", "", nil, nil, "", nil, true)
+	_, err = NewSubprocessRunnerFromConfigValues("audita", "bad", "synthetic-secret", "AUDITA_LLM_API_KEY", []string{"glossary"}, "https://openrouter.ai/api/v1", "openrouter/google/gemma-4-31b-it", "", "", "", "", nil, nil, "", nil, true)
 	if err == nil {
 		t.Fatal("expected timeout parse error")
 	}
 	_, err = NewSubprocessRunner(SubprocessRunnerConfig{
-		Binary:       "audita",
+		Binary:             "audita",
-		Timeout:      mustParseAuditaDuration(t, "1s"),
+		Timeout:            mustParseAuditaDuration(t, "1s"),
-		LLMAPIKeyEnv: "AUDITA_LLM_API_KEY",
+		LLMAPIKey:          "synthetic-secret",
-		Modules:      []string{"glossary"},
+		LLMAPIKeyReference: "AUDITA_LLM_API_KEY",
-		BaseURL:      "://",
+		Modules:            []string{"glossary"},
-		Model:        "openrouter/google/gemma-4-31b-it",
+		BaseURL:            "://",
+		Model:              "openrouter/google/gemma-4-31b-it",
 	})
 	if err == nil {
 		t.Fatal("expected base url validation error")

internal/adapters/notarius/fake.go

@@ -15,6 +15,10 @@ func (f *FakeRunner) Run(ctx context.Context, req RunRequest) (RunResult, error)
 		return RunResult{}, err
 	}
 	copyRequest := req
+	if req.ConfigProfile != nil {
+		profile := *req.ConfigProfile
+		copyRequest.ConfigProfile = &profile
+	}
 	copyRequest.References = append([]ReferenceBinding(nil), req.References...)
 	f.Requests = append(f.Requests, copyRequest)
 	if f.Err != nil {

internal/adapters/notarius/runner.go

@@ -4,6 +4,8 @@ package notarius
 import (
 	"context"
 	"time"
+
+	configkit "gitea.maximumdirect.net/eric/ConfigKit"
 )
 
 const ReceiptSchemaVersion = "notarius.run-result.v2"
@@ -22,7 +24,10 @@ type ReferenceBinding struct {
 
 // RunRequest contains the resolved inputs and diagnostic destinations for one invocation.
 type RunRequest struct {
-	Binary           string
+	Binary string
+	// LLMAPIKey is an optional already-resolved credential reference.
+	LLMAPIKey        *configkit.Secret
+	ConfigProfile    *string
 	ConfigPath       string
 	PipelineID       string
 	InputPath        string
@@ -31,6 +36,8 @@ type RunRequest struct {
 	ReceiptPath      string
 	LogPath          string
 	Timeout          time.Duration
+	Resume           bool
+	ChunkCache       string
 	References       []ReferenceBinding
 }
 

internal/adapters/notarius/subprocess.go

@@ -52,6 +52,20 @@ func (r *SubprocessRunner) Run(ctx context.Context, req RunRequest) (RunResult,
 	if err != nil {
 		return RunResult{}, err
 	}
+	var envOverrides map[string]string
+	var sensitiveEnvNames []string
+	if req.LLMAPIKey != nil {
+		if !req.LLMAPIKey.IsResolved() {
+			return RunResult{}, fmt.Errorf("notarius LLM API credential must be resolved and nonempty")
+		}
+		name := req.LLMAPIKey.EnvName()
+		value := req.LLMAPIKey.Value()
+		if value == "" {
+			return RunResult{}, fmt.Errorf("notarius LLM API credential must be resolved and nonempty")
+		}
+		envOverrides = map[string]string{name: value}
+		sensitiveEnvNames = []string{name}
+	}
 
 	args := []string{
 		"run", req.PipelineID,
@@ -59,18 +73,29 @@ func (r *SubprocessRunner) Run(ctx context.Context, req RunRequest) (RunResult,
 		"--input", req.InputPath,
 		"--output-dir", req.OutputRoot,
 	}
+	if req.Resume {
+		args = append(args, "--resume")
+	}
+	if req.ChunkCache != "" {
+		args = append(args, "--chunk_cache", req.ChunkCache)
+	}
 	for _, reference := range references {
 		args = append(args, "--reference", reference.Selector+"="+reference.Path)
 	}
+	if req.ConfigProfile != nil {
+		args = append(args, "--config-profile", *req.ConfigProfile)
+	}
 	args = append(args, "--json")
 	processResult, err := r.run(ctx, subprocess.RunRequest{
-		Executable:      req.Binary,
+		Executable:        req.Binary,
-		Args:            args,
+		Args:              args,
-		WorkingDir:      req.WorkingDirectory,
+		WorkingDir:        req.WorkingDirectory,
-		Timeout:         req.Timeout,
+		Timeout:           req.Timeout,
-		DiagnosticOwner: "notarius",
+		DiagnosticOwner:   "notarius",
-		StdoutLogPath:   req.ReceiptPath,
+		StdoutLogPath:     req.ReceiptPath,
-		StderrLogPath:   req.LogPath,
+		StderrLogPath:     req.LogPath,
+		EnvOverrides:      envOverrides,
+		SensitiveEnvNames: sensitiveEnvNames,
 	})
 	baseResult := RunResult{
 		ReceiptPath: req.ReceiptPath,
@@ -143,6 +168,9 @@ func validateRunRequest(req RunRequest) ([]ReferenceBinding, error) {
 	if req.Timeout <= 0 {
 		return nil, fmt.Errorf("notarius timeout must be positive")
 	}
+	if !validChunkCacheMode(req.ChunkCache) {
+		return nil, fmt.Errorf("notarius chunk cache mode must be empty, auto, bypass, or refresh")
+	}
 	for label, path := range map[string]string{
 		"config":            req.ConfigPath,
 		"input":             req.InputPath,
@@ -201,6 +229,15 @@ func validateRunRequest(req RunRequest) ([]ReferenceBinding, error) {
 	return references, nil
 }
 
+func validChunkCacheMode(mode string) bool {
+	switch mode {
+	case "", "auto", "bypass", "refresh":
+		return true
+	default:
+		return false
+	}
+}
+
 type receiptDocument struct {
 	SchemaVersion             string                      `json:"schema_version"`
 	RunID                     string                      `json:"run_id"`
@@ -248,7 +285,7 @@ func validateValidationSummaries(documents []validationSummaryDocument) ([]Valid
 		if document.Status != "complete" && document.Status != "rejected" && document.Status != "incomplete" {
 			return nil, fmt.Errorf("notarius validation summary status %q is invalid", document.Status)
 		}
-		if document.ProducerAttemptCount == nil || *document.ProducerAttemptCount <= 0 || !validTerminalAction(document.TerminalAction) {
+		if document.ProducerAttemptCount == nil || *document.ProducerAttemptCount < 0 || !validTerminalAction(document.TerminalAction) {
 			return nil, fmt.Errorf("notarius validation summary is missing required fields")
 		}
 		summaries = append(summaries, ValidationSummary{
@@ -671,7 +708,7 @@ func sumWarningOccurrences(values []WarningSummary) (int, error) {
 }
 
 func decodeBoundedJSON(path string, limit int64, destination any) error {
-	data, err := fileops.ReadRegularFile(path, limit)
+	data, err := fileops.ReadFileBounded(path, limit)
 	if err != nil {
 		return fmt.Errorf("notarius JSON result exceeds or cannot be read within %d-byte limit: %w", limit, err)
 	}

internal/adapters/notarius/subprocess_test.go

@@ -11,6 +11,7 @@ import (
 	"testing"
 	"time"
 
+	configkit "gitea.maximumdirect.net/eric/ConfigKit"
 	sharedsubprocess "gitea.maximumdirect.net/eric/narratio/internal/adapters/subprocess"
 )
 
@@ -76,6 +77,174 @@ func TestSubprocessRunnerBuildsExactInvocationAndDiscoversBundle(t *testing.T) {
 	}
 }
 
+func TestSubprocessRunnerMapsCacheControlsToArguments(t *testing.T) {
+	tests := []struct {
+		name       string
+		resume     bool
+		chunkCache string
+		want       []string
+	}{
+		{name: "resume", resume: true, want: []string{"--resume"}},
+		{name: "auto", chunkCache: "auto", want: []string{"--chunk_cache", "auto"}},
+		{name: "bypass", chunkCache: "bypass", want: []string{"--chunk_cache", "bypass"}},
+		{name: "refresh", chunkCache: "refresh", want: []string{"--chunk_cache", "refresh"}},
+		{name: "resume and mode", resume: true, chunkCache: "auto", want: []string{"--resume", "--chunk_cache", "auto"}},
+	}
+	for _, test := range tests {
+		t.Run(test.name, func(t *testing.T) {
+			req := validRunRequest(t)
+			req.Resume = test.resume
+			req.ChunkCache = test.chunkCache
+			var captured sharedsubprocess.RunRequest
+			runner := &SubprocessRunner{run: func(_ context.Context, processReq sharedsubprocess.RunRequest) (sharedsubprocess.RunResult, error) {
+				captured = processReq
+				writeValidBundleAndReceipt(t, req, false)
+				return sharedsubprocess.RunResult{ExitCode: 0}, nil
+			}}
+
+			if _, err := runner.Run(context.Background(), req); err != nil {
+				t.Fatalf("Run() error = %v", err)
+			}
+			want := []string{
+				"run", req.PipelineID, "--config", req.ConfigPath, "--input", req.InputPath,
+				"--output-dir", req.OutputRoot,
+			}
+			want = append(want, test.want...)
+			want = append(want, "--json")
+			if !reflect.DeepEqual(captured.Args, want) {
+				t.Fatalf("subprocess args = %#v, want %#v", captured.Args, want)
+			}
+		})
+	}
+}
+
+func TestSubprocessRunnerMapsConfigProfileToArgumentPair(t *testing.T) {
+	tests := []struct {
+		name     string
+		supplied bool
+		value    string
+	}{
+		{name: "omitted"},
+		{name: "empty", supplied: true},
+		{name: "default", supplied: true, value: "default"},
+		{name: "named", supplied: true, value: "inexpensive"},
+	}
+	for _, test := range tests {
+		t.Run(test.name, func(t *testing.T) {
+			req := validRunRequest(t)
+			if test.supplied {
+				req.ConfigProfile = &test.value
+			}
+			var captured sharedsubprocess.RunRequest
+			runner := &SubprocessRunner{run: func(_ context.Context, processReq sharedsubprocess.RunRequest) (sharedsubprocess.RunResult, error) {
+				captured = processReq
+				writeValidBundleAndReceipt(t, req, false)
+				return sharedsubprocess.RunResult{ExitCode: 0}, nil
+			}}
+
+			if _, err := runner.Run(context.Background(), req); err != nil {
+				t.Fatalf("Run() error = %v", err)
+			}
+			flagIndex := -1
+			for index, arg := range captured.Args {
+				if arg == "--config-profile" {
+					flagIndex = index
+					break
+				}
+			}
+			if !test.supplied {
+				if flagIndex >= 0 {
+					t.Fatalf("subprocess args unexpectedly contain config profile: %#v", captured.Args)
+				}
+				return
+			}
+			if flagIndex < 0 || flagIndex+2 >= len(captured.Args) || captured.Args[flagIndex+1] != test.value || captured.Args[flagIndex+2] != "--json" {
+				t.Fatalf("config profile argument boundary = %#v, want flag, exact value, then --json", captured.Args)
+			}
+		})
+	}
+}
+
+func TestSubprocessRunnerRejectsInvalidChunkCacheBeforeLaunch(t *testing.T) {
+	for _, mode := range []string{"AUTO", "invalid", " auto "} {
+		t.Run(mode, func(t *testing.T) {
+			req := validRunRequest(t)
+			req.ChunkCache = mode
+			started := false
+			runner := &SubprocessRunner{run: func(context.Context, sharedsubprocess.RunRequest) (sharedsubprocess.RunResult, error) {
+				started = true
+				return sharedsubprocess.RunResult{}, nil
+			}}
+			_, err := runner.Run(context.Background(), req)
+			if err == nil || !strings.Contains(err.Error(), "chunk cache mode") {
+				t.Fatalf("Run() error = %v, want chunk cache validation failure", err)
+			}
+			if started {
+				t.Fatal("subprocess started with invalid chunk cache mode")
+			}
+		})
+	}
+}
+
+func TestSubprocessRunnerForwardsLoadedCredentialWithoutRereadingEnvironment(t *testing.T) {
+	const envName = "NARRATIO_NOTARIUS_CUSTOM_LLM_KEY"
+	const loadedValue = "loaded-notarius-credential"
+	credential := loadTestCredential(t, envName, loadedValue)
+	t.Setenv(envName, "changed-parent-credential")
+
+	req := validRunRequest(t)
+	req.LLMAPIKey = credential
+	var captured sharedsubprocess.RunRequest
+	runner := &SubprocessRunner{run: func(_ context.Context, processReq sharedsubprocess.RunRequest) (sharedsubprocess.RunResult, error) {
+		captured = processReq
+		writeValidBundleAndReceipt(t, req, false)
+		return sharedsubprocess.RunResult{ExitCode: 0}, nil
+	}}
+	if _, err := runner.Run(context.Background(), req); err != nil {
+		t.Fatalf("Run() error = %v", err)
+	}
+	if len(captured.EnvOverrides) != 1 || captured.EnvOverrides[envName] != loadedValue {
+		t.Fatal("subprocess did not receive exactly the loaded credential override")
+	}
+	if !reflect.DeepEqual(captured.SensitiveEnvNames, []string{envName}) {
+		t.Fatalf("sensitive environment names = %#v, want configured reference name", captured.SensitiveEnvNames)
+	}
+	wantArgs := []string{
+		"run", req.PipelineID, "--config", req.ConfigPath, "--input", req.InputPath,
+		"--output-dir", req.OutputRoot, "--json",
+	}
+	if !reflect.DeepEqual(captured.Args, wantArgs) {
+		t.Fatalf("subprocess args = %#v, want %#v", captured.Args, wantArgs)
+	}
+}
+
+func TestSubprocessRunnerRejectsInvalidProgrammaticCredentialBeforeLaunch(t *testing.T) {
+	for _, test := range []struct {
+		name      string
+		reference configkit.Secret
+	}{
+		{name: "unresolved", reference: configkit.SecretEnv("NARRATIO_NOTARIUS_UNRESOLVED_KEY")},
+		{name: "empty", reference: configkit.SecretEnv("")},
+	} {
+		t.Run(test.name, func(t *testing.T) {
+			req := validRunRequest(t)
+			req.LLMAPIKey = &test.reference
+			started := false
+			runner := &SubprocessRunner{run: func(context.Context, sharedsubprocess.RunRequest) (sharedsubprocess.RunResult, error) {
+				started = true
+				return sharedsubprocess.RunResult{}, nil
+			}}
+			_, err := runner.Run(context.Background(), req)
+			if err == nil || !strings.Contains(err.Error(), "resolved and nonempty") {
+				t.Fatalf("Run() error = %v, want credential validation failure", err)
+			}
+			if started {
+				t.Fatal("subprocess started with an invalid credential")
+			}
+		})
+	}
+}
+
 func TestSubprocessRunnerBuildsOrderedReferenceArguments(t *testing.T) {
 	req := validRunRequest(t)
 	referenceRoot := t.TempDir()
@@ -172,16 +341,23 @@ func TestSubprocessRunnerRejectsInvalidReferencesBeforeLaunch(t *testing.T) {
 }
 
 func TestSubprocessRunnerUsesMinimalEnvironmentAndSeparatesStreams(t *testing.T) {
+	const envName = "NARRATIO_NOTARIUS_CUSTOM_CAPTURE_KEY"
+	const canary = "notarius-captured-stream-canary"
+	const unrelated = "NARRATIO_NOTARIUS_UNRELATED_PARENT_KEY"
 	req := validRunRequest(t)
+	req.LLMAPIKey = loadTestCredential(t, envName, canary)
 	writeValidBundleAndReceipt(t, req, false)
 	receiptFixture := req.ReceiptPath + ".fixture"
 	data, err := os.ReadFile(req.ReceiptPath)
 	if err != nil {
 		t.Fatalf("ReadFile(receipt) error = %v", err)
 	}
-	if err := os.WriteFile(receiptFixture, data, 0o644); err != nil {
+	var receipt map[string]any
-		t.Fatalf("WriteFile(receipt fixture) error = %v", err)
+	if err := json.Unmarshal(data, &receipt); err != nil {
+		t.Fatalf("decode receipt fixture: %v", err)
 	}
+	receipt["credential_echo"] = canary
+	writeJSONFile(t, receiptFixture, receipt)
 	if err := os.Remove(req.ReceiptPath); err != nil {
 		t.Fatalf("Remove(receipt) error = %v", err)
 	}
@@ -193,27 +369,55 @@ func TestSubprocessRunnerUsesMinimalEnvironmentAndSeparatesStreams(t *testing.T)
 	script := writeShellScript(t, `#!/bin/sh
 pwd > "$NOTARIUS_CAPTURE_DIR/working-directory"
 printf '%s' "$NOTARIUS_INHERITED_VALUE" > "$NOTARIUS_CAPTURE_DIR/environment"
-printf 'diagnostic stream\n' >&2
+if [ "$`+envName+`" != '`+canary+`' ]; then
+  printf 'selected credential mismatch\n' >&2
+  exit 8
+fi
+if [ -n "$`+unrelated+`" ]; then
+  printf 'unrelated credential inherited\n' >&2
+  exit 9
+fi
+printf '%s\n' "$`+envName+`" >&2
 cat "$NOTARIUS_RECEIPT_FIXTURE"
 `)
 	req.Binary = script
 	t.Setenv("NOTARIUS_CAPTURE_DIR", captureDir)
 	t.Setenv("NOTARIUS_INHERITED_VALUE", "inherited-value")
 	t.Setenv("NOTARIUS_RECEIPT_FIXTURE", receiptFixture)
+	t.Setenv(unrelated, "unrelated-parent-credential")
 
 	if _, err := NewSubprocessRunner().Run(context.Background(), req); err != nil {
 		t.Fatalf("Run() error = %v", err)
 	}
 	assertTextFile(t, filepath.Join(captureDir, "working-directory"), req.WorkingDirectory+"\n")
 	assertTextFile(t, filepath.Join(captureDir, "environment"), "")
-	assertTextFile(t, req.LogPath, "diagnostic stream\n")
+	assertRedactedCredentialFile(t, req.LogPath, canary)
 	receiptBytes, err := os.ReadFile(req.ReceiptPath)
 	if err != nil {
 		t.Fatalf("ReadFile(receipt) error = %v", err)
 	}
-	if strings.Contains(string(receiptBytes), "diagnostic stream") {
+	assertRedactedCredentialBytes(t, receiptBytes, canary)
-		t.Fatal("receipt contains stderr output")
+}
+
+func TestSubprocessRunnerRedactsCredentialOnProcessFailure(t *testing.T) {
+	const envName = "NARRATIO_NOTARIUS_CUSTOM_FAILURE_KEY"
+	const canary = "notarius-failure-stream-canary"
+	req := validRunRequest(t)
+	req.LLMAPIKey = loadTestCredential(t, envName, canary)
+	req.Binary = writeShellScript(t, `#!/bin/sh
+printf '%s\n' "$`+envName+`"
+printf '%s\n' "$`+envName+`" >&2
+exit 7
+`)
+	_, err := NewSubprocessRunner().Run(context.Background(), req)
+	if err == nil {
+		t.Fatal("Run() error = nil, want process failure")
 	}
+	if strings.Contains(err.Error(), canary) || !strings.Contains(err.Error(), "<redacted>") {
+		t.Fatal("process failure did not return redacted diagnostics")
+	}
+	assertRedactedCredentialFile(t, req.ReceiptPath, canary)
+	assertRedactedCredentialFile(t, req.LogPath, canary)
 }
 
 func TestSubprocessRunnerReturnsProcessFailuresWithoutParsingStdout(t *testing.T) {
@@ -231,6 +435,10 @@ func TestSubprocessRunnerReturnsProcessFailuresWithoutParsingStdout(t *testing.T
 	for _, test := range tests {
 		t.Run(test.name, func(t *testing.T) {
 			req := validRunRequest(t)
+			if test.name == "nonzero" {
+				profile := "missing-profile"
+				req.ConfigProfile = &profile
+			}
 			req.Binary = writeShellScript(t, "#!/bin/sh\n"+test.scriptBody)
 			req.Timeout = test.timeout
 			ctx := context.Background()
@@ -527,7 +735,8 @@ func TestLoadDiagnosticSummariesValidateBoundsAndTolerateUnknownFields(t *testin
 }
 
 func TestFakeRunnerCapturesRequestsAndHonorsContextAndError(t *testing.T) {
-	req := RunRequest{PipelineID: "pipeline", References: []ReferenceBinding{{Selector: "party", Path: "/references/party.json"}}}
+	profile := "testing"
+	req := RunRequest{PipelineID: "pipeline", ConfigProfile: &profile, References: []ReferenceBinding{{Selector: "party", Path: "/references/party.json"}}}
 	want := RunResult{BundleRoot: "/bundle"}
 	fake := &FakeRunner{Result: want}
 	got, err := fake.Run(context.Background(), req)
@@ -538,6 +747,10 @@ func TestFakeRunnerCapturesRequestsAndHonorsContextAndError(t *testing.T) {
 	if fake.Requests[0].References[0].Path != "/references/party.json" {
 		t.Fatalf("fake retained aliased request references: %#v", fake.Requests[0].References)
 	}
+	profile = "changed"
+	if *fake.Requests[0].ConfigProfile != "testing" {
+		t.Fatalf("fake retained aliased config profile: %q", *fake.Requests[0].ConfigProfile)
+	}
 
 	wantErr := errors.New("configured failure")
 	fake.Err = wantErr
@@ -579,6 +792,56 @@ func validRunRequest(t *testing.T) RunRequest {
 	}
 }
 
+func loadTestCredential(t *testing.T, envName, value string) *configkit.Secret {
+	t.Helper()
+	t.Setenv(envName, value)
+	type testConfig struct {
+		Credential *configkit.Secret `yaml:"credential"`
+	}
+	result, err := configkit.Load(configkit.Bytes("credential.yml", []byte("credential: "+envName+"\n")), configkit.Options[testConfig]{})
+	if err != nil {
+		t.Fatalf("load test credential: %v", err)
+	}
+	return result.Config.Credential
+}
+
+func assertRedactedCredentialFile(t *testing.T, path, canary string) {
+	t.Helper()
+	data, err := os.ReadFile(path)
+	if err != nil {
+		t.Fatalf("ReadFile(%q) error = %v", path, err)
+	}
+	assertRedactedCredentialBytes(t, data, canary)
+}
+
+func assertRedactedCredentialBytes(t *testing.T, data []byte, canary string) {
+	t.Helper()
+	text := string(data)
+	if strings.Contains(text, canary) || !strings.Contains(text, "<redacted>") {
+		t.Fatal("captured stream did not contain the required redaction marker")
+	}
+}
+
+func TestValidateValidationSummariesAcceptsReusedProducerWithoutAttempts(t *testing.T) {
+	zero := 0
+	summaries, err := validateValidationSummaries([]validationSummaryDocument{{
+		Stage: "chunk", Status: "complete", ProducerAttemptCount: &zero, TerminalAction: "accepted",
+	}})
+	if err != nil {
+		t.Fatalf("validateValidationSummaries() error = %v", err)
+	}
+	if len(summaries) != 1 || summaries[0].ProducerAttemptCount != 0 {
+		t.Fatalf("summaries = %#v, want one reused producer with zero attempts", summaries)
+	}
+
+	negative := -1
+	if _, err := validateValidationSummaries([]validationSummaryDocument{{
+		Stage: "chunk", Status: "complete", ProducerAttemptCount: &negative, TerminalAction: "accepted",
+	}}); err == nil {
+		t.Fatal("validateValidationSummaries() negative producer attempts error = nil")
+	}
+}
+
 func writeValidBundleAndReceipt(t *testing.T, req RunRequest, includeUnknown bool) {
 	t.Helper()
 	bundle := filepath.Join(req.OutputRoot, "notarius-run-1")

internal/adapters/scriptorium/fake.go

@@ -4,7 +4,6 @@ import (
 	"context"
 	"fmt"
 
-	"gitea.maximumdirect.net/eric/narratio/internal/adapters/subprocess"
 	"gitea.maximumdirect.net/eric/narratio/internal/fileops"
 )
 
@@ -143,7 +142,7 @@ func (f *FakeRunner) RenderArtifact(ctx context.Context, req RenderArtifactReque
 
 func materializeRunPlaceholders(req RunArtifactRequest) error {
 	if req.OutputPath != "" {
-		if err := subprocess.WriteFileAtomic(req.OutputPath, []byte("scriptorium noop/fake run artifact\n"), fileops.WorkspaceFileMode); err != nil {
+		if err := fileops.WriteFileAtomic(req.OutputPath, []byte("scriptorium noop/fake run artifact\n"), fileops.WorkspaceFileMode); err != nil {
 			return fmt.Errorf("write run output %q: %w", req.OutputPath, err)
 		}
 	}
@@ -155,17 +154,17 @@ func materializeRunPlaceholders(req RunArtifactRequest) error {
 			"prompt_id":   req.PromptID,
 			"output_path": req.OutputPath,
 		}
-		if err := subprocess.WriteYAMLAtomic(req.GeneratedConfigPath, payload, fileops.WorkspaceFileMode); err != nil {
+		if err := writeGeneratedYAML(req.GeneratedConfigPath, payload); err != nil {
 			return fmt.Errorf("write generated config %q: %w", req.GeneratedConfigPath, err)
 		}
 	}
 	if req.StdoutLogPath != "" {
-		if err := subprocess.WriteFileAtomic(req.StdoutLogPath, []byte("scriptorium noop/fake run stdout placeholder\n"), fileops.WorkspaceFileMode); err != nil {
+		if err := fileops.WriteFileAtomic(req.StdoutLogPath, []byte("scriptorium noop/fake run stdout placeholder\n"), fileops.WorkspaceFileMode); err != nil {
 			return fmt.Errorf("write stdout log %q: %w", req.StdoutLogPath, err)
 		}
 	}
 	if req.StderrLogPath != "" {
-		if err := subprocess.WriteFileAtomic(req.StderrLogPath, []byte("scriptorium noop/fake run stderr placeholder\n"), fileops.WorkspaceFileMode); err != nil {
+		if err := fileops.WriteFileAtomic(req.StderrLogPath, []byte("scriptorium noop/fake run stderr placeholder\n"), fileops.WorkspaceFileMode); err != nil {
 			return fmt.Errorf("write stderr log %q: %w", req.StderrLogPath, err)
 		}
 	}
@@ -174,7 +173,7 @@ func materializeRunPlaceholders(req RunArtifactRequest) error {
 
 func materializeRenderPlaceholders(req RenderArtifactRequest) error {
 	if req.OutputPath != "" {
-		if err := subprocess.WriteFileAtomic(req.OutputPath, []byte("{\"schema\":\"scriptorium.render.v1\",\"placeholder\":true}\n"), fileops.WorkspaceFileMode); err != nil {
+		if err := fileops.WriteFileAtomic(req.OutputPath, []byte("{\"schema\":\"scriptorium.render.v1\",\"placeholder\":true}\n"), fileops.WorkspaceFileMode); err != nil {
 			return fmt.Errorf("write render output %q: %w", req.OutputPath, err)
 		}
 	}
@@ -186,17 +185,17 @@ func materializeRenderPlaceholders(req RenderArtifactRequest) error {
 			"prompt_id":   req.PromptID,
 			"output_path": req.OutputPath,
 		}
-		if err := subprocess.WriteYAMLAtomic(req.GeneratedConfigPath, payload, fileops.WorkspaceFileMode); err != nil {
+		if err := writeGeneratedYAML(req.GeneratedConfigPath, payload); err != nil {
 			return fmt.Errorf("write generated config %q: %w", req.GeneratedConfigPath, err)
 		}
 	}
 	if req.StdoutLogPath != "" {
-		if err := subprocess.WriteFileAtomic(req.StdoutLogPath, []byte("scriptorium noop/fake render stdout placeholder\n"), fileops.WorkspaceFileMode); err != nil {
+		if err := fileops.WriteFileAtomic(req.StdoutLogPath, []byte("scriptorium noop/fake render stdout placeholder\n"), fileops.WorkspaceFileMode); err != nil {
 			return fmt.Errorf("write stdout log %q: %w", req.StdoutLogPath, err)
 		}
 	}
 	if req.StderrLogPath != "" {
-		if err := subprocess.WriteFileAtomic(req.StderrLogPath, []byte("scriptorium noop/fake render stderr placeholder\n"), fileops.WorkspaceFileMode); err != nil {
+		if err := fileops.WriteFileAtomic(req.StderrLogPath, []byte("scriptorium noop/fake render stderr placeholder\n"), fileops.WorkspaceFileMode); err != nil {
 			return fmt.Errorf("write stderr log %q: %w", req.StderrLogPath, err)
 		}
 	}

internal/adapters/scriptorium/generated_config.go

@@ -0,0 +1,19 @@
+package scriptorium
+
+import (
+	"fmt"
+
+	"gitea.maximumdirect.net/eric/narratio/internal/fileops"
+	"gopkg.in/yaml.v3"
+)
+
+func writeGeneratedYAML(path string, payload any) error {
+	data, err := yaml.Marshal(payload)
+	if err != nil {
+		return fmt.Errorf("marshal generated YAML for %q: %w", path, err)
+	}
+	if err := fileops.WriteFileAtomic(path, data, fileops.WorkspaceFileMode); err != nil {
+		return fmt.Errorf("write generated YAML %q: %w", path, err)
+	}
+	return nil
+}

internal/adapters/scriptorium/subprocess.go

@@ -342,11 +342,11 @@ func writeInvocationConfig(path string, payload invocationPayload) error {
 		"render_format":        payload.RenderFormat,
 		"render_prompt_logged": payload.RenderPromptStore,
 	}
-	return subprocess.WriteYAMLAtomic(path, data, fileops.WorkspaceFileMode)
+	return writeGeneratedYAML(path, data)
 }
 
 func validateNonEmptyOutput(path string) error {
-	data, err := fileops.ReadRegularFile(path, MaxOutputFileBytes)
+	data, err := fileops.ReadFileBounded(path, MaxOutputFileBytes)
 	if err != nil {
 		return fmt.Errorf("scriptorium artifact output exceeds or cannot be read within %d-byte limit: %w", MaxOutputFileBytes, err)
 	}

internal/adapters/scriptorium/subprocess_test.go

@@ -6,16 +6,12 @@ import (
 	"fmt"
 	"os"
 	"path/filepath"
-	"runtime"
 	"strings"
 	"testing"
 	"time"
 )
 
 func TestSubprocessRunnerRunSuccessBuildsDeterministicArgsAndCapturesLogs(t *testing.T) {
-	if runtime.GOOS == "windows" {
-		t.Skip("helper wrapper script uses /bin/sh")
-	}
 
 	t.Setenv("GO_WANT_SCRIPTORIUM_HELPER", "1")
 	t.Setenv("SCRIPTORIUM_HELPER_MODE", "run_success")
@@ -95,9 +91,6 @@ func TestSubprocessRunnerRunSuccessBuildsDeterministicArgsAndCapturesLogs(t *tes
 }
 
 func TestSubprocessRunnerRunExitCodeOneFails(t *testing.T) {
-	if runtime.GOOS == "windows" {
-		t.Skip("helper wrapper script uses /bin/sh")
-	}
 	t.Setenv("GO_WANT_SCRIPTORIUM_HELPER", "1")
 	t.Setenv("SCRIPTORIUM_HELPER_MODE", "fail_exit1")
 	t.Setenv("SCRIPTORIUM_HELPER_RECORD_PATH", filepath.Join(t.TempDir(), "record.json"))
@@ -118,9 +111,6 @@ func TestSubprocessRunnerRunExitCodeOneFails(t *testing.T) {
 }
 
 func TestSubprocessRunnerRunExitCodeTwoReturnsValidationFailureAndPreservesOutput(t *testing.T) {
-	if runtime.GOOS == "windows" {
-		t.Skip("helper wrapper script uses /bin/sh")
-	}
 	t.Setenv("GO_WANT_SCRIPTORIUM_HELPER", "1")
 	t.Setenv("SCRIPTORIUM_HELPER_MODE", "fail_exit2_with_output")
 	t.Setenv("SCRIPTORIUM_HELPER_RECORD_PATH", filepath.Join(t.TempDir(), "record.json"))
@@ -144,9 +134,6 @@ func TestSubprocessRunnerRunExitCodeTwoReturnsValidationFailureAndPreservesOutpu
 }
 
 func TestSubprocessRunnerRunMissingOutputOnSuccessFails(t *testing.T) {
-	if runtime.GOOS == "windows" {
-		t.Skip("helper wrapper script uses /bin/sh")
-	}
 	t.Setenv("GO_WANT_SCRIPTORIUM_HELPER", "1")
 	t.Setenv("SCRIPTORIUM_HELPER_MODE", "run_success_missing_output")
 	t.Setenv("SCRIPTORIUM_HELPER_RECORD_PATH", filepath.Join(t.TempDir(), "record.json"))
@@ -166,9 +153,6 @@ func TestSubprocessRunnerRunMissingOutputOnSuccessFails(t *testing.T) {
 }
 
 func TestSubprocessRunnerRenderSuccess(t *testing.T) {
-	if runtime.GOOS == "windows" {
-		t.Skip("helper wrapper script uses /bin/sh")
-	}
 	t.Setenv("GO_WANT_SCRIPTORIUM_HELPER", "1")
 	t.Setenv("SCRIPTORIUM_HELPER_MODE", "render_success")
 	recordPath := filepath.Join(t.TempDir(), "record.json")

internal/adapters/seriatim/fake.go

@@ -4,7 +4,6 @@ import (
 	"context"
 	"fmt"
 
-	"gitea.maximumdirect.net/eric/narratio/internal/adapters/subprocess"
 	"gitea.maximumdirect.net/eric/narratio/internal/fileops"
 )
 
@@ -261,7 +260,7 @@ func (f *FakeRunner) Render(ctx context.Context, req RenderRequest) (RenderResul
 
 func materializePlaceholders(req MergeRequest) error {
 	if req.OutputMergedTranscriptPath != "" {
-		if err := subprocess.WriteFileAtomic(req.OutputMergedTranscriptPath, []byte(`{"schema":"seriatim.intermediate.v1","segments":[]}`), fileops.WorkspaceFileMode); err != nil {
+		if err := fileops.WriteFileAtomic(req.OutputMergedTranscriptPath, []byte(`{"schema":"seriatim.intermediate.v1","segments":[]}`), fileops.WorkspaceFileMode); err != nil {
 			return fmt.Errorf("write merged transcript %q: %w", req.OutputMergedTranscriptPath, err)
 		}
 	}
@@ -272,22 +271,22 @@ func materializePlaceholders(req MergeRequest) error {
 			"input_transcript_paths": req.InputTranscriptPaths,
 			"output_path":            req.OutputMergedTranscriptPath,
 		}
-		if err := subprocess.WriteYAMLAtomic(req.GeneratedConfigPath, payload, fileops.WorkspaceFileMode); err != nil {
+		if err := writeGeneratedYAML(req.GeneratedConfigPath, payload); err != nil {
 			return fmt.Errorf("write generated config %q: %w", req.GeneratedConfigPath, err)
 		}
 	}
 	if req.StdoutLogPath != "" {
-		if err := subprocess.WriteFileAtomic(req.StdoutLogPath, []byte("seriatim noop/fake stdout placeholder\n"), fileops.WorkspaceFileMode); err != nil {
+		if err := fileops.WriteFileAtomic(req.StdoutLogPath, []byte("seriatim noop/fake stdout placeholder\n"), fileops.WorkspaceFileMode); err != nil {
 			return fmt.Errorf("write stdout log %q: %w", req.StdoutLogPath, err)
 		}
 	}
 	if req.StderrLogPath != "" {
-		if err := subprocess.WriteFileAtomic(req.StderrLogPath, []byte("seriatim noop/fake stderr placeholder\n"), fileops.WorkspaceFileMode); err != nil {
+		if err := fileops.WriteFileAtomic(req.StderrLogPath, []byte("seriatim noop/fake stderr placeholder\n"), fileops.WorkspaceFileMode); err != nil {
 			return fmt.Errorf("write stderr log %q: %w", req.StderrLogPath, err)
 		}
 	}
 	if req.ReportPath != "" {
-		if err := subprocess.WriteFileAtomic(req.ReportPath, []byte(`{"schema":"seriatim.report.v1","placeholder":true}`), fileops.WorkspaceFileMode); err != nil {
+		if err := fileops.WriteFileAtomic(req.ReportPath, []byte(`{"schema":"seriatim.report.v1","placeholder":true}`), fileops.WorkspaceFileMode); err != nil {
 			return fmt.Errorf("write report %q: %w", req.ReportPath, err)
 		}
 	}
@@ -296,7 +295,7 @@ func materializePlaceholders(req MergeRequest) error {
 
 func materializeTrimPlaceholders(req TrimRequest) error {
 	if req.OutputTrimmedPath != "" {
-		if err := subprocess.WriteFileAtomic(req.OutputTrimmedPath, []byte(`{"schema":"seriatim.intermediate.v1","segments":[]}`), fileops.WorkspaceFileMode); err != nil {
+		if err := fileops.WriteFileAtomic(req.OutputTrimmedPath, []byte(`{"schema":"seriatim.intermediate.v1","segments":[]}`), fileops.WorkspaceFileMode); err != nil {
 			return fmt.Errorf("write trimmed transcript %q: %w", req.OutputTrimmedPath, err)
 		}
 	}
@@ -309,17 +308,17 @@ func materializeTrimPlaceholders(req TrimRequest) error {
 			"output_path":   req.OutputTrimmedPath,
 			"keep_selector": req.KeepSelector,
 		}
-		if err := subprocess.WriteYAMLAtomic(req.GeneratedConfigPath, payload, fileops.WorkspaceFileMode); err != nil {
+		if err := writeGeneratedYAML(req.GeneratedConfigPath, payload); err != nil {
 			return fmt.Errorf("write generated config %q: %w", req.GeneratedConfigPath, err)
 		}
 	}
 	if req.StdoutLogPath != "" {
-		if err := subprocess.WriteFileAtomic(req.StdoutLogPath, []byte("seriatim noop/fake trim stdout placeholder\n"), fileops.WorkspaceFileMode); err != nil {
+		if err := fileops.WriteFileAtomic(req.StdoutLogPath, []byte("seriatim noop/fake trim stdout placeholder\n"), fileops.WorkspaceFileMode); err != nil {
 			return fmt.Errorf("write stdout log %q: %w", req.StdoutLogPath, err)
 		}
 	}
 	if req.StderrLogPath != "" {
-		if err := subprocess.WriteFileAtomic(req.StderrLogPath, []byte("seriatim noop/fake trim stderr placeholder\n"), fileops.WorkspaceFileMode); err != nil {
+		if err := fileops.WriteFileAtomic(req.StderrLogPath, []byte("seriatim noop/fake trim stderr placeholder\n"), fileops.WorkspaceFileMode); err != nil {
 			return fmt.Errorf("write stderr log %q: %w", req.StderrLogPath, err)
 		}
 	}
@@ -328,7 +327,7 @@ func materializeTrimPlaceholders(req TrimRequest) error {
 
 func materializeNormalizePlaceholders(req NormalizeRequest) error {
 	if req.OutputNormalizedPath != "" {
-		if err := subprocess.WriteFileAtomic(req.OutputNormalizedPath, []byte(`{"schema":"seriatim.intermediate.v1","segments":[]}`), fileops.WorkspaceFileMode); err != nil {
+		if err := fileops.WriteFileAtomic(req.OutputNormalizedPath, []byte(`{"schema":"seriatim.intermediate.v1","segments":[]}`), fileops.WorkspaceFileMode); err != nil {
 			return fmt.Errorf("write normalized transcript %q: %w", req.OutputNormalizedPath, err)
 		}
 	}
@@ -344,22 +343,22 @@ func materializeNormalizePlaceholders(req NormalizeRequest) error {
 		if req.ReportPath != "" {
 			payload["report_path"] = req.ReportPath
 		}
-		if err := subprocess.WriteYAMLAtomic(req.GeneratedConfigPath, payload, fileops.WorkspaceFileMode); err != nil {
+		if err := writeGeneratedYAML(req.GeneratedConfigPath, payload); err != nil {
 			return fmt.Errorf("write generated config %q: %w", req.GeneratedConfigPath, err)
 		}
 	}
 	if req.StdoutLogPath != "" {
-		if err := subprocess.WriteFileAtomic(req.StdoutLogPath, []byte("seriatim noop/fake normalize stdout placeholder\n"), fileops.WorkspaceFileMode); err != nil {
+		if err := fileops.WriteFileAtomic(req.StdoutLogPath, []byte("seriatim noop/fake normalize stdout placeholder\n"), fileops.WorkspaceFileMode); err != nil {
 			return fmt.Errorf("write stdout log %q: %w", req.StdoutLogPath, err)
 		}
 	}
 	if req.StderrLogPath != "" {
-		if err := subprocess.WriteFileAtomic(req.StderrLogPath, []byte("seriatim noop/fake normalize stderr placeholder\n"), fileops.WorkspaceFileMode); err != nil {
+		if err := fileops.WriteFileAtomic(req.StderrLogPath, []byte("seriatim noop/fake normalize stderr placeholder\n"), fileops.WorkspaceFileMode); err != nil {
 			return fmt.Errorf("write stderr log %q: %w", req.StderrLogPath, err)
 		}
 	}
 	if req.ReportPath != "" {
-		if err := subprocess.WriteFileAtomic(req.ReportPath, []byte(`{"schema":"seriatim.report.v1","placeholder":true}`), fileops.WorkspaceFileMode); err != nil {
+		if err := fileops.WriteFileAtomic(req.ReportPath, []byte(`{"schema":"seriatim.report.v1","placeholder":true}`), fileops.WorkspaceFileMode); err != nil {
 			return fmt.Errorf("write report %q: %w", req.ReportPath, err)
 		}
 	}
@@ -368,7 +367,7 @@ func materializeNormalizePlaceholders(req NormalizeRequest) error {
 
 func materializeRenderPlaceholders(req RenderRequest) error {
 	if req.OutputRenderedPath != "" {
-		if err := subprocess.WriteFileAtomic(req.OutputRenderedPath, []byte("# Transcript\n\nRendered markdown placeholder.\n"), fileops.WorkspaceFileMode); err != nil {
+		if err := fileops.WriteFileAtomic(req.OutputRenderedPath, []byte("# Transcript\n\nRendered markdown placeholder.\n"), fileops.WorkspaceFileMode); err != nil {
 			return fmt.Errorf("write rendered transcript %q: %w", req.OutputRenderedPath, err)
 		}
 	}
@@ -385,17 +384,17 @@ func materializeRenderPlaceholders(req RenderRequest) error {
 			"include_segment_ids": req.IncludeSegmentIDs,
 			"include_metadata":    req.IncludeMetadata,
 		}
-		if err := subprocess.WriteYAMLAtomic(req.GeneratedConfigPath, payload, fileops.WorkspaceFileMode); err != nil {
+		if err := writeGeneratedYAML(req.GeneratedConfigPath, payload); err != nil {
 			return fmt.Errorf("write generated config %q: %w", req.GeneratedConfigPath, err)
 		}
 	}
 	if req.StdoutLogPath != "" {
-		if err := subprocess.WriteFileAtomic(req.StdoutLogPath, []byte("seriatim noop/fake render stdout placeholder\n"), fileops.WorkspaceFileMode); err != nil {
+		if err := fileops.WriteFileAtomic(req.StdoutLogPath, []byte("seriatim noop/fake render stdout placeholder\n"), fileops.WorkspaceFileMode); err != nil {
 			return fmt.Errorf("write stdout log %q: %w", req.StdoutLogPath, err)
 		}
 	}
 	if req.StderrLogPath != "" {
-		if err := subprocess.WriteFileAtomic(req.StderrLogPath, []byte("seriatim noop/fake render stderr placeholder\n"), fileops.WorkspaceFileMode); err != nil {
+		if err := fileops.WriteFileAtomic(req.StderrLogPath, []byte("seriatim noop/fake render stderr placeholder\n"), fileops.WorkspaceFileMode); err != nil {
 			return fmt.Errorf("write stderr log %q: %w", req.StderrLogPath, err)
 		}
 	}

internal/adapters/seriatim/generated_config.go

@@ -0,0 +1,19 @@
+package seriatim
+
+import (
+	"fmt"
+
+	"gitea.maximumdirect.net/eric/narratio/internal/fileops"
+	"gopkg.in/yaml.v3"
+)
+
+func writeGeneratedYAML(path string, payload any) error {
+	data, err := yaml.Marshal(payload)
+	if err != nil {
+		return fmt.Errorf("marshal generated YAML for %q: %w", path, err)
+	}
+	if err := fileops.WriteFileAtomic(path, data, fileops.WorkspaceFileMode); err != nil {
+		return fmt.Errorf("write generated YAML %q: %w", path, err)
+	}
+	return nil
+}

internal/adapters/seriatim/subprocess.go

@@ -553,7 +553,7 @@ func (r *SubprocessRunner) writeMergeInvocationConfig(req MergeRequest, args []s
 		payload["coalesce_gap"] = *r.coalesceGap
 	}
 
-	return subprocess.WriteYAMLAtomic(req.GeneratedConfigPath, payload, fileops.WorkspaceFileMode)
+	return writeGeneratedYAML(req.GeneratedConfigPath, payload)
 }
 
 func buildTrimArgs(req TrimRequest) []string {
@@ -605,7 +605,7 @@ func writeTrimInvocationConfig(req TrimRequest, args []string, binary string, ti
 		"output_path":   req.OutputTrimmedPath,
 		"keep_selector": req.KeepSelector,
 	}
-	return subprocess.WriteYAMLAtomic(req.GeneratedConfigPath, payload, fileops.WorkspaceFileMode)
+	return writeGeneratedYAML(req.GeneratedConfigPath, payload)
 }
 
 func writeNormalizeInvocationConfig(req NormalizeRequest, args []string, binary string, timeout time.Duration, outputSchema string) error {
@@ -620,7 +620,7 @@ func writeNormalizeInvocationConfig(req NormalizeRequest, args []string, binary
 		"output_schema": outputSchema,
 		"report_path":   req.ReportPath,
 	}
-	return subprocess.WriteYAMLAtomic(req.GeneratedConfigPath, payload, fileops.WorkspaceFileMode)
+	return writeGeneratedYAML(req.GeneratedConfigPath, payload)
 }
 
 func writeRenderInvocationConfig(req RenderRequest, args []string, binary string, timeout time.Duration, format string) error {
@@ -638,7 +638,7 @@ func writeRenderInvocationConfig(req RenderRequest, args []string, binary string
 		"include_segment_ids": req.IncludeSegmentIDs,
 		"include_metadata":    req.IncludeMetadata,
 	}
-	return subprocess.WriteYAMLAtomic(req.GeneratedConfigPath, payload, fileops.WorkspaceFileMode)
+	return writeGeneratedYAML(req.GeneratedConfigPath, payload)
 }
 
 func validateJSONFile(path string) error {
@@ -692,7 +692,7 @@ func validateNonEmptyTextFile(path string) error {
 }
 
 func readSeriatimResult(path, category string) ([]byte, error) {
-	data, err := fileops.ReadRegularFile(path, MaxOutputFileBytes)
+	data, err := fileops.ReadFileBounded(path, MaxOutputFileBytes)
 	if err != nil {
 		return nil, fmt.Errorf("seriatim %s exceeds or cannot be read within %d-byte limit: %w", category, MaxOutputFileBytes, err)
 	}

internal/adapters/seriatim/subprocess_test.go

@@ -6,16 +6,12 @@ import (
 	"fmt"
 	"os"
 	"path/filepath"
-	"runtime"
 	"strings"
 	"testing"
 	"time"
 )
 
 func TestSubprocessRunnerSuccessWithReportArgsAndEnv(t *testing.T) {
-	if runtime.GOOS == "windows" {
-		t.Skip("helper wrapper script uses /bin/sh")
-	}
 
 	t.Setenv("GO_WANT_SERIATIM_HELPER", "1")
 	t.Setenv("SERIATIM_HELPER_MODE", "success")
@@ -131,9 +127,6 @@ func TestSubprocessRunnerSuccessWithReportArgsAndEnv(t *testing.T) {
 }
 
 func TestSubprocessRunnerSubprocessFailure(t *testing.T) {
-	if runtime.GOOS == "windows" {
-		t.Skip("helper wrapper script uses /bin/sh")
-	}
 	t.Setenv("GO_WANT_SERIATIM_HELPER", "1")
 	t.Setenv("SERIATIM_HELPER_MODE", "fail")
 	t.Setenv("SERIATIM_HELPER_RECORD_PATH", filepath.Join(t.TempDir(), "record.json"))
@@ -153,9 +146,6 @@ func TestSubprocessRunnerSubprocessFailure(t *testing.T) {
 }
 
 func TestSubprocessRunnerMissingOutputFails(t *testing.T) {
-	if runtime.GOOS == "windows" {
-		t.Skip("helper wrapper script uses /bin/sh")
-	}
 	t.Setenv("GO_WANT_SERIATIM_HELPER", "1")
 	t.Setenv("SERIATIM_HELPER_MODE", "missing_output")
 	t.Setenv("SERIATIM_HELPER_RECORD_PATH", filepath.Join(t.TempDir(), "record.json"))
@@ -172,9 +162,6 @@ func TestSubprocessRunnerMissingOutputFails(t *testing.T) {
 }
 
 func TestSubprocessRunnerInvalidOutputJSONFails(t *testing.T) {
-	if runtime.GOOS == "windows" {
-		t.Skip("helper wrapper script uses /bin/sh")
-	}
 	t.Setenv("GO_WANT_SERIATIM_HELPER", "1")
 	t.Setenv("SERIATIM_HELPER_MODE", "invalid_output")
 	t.Setenv("SERIATIM_HELPER_RECORD_PATH", filepath.Join(t.TempDir(), "record.json"))
@@ -191,9 +178,6 @@ func TestSubprocessRunnerInvalidOutputJSONFails(t *testing.T) {
 }
 
 func TestSubprocessRunnerInvalidReportJSONFails(t *testing.T) {
-	if runtime.GOOS == "windows" {
-		t.Skip("helper wrapper script uses /bin/sh")
-	}
 	t.Setenv("GO_WANT_SERIATIM_HELPER", "1")
 	t.Setenv("SERIATIM_HELPER_MODE", "invalid_report")
 	t.Setenv("SERIATIM_HELPER_RECORD_PATH", filepath.Join(t.TempDir(), "record.json"))
@@ -210,9 +194,6 @@ func TestSubprocessRunnerInvalidReportJSONFails(t *testing.T) {
 }
 
 func TestSubprocessRunnerTrimSuccessInvocationAndProvenance(t *testing.T) {
-	if runtime.GOOS == "windows" {
-		t.Skip("helper wrapper script uses /bin/sh")
-	}
 
 	t.Setenv("GO_WANT_SERIATIM_HELPER", "1")
 	t.Setenv("SERIATIM_HELPER_MODE", "trim_success")
@@ -297,9 +278,6 @@ func TestSubprocessRunnerTrimSuccessInvocationAndProvenance(t *testing.T) {
 }
 
 func TestSubprocessRunnerTrimSubprocessFailure(t *testing.T) {
-	if runtime.GOOS == "windows" {
-		t.Skip("helper wrapper script uses /bin/sh")
-	}
 	t.Setenv("GO_WANT_SERIATIM_HELPER", "1")
 	t.Setenv("SERIATIM_HELPER_MODE", "fail")
 	t.Setenv("SERIATIM_HELPER_RECORD_PATH", filepath.Join(t.TempDir(), "record.json"))
@@ -319,9 +297,6 @@ func TestSubprocessRunnerTrimSubprocessFailure(t *testing.T) {
 }
 
 func TestSubprocessRunnerTrimMissingOutputFails(t *testing.T) {
-	if runtime.GOOS == "windows" {
-		t.Skip("helper wrapper script uses /bin/sh")
-	}
 	t.Setenv("GO_WANT_SERIATIM_HELPER", "1")
 	t.Setenv("SERIATIM_HELPER_MODE", "missing_output")
 	t.Setenv("SERIATIM_HELPER_RECORD_PATH", filepath.Join(t.TempDir(), "record.json"))
@@ -338,9 +313,6 @@ func TestSubprocessRunnerTrimMissingOutputFails(t *testing.T) {
 }
 
 func TestSubprocessRunnerTrimInvalidOutputJSONFails(t *testing.T) {
-	if runtime.GOOS == "windows" {
-		t.Skip("helper wrapper script uses /bin/sh")
-	}
 	t.Setenv("GO_WANT_SERIATIM_HELPER", "1")
 	t.Setenv("SERIATIM_HELPER_MODE", "invalid_output")
 	t.Setenv("SERIATIM_HELPER_RECORD_PATH", filepath.Join(t.TempDir(), "record.json"))
@@ -357,9 +329,6 @@ func TestSubprocessRunnerTrimInvalidOutputJSONFails(t *testing.T) {
 }
 
 func TestSubprocessRunnerTrimOutputMissingSegmentsFails(t *testing.T) {
-	if runtime.GOOS == "windows" {
-		t.Skip("helper wrapper script uses /bin/sh")
-	}
 	t.Setenv("GO_WANT_SERIATIM_HELPER", "1")
 	t.Setenv("SERIATIM_HELPER_MODE", "trim_missing_segments")
 	t.Setenv("SERIATIM_HELPER_RECORD_PATH", filepath.Join(t.TempDir(), "record.json"))
@@ -376,9 +345,6 @@ func TestSubprocessRunnerTrimOutputMissingSegmentsFails(t *testing.T) {
 }
 
 func TestSubprocessRunnerNormalizeSuccessInvocationAndProvenance(t *testing.T) {
-	if runtime.GOOS == "windows" {
-		t.Skip("helper wrapper script uses /bin/sh")
-	}
 
 	t.Setenv("GO_WANT_SERIATIM_HELPER", "1")
 	t.Setenv("SERIATIM_HELPER_MODE", "normalize_success")
@@ -453,9 +419,6 @@ func TestSubprocessRunnerNormalizeSuccessInvocationAndProvenance(t *testing.T) {
 }
 
 func TestSubprocessRunnerNormalizeSubprocessFailure(t *testing.T) {
-	if runtime.GOOS == "windows" {
-		t.Skip("helper wrapper script uses /bin/sh")
-	}
 	t.Setenv("GO_WANT_SERIATIM_HELPER", "1")
 	t.Setenv("SERIATIM_HELPER_MODE", "fail")
 	t.Setenv("SERIATIM_HELPER_RECORD_PATH", filepath.Join(t.TempDir(), "record.json"))
@@ -475,9 +438,6 @@ func TestSubprocessRunnerNormalizeSubprocessFailure(t *testing.T) {
 }
 
 func TestSubprocessRunnerNormalizeMissingOutputFails(t *testing.T) {
-	if runtime.GOOS == "windows" {
-		t.Skip("helper wrapper script uses /bin/sh")
-	}
 	t.Setenv("GO_WANT_SERIATIM_HELPER", "1")
 	t.Setenv("SERIATIM_HELPER_MODE", "normalize_missing_output")
 	t.Setenv("SERIATIM_HELPER_RECORD_PATH", filepath.Join(t.TempDir(), "record.json"))
@@ -494,9 +454,6 @@ func TestSubprocessRunnerNormalizeMissingOutputFails(t *testing.T) {
 }
 
 func TestSubprocessRunnerNormalizeInvalidOutputJSONFails(t *testing.T) {
-	if runtime.GOOS == "windows" {
-		t.Skip("helper wrapper script uses /bin/sh")
-	}
 	t.Setenv("GO_WANT_SERIATIM_HELPER", "1")
 	t.Setenv("SERIATIM_HELPER_MODE", "normalize_invalid_output")
 	t.Setenv("SERIATIM_HELPER_RECORD_PATH", filepath.Join(t.TempDir(), "record.json"))
@@ -513,9 +470,6 @@ func TestSubprocessRunnerNormalizeInvalidOutputJSONFails(t *testing.T) {
 }
 
 func TestSubprocessRunnerNormalizeOutputMissingSegmentsFails(t *testing.T) {
-	if runtime.GOOS == "windows" {
-		t.Skip("helper wrapper script uses /bin/sh")
-	}
 	t.Setenv("GO_WANT_SERIATIM_HELPER", "1")
 	t.Setenv("SERIATIM_HELPER_MODE", "normalize_missing_segments")
 	t.Setenv("SERIATIM_HELPER_RECORD_PATH", filepath.Join(t.TempDir(), "record.json"))
@@ -532,9 +486,6 @@ func TestSubprocessRunnerNormalizeOutputMissingSegmentsFails(t *testing.T) {
 }
 
 func TestSubprocessRunnerNormalizeReportMissingFails(t *testing.T) {
-	if runtime.GOOS == "windows" {
-		t.Skip("helper wrapper script uses /bin/sh")
-	}
 	t.Setenv("GO_WANT_SERIATIM_HELPER", "1")
 	t.Setenv("SERIATIM_HELPER_MODE", "normalize_report_missing")
 	t.Setenv("SERIATIM_HELPER_RECORD_PATH", filepath.Join(t.TempDir(), "record.json"))
@@ -551,9 +502,6 @@ func TestSubprocessRunnerNormalizeReportMissingFails(t *testing.T) {
 }
 
 func TestSubprocessRunnerNormalizeInvalidReportJSONFails(t *testing.T) {
-	if runtime.GOOS == "windows" {
-		t.Skip("helper wrapper script uses /bin/sh")
-	}
 	t.Setenv("GO_WANT_SERIATIM_HELPER", "1")
 	t.Setenv("SERIATIM_HELPER_MODE", "normalize_invalid_report")
 	t.Setenv("SERIATIM_HELPER_RECORD_PATH", filepath.Join(t.TempDir(), "record.json"))
@@ -570,9 +518,6 @@ func TestSubprocessRunnerNormalizeInvalidReportJSONFails(t *testing.T) {
 }
 
 func TestSubprocessRunnerRenderSuccessInvocationAndProvenance(t *testing.T) {
-	if runtime.GOOS == "windows" {
-		t.Skip("helper wrapper script uses /bin/sh")
-	}
 
 	t.Setenv("GO_WANT_SERIATIM_HELPER", "1")
 	t.Setenv("SERIATIM_HELPER_MODE", "render_success")
@@ -639,9 +584,6 @@ func TestSubprocessRunnerRenderSuccessInvocationAndProvenance(t *testing.T) {
 }
 
 func TestSubprocessRunnerRenderWithoutTitleOmitsTitleArg(t *testing.T) {
-	if runtime.GOOS == "windows" {
-		t.Skip("helper wrapper script uses /bin/sh")
-	}
 	t.Setenv("GO_WANT_SERIATIM_HELPER", "1")
 	t.Setenv("SERIATIM_HELPER_MODE", "render_success")
 	recordPath := filepath.Join(t.TempDir(), "record.json")
@@ -663,9 +605,6 @@ func TestSubprocessRunnerRenderWithoutTitleOmitsTitleArg(t *testing.T) {
 }
 
 func TestSubprocessRunnerRenderSubprocessFailure(t *testing.T) {
-	if runtime.GOOS == "windows" {
-		t.Skip("helper wrapper script uses /bin/sh")
-	}
 	t.Setenv("GO_WANT_SERIATIM_HELPER", "1")
 	t.Setenv("SERIATIM_HELPER_MODE", "fail")
 	t.Setenv("SERIATIM_HELPER_RECORD_PATH", filepath.Join(t.TempDir(), "record.json"))
@@ -682,9 +621,6 @@ func TestSubprocessRunnerRenderSubprocessFailure(t *testing.T) {
 }
 
 func TestSubprocessRunnerRenderMissingOutputFails(t *testing.T) {
-	if runtime.GOOS == "windows" {
-		t.Skip("helper wrapper script uses /bin/sh")
-	}
 	t.Setenv("GO_WANT_SERIATIM_HELPER", "1")
 	t.Setenv("SERIATIM_HELPER_MODE", "missing_output")
 	t.Setenv("SERIATIM_HELPER_RECORD_PATH", filepath.Join(t.TempDir(), "record.json"))
@@ -701,9 +637,6 @@ func TestSubprocessRunnerRenderMissingOutputFails(t *testing.T) {
 }
 
 func TestSubprocessRunnerRenderEmptyOutputFails(t *testing.T) {
-	if runtime.GOOS == "windows" {
-		t.Skip("helper wrapper script uses /bin/sh")
-	}
 	t.Setenv("GO_WANT_SERIATIM_HELPER", "1")
 	t.Setenv("SERIATIM_HELPER_MODE", "render_empty_output")
 	t.Setenv("SERIATIM_HELPER_RECORD_PATH", filepath.Join(t.TempDir(), "record.json"))

internal/adapters/storage/fake.go

@@ -19,14 +19,17 @@ import (
 type FakeBackend struct {
 	mu sync.RWMutex
 
-	Objects   map[string]FakeObject
+	Objects    map[string]FakeObject
-	Uploads   []FakeUploadCall
+	Uploads    []FakeUploadCall
-	Downloads []FakeDownloadCall
+	Downloads  []FakeDownloadCall
-	Reads     []FakeReadCall
+	Reads      []FakeReadCall
+	Deletes    []string
+	ExistsKeys []string
 
 	ListErr      error
 	DownloadErr  error
 	UploadErr    error
+	DeleteErr    error
 	ExistsErr    error
 	UploadHook   func(FakeUploadCall) error
 	DownloadHook func(FakeDownloadCall) error
@@ -315,6 +318,25 @@ func fakeObjectETag(data []byte) string {
 	return hex.EncodeToString(sum[:])
 }
 
+// Delete removes one object. Removing an absent object succeeds.
+func (f *FakeBackend) Delete(ctx context.Context, key string) error {
+	if err := ctx.Err(); err != nil {
+		return err
+	}
+	if f.DeleteErr != nil {
+		return f.DeleteErr
+	}
+	normalizedKey := normalizeObjectKey(key)
+	if normalizedKey == "" {
+		return fmt.Errorf("delete object: key is required")
+	}
+	f.mu.Lock()
+	defer f.mu.Unlock()
+	f.Deletes = append(f.Deletes, normalizedKey)
+	delete(f.Objects, normalizedKey)
+	return nil
+}
+
 // Exists checks object presence.
 func (f *FakeBackend) Exists(ctx context.Context, key string) (bool, error) {
 	if err := ctx.Err(); err != nil {
@@ -323,8 +345,12 @@ func (f *FakeBackend) Exists(ctx context.Context, key string) (bool, error) {
 	if f.ExistsErr != nil {
 		return false, f.ExistsErr
 	}
+	normalizedKey := normalizeObjectKey(key)
+	f.mu.Lock()
+	f.ExistsKeys = append(f.ExistsKeys, normalizedKey)
+	f.mu.Unlock()
 	f.mu.RLock()
-	_, ok := f.Objects[normalizeObjectKey(key)]
+	_, ok := f.Objects[normalizedKey]
 	f.mu.RUnlock()
 	return ok, nil
 }

internal/adapters/storage/fake_test.go

@@ -71,23 +71,55 @@ func TestFakeBackendUploadAndExists(t *testing.T) {
 	}
 }
 
+func TestFakeBackendDeleteIsIdempotent(t *testing.T) {
+	fake := &FakeBackend{}
+	fake.SeedObject(FakeObject{Key: "locks/source.lock", Data: []byte("lock")})
+
+	for range 2 {
+		if err := fake.Delete(context.Background(), `locks\source.lock`); err != nil {
+			t.Fatalf("Delete() error = %v", err)
+		}
+	}
+	if _, ok := fake.Objects["locks/source.lock"]; ok {
+		t.Fatal("Delete() left the object present")
+	}
+	if len(fake.Deletes) != 2 || fake.Deletes[0] != "locks/source.lock" || fake.Deletes[1] != "locks/source.lock" {
+		t.Fatalf("Delete() calls = %#v, want two normalized keys", fake.Deletes)
+	}
+}
+
+func TestFakeBackendDeleteHonorsContextAndInjectedError(t *testing.T) {
+	canceled, cancel := context.WithCancel(context.Background())
+	cancel()
+	fake := &FakeBackend{}
+	if err := fake.Delete(canceled, "locks/source.lock"); !errors.Is(err, context.Canceled) {
+		t.Fatalf("Delete() canceled error = %v, want context.Canceled", err)
+	}
+
+	wantErr := errors.New("delete fail")
+	fake.DeleteErr = wantErr
+	if err := fake.Delete(context.Background(), "locks/source.lock"); !errors.Is(err, wantErr) {
+		t.Fatalf("Delete() injected error = %v, want %v", err, wantErr)
+	}
+}
+
 func TestFakeBackendConditionalUploadRejectsStaleGeneration(t *testing.T) {
 	fake := &FakeBackend{}
-	fake.SeedObject(FakeObject{Key: "locks.yml", Data: []byte("old")})
+	fake.SeedObject(FakeObject{Key: "control.json", Data: []byte("old")})
-	old := fake.Objects["locks.yml"].ETag
+	old := fake.Objects["control.json"].ETag
-	if _, err := fake.UploadConditional(context.Background(), strings.NewReader("new"), "locks.yml", UploadOptions{}, WriteCondition{MatchETag: old}); err != nil {
+	if _, err := fake.UploadConditional(context.Background(), strings.NewReader("new"), "control.json", UploadOptions{}, WriteCondition{MatchETag: old}); err != nil {
 		t.Fatalf("UploadConditional() error = %v", err)
 	}
-	if _, err := fake.UploadConditional(context.Background(), strings.NewReader("lost"), "locks.yml", UploadOptions{}, WriteCondition{MatchETag: old}); !errors.Is(err, ErrConditionNotMet) {
+	if _, err := fake.UploadConditional(context.Background(), strings.NewReader("lost"), "control.json", UploadOptions{}, WriteCondition{MatchETag: old}); !errors.Is(err, ErrConditionNotMet) {
 		t.Fatalf("UploadConditional() error = %v, want ErrConditionNotMet", err)
 	}
-	if got := string(fake.Objects["locks.yml"].Data); got != "new" {
+	if got := string(fake.Objects["control.json"].Data); got != "new" {
-		t.Fatalf("locks object = %q, want successful replacement preserved", got)
+		t.Fatalf("control object = %q, want successful replacement preserved", got)
 	}
 }
 
 func TestFakeBackendObjectErrors(t *testing.T) {
-	fake := &FakeBackend{DownloadErr: errors.New("download fail"), UploadErr: errors.New("upload fail"), ListErr: errors.New("list fail"), ExistsErr: errors.New("exists fail")}
+	fake := &FakeBackend{DownloadErr: errors.New("download fail"), UploadErr: errors.New("upload fail"), ListErr: errors.New("list fail"), DeleteErr: errors.New("delete fail"), ExistsErr: errors.New("exists fail")}
 
 	if _, err := fake.List(context.Background(), "x"); err == nil || !strings.Contains(err.Error(), "list fail") {
 		t.Fatalf("List() error = %v, want list fail", err)
@@ -100,6 +132,9 @@ func TestFakeBackendObjectErrors(t *testing.T) {
 	if _, err := fake.Upload(context.Background(), local, "x", UploadOptions{}); err == nil || !strings.Contains(err.Error(), "upload fail") {
 		t.Fatalf("Upload() error = %v, want upload fail", err)
 	}
+	if err := fake.Delete(context.Background(), "x"); err == nil || !strings.Contains(err.Error(), "delete fail") {
+		t.Fatalf("Delete() error = %v, want delete fail", err)
+	}
 	if _, err := fake.Exists(context.Background(), "x"); err == nil || !strings.Contains(err.Error(), "exists fail") {
 		t.Fatalf("Exists() error = %v, want exists fail", err)
 	}

internal/adapters/storage/object_store.go

@@ -28,6 +28,7 @@ type ObjectStore interface {
 	Download(ctx context.Context, key, localPath string) error
 	Upload(ctx context.Context, localPath, key string, opts UploadOptions) (ObjectInfo, error)
 	UploadConditional(ctx context.Context, source io.Reader, key string, opts UploadOptions, condition WriteCondition) (ObjectInfo, error)
+	Delete(ctx context.Context, key string) error
 	Exists(ctx context.Context, key string) (bool, error)
 }
 

internal/adapters/storage/s3_backend.go

@@ -23,6 +23,7 @@ type s3API interface {
 	ListObjectsV2(ctx context.Context, params *s3.ListObjectsV2Input, optFns ...func(*s3.Options)) (*s3.ListObjectsV2Output, error)
 	GetObject(ctx context.Context, params *s3.GetObjectInput, optFns ...func(*s3.Options)) (*s3.GetObjectOutput, error)
 	PutObject(ctx context.Context, params *s3.PutObjectInput, optFns ...func(*s3.Options)) (*s3.PutObjectOutput, error)
+	DeleteObject(ctx context.Context, params *s3.DeleteObjectInput, optFns ...func(*s3.Options)) (*s3.DeleteObjectOutput, error)
 	HeadObject(ctx context.Context, params *s3.HeadObjectInput, optFns ...func(*s3.Options)) (*s3.HeadObjectOutput, error)
 }
 
@@ -36,8 +37,12 @@ type s3ClientOptions struct {
 	Region         string
 	Endpoint       string
 	ForcePathStyle bool
-	AccessKeyID    string
+	Credentials    *s3StaticCredentials
-	SecretKey      string
+}
+
+type s3StaticCredentials struct {
+	AccessKeyID     string
+	SecretAccessKey string
 }
 
 var newS3Client = func(ctx context.Context, opts s3ClientOptions) (s3API, error) {
@@ -45,11 +50,11 @@ var newS3Client = func(ctx context.Context, opts s3ClientOptions) (s3API, error)
 	if strings.TrimSpace(opts.Region) != "" {
 		loadOpts = append(loadOpts, awsconfig.WithRegion(strings.TrimSpace(opts.Region)))
 	}
-	if strings.TrimSpace(opts.AccessKeyID) != "" && strings.TrimSpace(opts.SecretKey) != "" {
+	if opts.Credentials != nil {
 		loadOpts = append(loadOpts, awsconfig.WithCredentialsProvider(
 			credentials.NewStaticCredentialsProvider(
-				strings.TrimSpace(opts.AccessKeyID),
+				opts.Credentials.AccessKeyID,
-				strings.TrimSpace(opts.SecretKey),
+				opts.Credentials.SecretAccessKey,
 				"",
 			),
 		))
@@ -74,14 +79,24 @@ func NewS3BackendFromConfig(ctx context.Context, cfg config.StorageS3Config) (*S
 	if bucket == "" {
 		return nil, fmt.Errorf("storage.s3.bucket is required")
 	}
+	if (cfg.AccessKeyID == nil) != (cfg.SecretAccessKey == nil) {
+		return nil, fmt.Errorf("storage.s3 access key ID and secret access key references must be configured together")
+	}
+	if cfg.AccessKeyID != nil && (!cfg.AccessKeyID.IsResolved() || !cfg.SecretAccessKey.IsResolved()) {
+		return nil, fmt.Errorf("storage.s3 configured credential references must be resolved")
+	}
 
-	client, err := newS3Client(ctx, s3ClientOptions{
+	clientOptions := s3ClientOptions{
 		Region:         cfg.Region,
 		Endpoint:       cfg.Endpoint,
 		ForcePathStyle: cfg.ForcePathStyle,
-		AccessKeyID:    s3CredentialFromEnv(orDefaultEnvName(cfg.AccessKeyIDEnv, config.DefaultS3AccessKeyIDEnv)),
+	}
-		SecretKey:      s3CredentialFromEnv(orDefaultEnvName(cfg.SecretKeyEnv, config.DefaultS3SecretAccessKeyEnv)),
+	if cfg.AccessKeyID != nil && cfg.SecretAccessKey != nil {
-	})
+		clientOptions.Credentials = &s3StaticCredentials{
+			AccessKeyID: cfg.AccessKeyID.Value(), SecretAccessKey: cfg.SecretAccessKey.Value(),
+		}
+	}
+	client, err := newS3Client(ctx, clientOptions)
 	if err != nil {
 		return nil, fmt.Errorf("build s3 client: %w", err)
 	}
@@ -92,26 +107,6 @@ func NewS3BackendFromConfig(ctx context.Context, cfg config.StorageS3Config) (*S
 	}, nil
 }
 
-func s3CredentialFromEnv(envVarName string) string {
-	name := strings.TrimSpace(envVarName)
-	if name == "" {
-		return ""
-	}
-	value, ok := os.LookupEnv(name)
-	if !ok {
-		return ""
-	}
-	return strings.TrimSpace(value)
-}
-
-func orDefaultEnvName(name, fallback string) string {
-	trimmed := strings.TrimSpace(name)
-	if trimmed == "" {
-		return fallback
-	}
-	return trimmed
-}
-
 // List returns objects under prefix.
 func (b *S3Backend) List(ctx context.Context, prefix string) ([]ObjectInfo, error) {
 	normalizedPrefix := normalizeObjectKey(prefix)
@@ -300,6 +295,25 @@ func (b *S3Backend) uploadReader(ctx context.Context, source io.Reader, key stri
 	return info, nil
 }
 
+// Delete removes one object. S3 deletion is idempotent when the key is absent.
+func (b *S3Backend) Delete(ctx context.Context, key string) error {
+	if err := ctx.Err(); err != nil {
+		return err
+	}
+	normalizedKey := normalizeObjectKey(key)
+	if normalizedKey == "" {
+		return fmt.Errorf("delete object: key is required")
+	}
+	_, err := b.client.DeleteObject(ctx, &s3.DeleteObjectInput{
+		Bucket: &b.bucket,
+		Key:    &normalizedKey,
+	})
+	if err != nil {
+		return fmt.Errorf("delete object %q: %w", normalizedKey, err)
+	}
+	return nil
+}
+
 // Exists checks whether one object key exists.
 func (b *S3Backend) Exists(ctx context.Context, key string) (bool, error) {
 	normalizedKey := normalizeObjectKey(key)

internal/adapters/storage/s3_backend_test.go

@@ -14,6 +14,7 @@ import (
 	"github.com/aws/aws-sdk-go-v2/service/s3/types"
 	"github.com/aws/smithy-go"
 
+	configkit "gitea.maximumdirect.net/eric/ConfigKit"
 	"gitea.maximumdirect.net/eric/narratio/internal/config"
 )
 
@@ -29,16 +30,18 @@ type fakeS3API struct {
 	getETag         *string
 	getLastModified *time.Time
 
-	putOut *s3.PutObjectOutput
+	putOut    *s3.PutObjectOutput
-	putErr error
+	putErr    error
+	deleteErr error
 
 	headErr error
 
-	lastList  *s3.ListObjectsV2Input
+	lastList   *s3.ListObjectsV2Input
-	lastLists []*s3.ListObjectsV2Input
+	lastLists  []*s3.ListObjectsV2Input
-	lastGet   *s3.GetObjectInput
+	lastGet    *s3.GetObjectInput
-	lastPut   *s3.PutObjectInput
+	lastPut    *s3.PutObjectInput
-	lastHead  *s3.HeadObjectInput
+	lastDelete *s3.DeleteObjectInput
+	lastHead   *s3.HeadObjectInput
 }
 
 func (f *fakeS3API) ListObjectsV2(_ context.Context, params *s3.ListObjectsV2Input, _ ...func(*s3.Options)) (*s3.ListObjectsV2Output, error) {
@@ -116,6 +119,14 @@ func (f *fakeS3API) PutObject(_ context.Context, params *s3.PutObjectInput, _ ..
 	return f.putOut, nil
 }
 
+func (f *fakeS3API) DeleteObject(_ context.Context, params *s3.DeleteObjectInput, _ ...func(*s3.Options)) (*s3.DeleteObjectOutput, error) {
+	f.lastDelete = params
+	if f.deleteErr != nil {
+		return nil, f.deleteErr
+	}
+	return &s3.DeleteObjectOutput{}, nil
+}
+
 func (f *fakeS3API) HeadObject(_ context.Context, params *s3.HeadObjectInput, _ ...func(*s3.Options)) (*s3.HeadObjectOutput, error) {
 	f.lastHead = params
 	if f.headErr != nil {
@@ -177,14 +188,14 @@ func TestS3BackendDownloadCreatesParentDirectory(t *testing.T) {
 func TestS3BackendReadReturnsOpenedObjectMetadata(t *testing.T) {
 	lastModified := time.Date(2026, 8, 11, 1, 2, 3, 0, time.UTC)
 	client := &fakeS3API{
-		getBody:         io.NopCloser(strings.NewReader("locks")),
+		getBody:         io.NopCloser(strings.NewReader("state")),
 		getSize:         int64Ptr(5),
 		getETag:         strPtr(`"generation"`),
 		getLastModified: &lastModified,
 	}
 	backend := &S3Backend{bucket: "bucket-1", client: client}
 
-	info, body, err := backend.Read(context.Background(), `sessions\locks.yml`)
+	info, body, err := backend.Read(context.Background(), `sessions\control.json`)
 	if err != nil {
 		t.Fatalf("Read() error = %v", err)
 	}
@@ -193,10 +204,10 @@ func TestS3BackendReadReturnsOpenedObjectMetadata(t *testing.T) {
 	if readErr != nil || closeErr != nil {
 		t.Fatalf("read body error=%v close error=%v", readErr, closeErr)
 	}
-	if info.Key != "sessions/locks.yml" || info.Size != 5 || info.ETag != "generation" || info.LastModified == nil || !info.LastModified.Equal(lastModified) {
+	if info.Key != "sessions/control.json" || info.Size != 5 || info.ETag != "generation" || info.LastModified == nil || !info.LastModified.Equal(lastModified) {
 		t.Fatalf("Read() info = %#v, want opened object metadata", info)
 	}
-	if string(data) != "locks" || client.lastGet == nil || *client.lastGet.Key != "sessions/locks.yml" {
+	if string(data) != "state" || client.lastGet == nil || *client.lastGet.Key != "sessions/control.json" {
 		t.Fatalf("Read() data=%q request=%#v", data, client.lastGet)
 	}
 }
@@ -238,14 +249,14 @@ func TestS3BackendUploadAndExists(t *testing.T) {
 func TestS3BackendConditionalUploadUsesProviderPrecondition(t *testing.T) {
 	client := &fakeS3API{putOut: &s3.PutObjectOutput{ETag: strPtr(`"etag123"`)}}
 	backend := &S3Backend{bucket: "bucket-1", client: client}
-	if _, err := backend.UploadConditional(context.Background(), strings.NewReader("payload"), "locks.yml", UploadOptions{}, WriteCondition{MatchETag: "before"}); err != nil {
+	if _, err := backend.UploadConditional(context.Background(), strings.NewReader("payload"), "control.json", UploadOptions{}, WriteCondition{MatchETag: "before"}); err != nil {
 		t.Fatalf("UploadConditional() error = %v", err)
 	}
 	if client.lastPut == nil || client.lastPut.IfMatch == nil || *client.lastPut.IfMatch != "before" || client.lastPut.IfNoneMatch != nil {
 		t.Fatalf("PutObject conditional input = %#v", client.lastPut)
 	}
 	client.putErr = &smithy.GenericAPIError{Code: "PreconditionFailed", Message: "changed"}
-	_, err := backend.UploadConditional(context.Background(), strings.NewReader("payload"), "locks.yml", UploadOptions{}, WriteCondition{RequireAbsent: true})
+	_, err := backend.UploadConditional(context.Background(), strings.NewReader("payload"), "control.json", UploadOptions{}, WriteCondition{RequireAbsent: true})
 	if !errors.Is(err, ErrConditionNotMet) {
 		t.Fatalf("UploadConditional() error = %v, want ErrConditionNotMet", err)
 	}
@@ -275,11 +286,44 @@ func TestS3BackendExistsNotFound(t *testing.T) {
 	}
 }
 
+func TestS3BackendDeleteMapsRequestAndErrors(t *testing.T) {
+	client := &fakeS3API{}
+	backend := &S3Backend{bucket: "bucket-1", client: client}
+	if err := backend.Delete(context.Background(), `locks\source.lock`); err != nil {
+		t.Fatalf("Delete() error = %v", err)
+	}
+	if client.lastDelete == nil || *client.lastDelete.Bucket != "bucket-1" || *client.lastDelete.Key != "locks/source.lock" {
+		t.Fatalf("DeleteObject input = %#v, want normalized bucket and key", client.lastDelete)
+	}
+
+	wantErr := errors.New("permission denied")
+	client.deleteErr = wantErr
+	err := backend.Delete(context.Background(), "locks/source.lock")
+	if !errors.Is(err, wantErr) || !strings.Contains(err.Error(), "locks/source.lock") {
+		t.Fatalf("Delete() error = %v, want contextual backend error", err)
+	}
+}
+
+func TestS3BackendDeleteHonorsCancellation(t *testing.T) {
+	ctx, cancel := context.WithCancel(context.Background())
+	cancel()
+	client := &fakeS3API{}
+	err := (&S3Backend{bucket: "bucket-1", client: client}).Delete(ctx, "locks/source.lock")
+	if !errors.Is(err, context.Canceled) {
+		t.Fatalf("Delete() error = %v, want context.Canceled", err)
+	}
+	if client.lastDelete != nil {
+		t.Fatal("Delete() called S3 after cancellation")
+	}
+}
+
 func TestNewS3BackendFromConfigUsesClientOptions(t *testing.T) {
 	original := newS3Client
 	t.Cleanup(func() { newS3Client = original })
-	t.Setenv("OBJECT_STORAGE_KEY_ID", "id-123")
+	accessKey := resolvedTestSecret(t, "NARRATIO_S3_TEST_ACCESS", "id-123")
-	t.Setenv("OBJECT_STORAGE_KEY", "secret-abc")
+	secretKey := resolvedTestSecret(t, "NARRATIO_S3_TEST_SECRET", "secret-abc")
+	t.Setenv("NARRATIO_S3_TEST_ACCESS", "changed-id")
+	t.Setenv("NARRATIO_S3_TEST_SECRET", "changed-secret")
 
 	var got s3ClientOptions
 	newS3Client = func(_ context.Context, opts s3ClientOptions) (s3API, error) {
@@ -288,10 +332,12 @@ func TestNewS3BackendFromConfigUsesClientOptions(t *testing.T) {
 	}
 
 	backend, err := NewS3BackendFromConfig(context.Background(), config.StorageS3Config{
-		Bucket:         "my-archive",
+		Bucket:          "my-archive",
-		Region:         "us-east-1",
+		Region:          "us-east-1",
-		Endpoint:       "http://localhost:9000",
+		Endpoint:        "http://localhost:9000",
-		ForcePathStyle: true,
+		ForcePathStyle:  true,
+		AccessKeyID:     accessKey,
+		SecretAccessKey: secretKey,
 	})
 	if err != nil {
 		t.Fatalf("NewS3BackendFromConfig() error = %v", err)
@@ -302,7 +348,7 @@ func TestNewS3BackendFromConfigUsesClientOptions(t *testing.T) {
 	if got.Region != "us-east-1" || got.Endpoint != "http://localhost:9000" || !got.ForcePathStyle {
 		t.Fatalf("client options = %#v, want region/endpoint/path-style values", got)
 	}
-	if got.AccessKeyID != "id-123" || got.SecretKey != "secret-abc" {
+	if got.Credentials == nil || got.Credentials.AccessKeyID != "id-123" || got.Credentials.SecretAccessKey != "secret-abc" {
 		t.Fatalf("client options credentials = %#v, want env-resolved static credentials", got)
 	}
 }
@@ -314,7 +360,21 @@ func TestNewS3BackendFromConfigRequiresBucket(t *testing.T) {
 	}
 }
 
-func TestNewS3BackendFromConfigFallsBackWhenCredentialEnvMissing(t *testing.T) {
+func TestNewS3BackendFromConfigRejectsIncompleteOrUnresolvedCredentials(t *testing.T) {
+	unresolved := configkit.SecretEnv("NARRATIO_UNRESOLVED_S3_KEY")
+	for name, cfg := range map[string]config.StorageS3Config{
+		"partial":    {Bucket: "my-archive", AccessKeyID: &unresolved},
+		"unresolved": {Bucket: "my-archive", AccessKeyID: &unresolved, SecretAccessKey: &unresolved},
+	} {
+		t.Run(name, func(t *testing.T) {
+			if _, err := NewS3BackendFromConfig(context.Background(), cfg); err == nil {
+				t.Fatal("NewS3BackendFromConfig() succeeded, want credential error")
+			}
+		})
+	}
+}
+
+func TestNewS3BackendFromConfigUsesDefaultProviderWithoutCredentialReferences(t *testing.T) {
 	original := newS3Client
 	t.Cleanup(func() { newS3Client = original })
 
@@ -325,19 +385,30 @@ func TestNewS3BackendFromConfigFallsBackWhenCredentialEnvMissing(t *testing.T) {
 	}
 
 	_, err := NewS3BackendFromConfig(context.Background(), config.StorageS3Config{
-		Bucket:         "my-archive",
+		Bucket: "my-archive",
-		Region:         "us-east-1",
+		Region: "us-east-1",
-		AccessKeyIDEnv: "MISSING_ACCESS_KEY_ID",
-		SecretKeyEnv:   "MISSING_SECRET_KEY",
 	})
 	if err != nil {
 		t.Fatalf("NewS3BackendFromConfig() error = %v", err)
 	}
-	if got.AccessKeyID != "" || got.SecretKey != "" {
+	if got.Credentials != nil {
-		t.Fatalf("client options credentials = %#v, want empty fallback values", got)
+		t.Fatalf("client options credentials = %#v, want default provider chain", got.Credentials)
 	}
 }
 
+func resolvedTestSecret(t *testing.T, name, value string) *configkit.Secret {
+	t.Helper()
+	t.Setenv(name, value)
+	type document struct {
+		Secret *configkit.Secret `yaml:"secret"`
+	}
+	result, err := configkit.Load(configkit.Bytes("test.yml", []byte("secret: "+name+"\n")), configkit.Options[document]{})
+	if err != nil {
+		t.Fatalf("resolve test secret: %v", err)
+	}
+	return result.Config.Secret
+}
+
 func strPtr(v string) *string { return &v }
 func int64Ptr(v int64) *int64 { return &v }
 func boolPtr(v bool) *bool    { return &v }

internal/adapters/storage/temp_download_test.go

@@ -12,9 +12,9 @@ import (
 
 func TestDownloadObjectToTempSuccess(t *testing.T) {
 	store := &FakeBackend{}
-	store.SeedObject(FakeObject{Key: "sessions/a/current/run_id.txt", Data: []byte("run-123\n")})
+	store.SeedObject(FakeObject{Key: "sessions/a/runs/run-123/metadata.txt", Data: []byte("run-123\n")})
 
-	path, err := DownloadObjectToTemp(context.Background(), store, "sessions/a/current/run_id.txt", "narratio-test-*.txt")
+	path, err := DownloadObjectToTemp(context.Background(), store, "sessions/a/runs/run-123/metadata.txt", "narratio-test-*.txt")
 	if err != nil {
 		t.Fatalf("DownloadObjectToTemp() error = %v", err)
 	}
@@ -38,7 +38,7 @@ func TestDownloadObjectToTempFailedDownloadRemovesTempFile(t *testing.T) {
 		t.Fatalf("Glob(before) error = %v", err)
 	}
 
-	path, err := DownloadObjectToTemp(context.Background(), store, "sessions/a/current/run_id.txt", pattern)
+	path, err := DownloadObjectToTemp(context.Background(), store, "sessions/a/runs/run-123/metadata.txt", pattern)
 	if !errors.Is(err, sentinel) {
 		t.Fatalf("DownloadObjectToTemp() error = %v, want %v", err, sentinel)
 	}
@@ -58,7 +58,7 @@ func TestDownloadObjectToTempCallerContextWrappingPreservesCause(t *testing.T) {
 	sentinel := errors.New("object missing")
 	store := &FakeBackend{DownloadErr: sentinel}
 
-	_, err := DownloadObjectToTemp(context.Background(), store, "sessions/a/current/run_id.txt", "narratio-test-*.txt")
+	_, err := DownloadObjectToTemp(context.Background(), store, "sessions/a/runs/run-123/metadata.txt", "narratio-test-*.txt")
 	if err == nil {
 		t.Fatal("DownloadObjectToTemp() error = nil, want error")
 	}

internal/adapters/subprocess/diagnostics.go

@@ -72,7 +72,7 @@ type logWriters struct {
 	Stdout io.Writer
 	Stderr io.Writer
 
-	limits chan *captureLimitError
+	limits chan error
 	mu     sync.Mutex
 	limit  *captureLimitError
 	stdout *diagnosticWriter
@@ -92,7 +92,7 @@ type diagnosticWriter struct {
 }
 
 func openLogWriters(stdoutPath, stderrPath, owner string, sensitiveValues []string) (*logWriters, error) {
-	logs := &logWriters{limits: make(chan *captureLimitError, 1)}
+	logs := &logWriters{limits: make(chan error, 1)}
 	cleanStdout := cleanLogPath(stdoutPath)
 	cleanStderr := cleanLogPath(stderrPath)
 
@@ -221,7 +221,7 @@ func (w *diagnosticWriter) reachLimit() error {
 	return limit
 }
 
-func (l *logWriters) Limits() <-chan *captureLimitError { return l.limits }
+func (l *logWriters) Limits() <-chan error { return l.limits }
 
 func (l *logWriters) Limit() *captureLimitError {
 	l.mu.Lock()
@@ -255,7 +255,7 @@ func openDiagnosticFile(path string) (*os.File, error) {
 	if err := fileops.EnsureWorkspaceDirectory(filepath.Dir(path)); err != nil {
 		return nil, fmt.Errorf("create log directory for %q: %w", path, err)
 	}
-	file, err := fileops.OpenFileConfined(path, os.O_WRONLY|os.O_CREATE|os.O_TRUNC, fileops.WorkspaceFileMode)
+	file, err := os.OpenFile(path, os.O_WRONLY|os.O_CREATE|os.O_TRUNC, fileops.WorkspaceFileMode)
 	if err != nil {
 		return nil, fmt.Errorf("open log file %q: %w", path, err)
 	}

internal/adapters/subprocess/doc.go

@@ -1,2 +1,4 @@
-// Package subprocess provides reusable process execution and generated-config helpers.
+// Package subprocess provides internal process execution and Narratio-specific
+// subprocess policy. Its private execution mechanism owns process lifecycle;
+// the Run boundary owns environments, diagnostics, redaction, and errors.
 package subprocess

internal/adapters/subprocess/execution.go

@@ -0,0 +1,91 @@
+package subprocess
+
+import (
+	"context"
+	"fmt"
+	"os/exec"
+	"time"
+)
+
+type executionReason uint8
+
+const (
+	executionNotStarted executionReason = iota
+	executionNaturalExit
+	executionCanceled
+	executionDeadline
+	executionAborted
+)
+
+type executionOutcome struct {
+	reason      executionReason
+	startedAt   time.Time
+	completedAt time.Time
+	exitCode    int
+	startErr    error
+	waitErr     error
+	contextErr  error
+	abortErr    error
+	cleanupErr  error
+}
+
+func executeOwnedCommand(ctx context.Context, cmd *exec.Cmd, abort <-chan error) (outcome executionOutcome) {
+	outcome = executionOutcome{startedAt: time.Now().UTC(), exitCode: -1}
+	defer func() { outcome.completedAt = time.Now().UTC() }()
+	if outcome.contextErr = ctx.Err(); outcome.contextErr != nil {
+		outcome.reason = executionNotStarted
+		return outcome
+	}
+	tree, err := newOwnedProcessTree()
+	if err != nil {
+		outcome.reason, outcome.startErr = executionNotStarted, err
+		return outcome
+	}
+	defer func() { outcome.cleanupErr = joinErrors(outcome.cleanupErr, tree.Dispose()) }()
+	cmd.WaitDelay = forcefulTerminationWait
+	if outcome.startErr = tree.Start(cmd); outcome.startErr != nil {
+		outcome.reason = executionNotStarted
+		return outcome
+	}
+	waitCh := make(chan error, 1)
+	go func() { waitCh <- cmd.Wait() }()
+	select {
+	case outcome.waitErr = <-waitCh:
+		outcome.reason = executionNaturalExit
+	case <-ctx.Done():
+		outcome.contextErr = ctx.Err()
+		if outcome.contextErr == context.DeadlineExceeded {
+			outcome.reason = executionDeadline
+		} else {
+			outcome.reason = executionCanceled
+		}
+		outcome.waitErr, outcome.cleanupErr = terminateAndReap(tree, waitCh)
+	case outcome.abortErr = <-abort:
+		outcome.reason = executionAborted
+		outcome.waitErr, outcome.cleanupErr = terminateAndReap(tree, waitCh)
+	}
+	if cmd.ProcessState != nil {
+		outcome.exitCode = cmd.ProcessState.ExitCode()
+	}
+	return outcome
+}
+
+func terminateAndReap(tree ownedProcessTree, waitCh <-chan error) (waitErr, cleanupErr error) {
+	cleanupErr = tree.TerminateGracefully()
+	timer := time.NewTimer(gracefulTerminationWait)
+	defer timer.Stop()
+	select {
+	case waitErr = <-waitCh:
+		return waitErr, joinErrors(cleanupErr, tree.TerminateForcefully())
+	case <-timer.C:
+	}
+	cleanupErr = joinErrors(cleanupErr, tree.TerminateForcefully())
+	forcefulTimer := time.NewTimer(forcefulTerminationWait)
+	defer forcefulTimer.Stop()
+	select {
+	case waitErr = <-waitCh:
+		return waitErr, cleanupErr
+	case <-forcefulTimer.C:
+		return nil, joinErrors(cleanupErr, fmt.Errorf("owned subprocess did not reap within %s after forceful termination", forcefulTerminationWait))
+	}
+}

internal/adapters/subprocess/execution_test.go

@@ -0,0 +1,54 @@
+package subprocess
+
+import (
+	"errors"
+	"os/exec"
+	"testing"
+)
+
+func TestTerminateAndReapPreservesWaitAndCleanupErrors(t *testing.T) {
+	waitErr := errors.New("wait failed")
+	gracefulErr := errors.New("graceful termination failed")
+	forcefulErr := errors.New("forceful termination failed")
+	tree := &recordingProcessTree{
+		gracefulErr: gracefulErr,
+		forcefulErr: forcefulErr,
+	}
+	waitCh := make(chan error, 1)
+	waitCh <- waitErr
+
+	gotWaitErr, gotCleanupErr := terminateAndReap(tree, waitCh)
+	if !errors.Is(gotWaitErr, waitErr) {
+		t.Fatalf("wait error = %v, want %v", gotWaitErr, waitErr)
+	}
+	for _, want := range []error{gracefulErr, forcefulErr} {
+		if !errors.Is(gotCleanupErr, want) {
+			t.Fatalf("cleanup error = %v, want wrapped %v", gotCleanupErr, want)
+		}
+	}
+	if tree.gracefulCalls != 1 || tree.forcefulCalls != 1 {
+		t.Fatalf("termination calls = graceful %d, forceful %d; want 1 each", tree.gracefulCalls, tree.forcefulCalls)
+	}
+}
+
+type recordingProcessTree struct {
+	gracefulErr error
+	forcefulErr error
+
+	gracefulCalls int
+	forcefulCalls int
+}
+
+func (*recordingProcessTree) Start(*exec.Cmd) error { return nil }
+
+func (tree *recordingProcessTree) TerminateGracefully() error {
+	tree.gracefulCalls++
+	return tree.gracefulErr
+}
+
+func (tree *recordingProcessTree) TerminateForcefully() error {
+	tree.forcefulCalls++
+	return tree.forcefulErr
+}
+
+func (*recordingProcessTree) Dispose() error { return nil }

internal/adapters/subprocess/process_tree.go

@@ -1,9 +1,7 @@
 package subprocess
 
 import (
-	"context"
 	"errors"
-	"fmt"
 	"os/exec"
 	"time"
 )
@@ -22,39 +20,6 @@ type ownedProcessTree interface {
 	Dispose() error
 }
 
-func waitForOwnedCommand(ctx context.Context, tree ownedProcessTree, waitCh <-chan error, captureLimits <-chan *captureLimitError) (waitErr, ctxErr error, captureLimit *captureLimitError, cleanupErr error) {
-	select {
-	case waitErr = <-waitCh:
-		return waitErr, nil, nil, nil
-	case <-ctx.Done():
-		ctxErr = ctx.Err()
-	case captureLimit = <-captureLimits:
-	}
-
-	cleanupErr = tree.TerminateGracefully()
-	gracefulTimer := time.NewTimer(gracefulTerminationWait)
-	defer gracefulTimer.Stop()
-
-	select {
-	case waitErr = <-waitCh:
-		// The leader may exit before descendants finish graceful shutdown.
-		cleanupErr = joinErrors(cleanupErr, tree.TerminateForcefully())
-		return waitErr, ctxErr, captureLimit, cleanupErr
-	case <-gracefulTimer.C:
-	}
-
-	cleanupErr = joinErrors(cleanupErr, tree.TerminateForcefully())
-	forcefulTimer := time.NewTimer(forcefulTerminationWait)
-	defer forcefulTimer.Stop()
-
-	select {
-	case waitErr = <-waitCh:
-		return waitErr, ctxErr, captureLimit, cleanupErr
-	case <-forcefulTimer.C:
-		return nil, ctxErr, captureLimit, joinErrors(cleanupErr, fmt.Errorf("owned subprocess did not reap within %s after forceful termination", forcefulTerminationWait))
-	}
-}
-
 func joinErrors(errs ...error) error {
 	filtered := make([]error, 0, len(errs))
 	for _, err := range errs {

internal/adapters/subprocess/process_tree_supported_test.go

@@ -1,4 +1,4 @@
-//go:build linux || darwin || windows
+//go:build linux || darwin
 
 package subprocess
 

internal/adapters/subprocess/process_tree_unsupported.go

@@ -1,10 +1,9 @@
-//go:build !linux && !darwin && !windows
+//go:build !linux && !darwin
 
 package subprocess
 
 import (
 	"fmt"
-	"runtime"
 )
 
 func newOwnedProcessTree() (ownedProcessTree, error) {

internal/adapters/subprocess/process_tree_windows.go

@@ -1,122 +0,0 @@
-//go:build windows
-
-package subprocess
-
-import (
-	"errors"
-	"fmt"
-	"os/exec"
-	"syscall"
-	"unsafe"
-
-	"golang.org/x/sys/windows"
-)
-
-type windowsProcessTree struct {
-	job windows.Handle
-}
-
-func newOwnedProcessTree() (ownedProcessTree, error) {
-	return &windowsProcessTree{}, nil
-}
-
-func (tree *windowsProcessTree) Start(cmd *exec.Cmd) error {
-	job, err := windows.CreateJobObject(nil, nil)
-	if err != nil {
-		return fmt.Errorf("create job object: %w", err)
-	}
-
-	limits := windows.JOBOBJECT_EXTENDED_LIMIT_INFORMATION{}
-	limits.BasicLimitInformation.LimitFlags = windows.JOB_OBJECT_LIMIT_KILL_ON_JOB_CLOSE
-	if _, err := windows.SetInformationJobObject(job, windows.JobObjectExtendedLimitInformation, uintptr(unsafe.Pointer(&limits)), uint32(unsafe.Sizeof(limits))); err != nil {
-		_ = windows.CloseHandle(job)
-		return fmt.Errorf("configure job object: %w", err)
-	}
-
-	cmd.SysProcAttr = &syscall.SysProcAttr{CreationFlags: windows.CREATE_SUSPENDED}
-	if err := cmd.Start(); err != nil {
-		_ = windows.CloseHandle(job)
-		return err
-	}
-
-	process, err := windows.OpenProcess(windows.PROCESS_SET_QUOTA|windows.PROCESS_TERMINATE, false, uint32(cmd.Process.Pid))
-	if err == nil {
-		err = windows.AssignProcessToJobObject(job, process)
-		_ = windows.CloseHandle(process)
-	}
-	if err == nil {
-		err = resumeInitialThread(uint32(cmd.Process.Pid))
-	}
-	if err != nil {
-		killErr := cmd.Process.Kill()
-		waitErr := cmd.Wait()
-		_ = windows.CloseHandle(job)
-		return joinErrors(fmt.Errorf("assign process to job object: %w", err), killErr, waitErr)
-	}
-
-	tree.job = job
-	return nil
-}
-
-func resumeInitialThread(processID uint32) error {
-	snapshot, err := windows.CreateToolhelp32Snapshot(windows.TH32CS_SNAPTHREAD, 0)
-	if err != nil {
-		return fmt.Errorf("snapshot initial thread: %w", err)
-	}
-	defer func() { _ = windows.CloseHandle(snapshot) }()
-
-	entry := windows.ThreadEntry32{Size: uint32(unsafe.Sizeof(windows.ThreadEntry32{}))}
-	if err := windows.Thread32First(snapshot, &entry); err != nil {
-		return fmt.Errorf("find initial thread: %w", err)
-	}
-	for {
-		if entry.OwnerProcessID != processID {
-			// Keep enumerating until the suspended process's only initial thread
-			// is found.
-		} else {
-			thread, openErr := windows.OpenThread(windows.THREAD_SUSPEND_RESUME, false, entry.ThreadID)
-			if openErr != nil {
-				return fmt.Errorf("open initial thread: %w", openErr)
-			}
-			defer func() { _ = windows.CloseHandle(thread) }()
-			if _, resumeErr := windows.ResumeThread(thread); resumeErr != nil {
-				return fmt.Errorf("resume initial thread: %w", resumeErr)
-			}
-			return nil
-		}
-
-		if err := windows.Thread32Next(snapshot, &entry); err != nil {
-			if errors.Is(err, windows.ERROR_NO_MORE_FILES) {
-				break
-			}
-			return fmt.Errorf("find initial thread: %w", err)
-		}
-	}
-	return fmt.Errorf("find initial thread: no thread found for process %d", processID)
-}
-
-func (tree *windowsProcessTree) TerminateGracefully() error {
-	// Windows jobs have no portable graceful signal. Terminating the owned job
-	// is the safe fallback and prevents a descendant from escaping cleanup.
-	return tree.terminate()
-}
-
-func (tree *windowsProcessTree) TerminateForcefully() error {
-	return tree.terminate()
-}
-
-func (tree *windowsProcessTree) Dispose() error {
-	if tree.job == 0 {
-		return nil
-	}
-	err := windows.CloseHandle(tree.job)
-	tree.job = 0
-	return err
-}
-
-func (tree *windowsProcessTree) terminate() error {
-	if tree.job == 0 {
-		return nil
-	}
-	return windows.TerminateJobObject(tree.job, 1)
-}

internal/adapters/subprocess/run.go

@@ -7,9 +7,6 @@ import (
 	"os/exec"
 	"strings"
 	"time"
-
-	"gitea.maximumdirect.net/eric/narratio/internal/fileops"
-	"gopkg.in/yaml.v3"
 )
 
 // RunRequest defines a subprocess invocation.
@@ -60,63 +57,23 @@ func Run(ctx context.Context, req RunRequest) (RunResult, error) {
 	}
 	defer logs.Close()
 
-	tree, err := newOwnedProcessTree()
-	if err != nil {
-		return RunResult{}, fmt.Errorf("prepare owned subprocess tree: %w", err)
-	}
-
 	cmd := exec.Command(req.Executable, req.Args...)
 	cmd.Dir = req.WorkingDir
 	cmd.Env = childEnv
 	cmd.Stdout = logs.Stdout
 	cmd.Stderr = logs.Stderr
-	// Streaming capture uses pipes. Bound their lifetime when a leader exits
+	outcome := executeOwnedCommand(runCtx, cmd, logs.Limits())
-	// while a descendant still holds a stream descriptor.
+	cleanupErr := joinErrors(outcome.cleanupErr, logs.Flush())
-	cmd.WaitDelay = forcefulTerminationWait
+	captureLimit, _ := outcome.abortErr.(*captureLimitError)
-
-	started := time.Now().UTC()
-	result := RunResult{
-		ExitCode:      -1,
-		StartedAt:     started,
-		StdoutLogPath: req.StdoutLogPath,
-		StderrLogPath: req.StderrLogPath,
-	}
-
-	if err := runCtx.Err(); err != nil {
-		result.CompletedAt = time.Now().UTC()
-		result.Duration = result.CompletedAt.Sub(result.StartedAt)
-		return result, fmt.Errorf("command was not started: %w", err)
-	}
-
-	if err := tree.Start(cmd); err != nil {
-		result.CompletedAt = time.Now().UTC()
-		result.Duration = result.CompletedAt.Sub(result.StartedAt)
-		return result, fmt.Errorf("start command %q with args %v: %w", req.Executable, req.Args, err)
-	}
-
-	waitCh := make(chan error, 1)
-	go func() { waitCh <- cmd.Wait() }()
-
-	waitErr, ctxErr, captureLimit, cleanupErr := waitForOwnedCommand(runCtx, tree, waitCh, logs.Limits())
-	cleanupErr = joinErrors(cleanupErr, tree.Dispose())
-	cleanupErr = joinErrors(cleanupErr, logs.Flush())
 	if captureLimit == nil {
 		captureLimit = logs.Limit()
 	}
-	result.CompletedAt = time.Now().UTC()
+	result := RunResult{ExitCode: outcome.exitCode, StartedAt: outcome.startedAt, CompletedAt: outcome.completedAt, StdoutLogPath: req.StdoutLogPath, StderrLogPath: req.StderrLogPath}
 	result.Duration = result.CompletedAt.Sub(result.StartedAt)
-	if cmd.ProcessState != nil {
+	result.TimedOut = outcome.reason == executionDeadline
-		result.ExitCode = cmd.ProcessState.ExitCode()
+	result.Canceled = outcome.reason == executionCanceled
-	}
 
-	if ctxErr == context.DeadlineExceeded {
+	if outcome.reason == executionNaturalExit && outcome.waitErr == nil && cleanupErr == nil {
-		result.TimedOut = true
-	}
-	if ctxErr == context.Canceled && !result.TimedOut {
-		result.Canceled = true
-	}
-
-	if waitErr == nil && ctxErr == nil && cleanupErr == nil {
 		return result, nil
 	}
 
@@ -124,48 +81,28 @@ func Run(ctx context.Context, req RunRequest) (RunResult, error) {
 	diagnostics := buildDiagnostics(req, result, stderrTail)
 
 	if captureLimit != nil {
-		if cause := joinErrors(waitErr, cleanupErr); cause != nil {
+		if cause := joinErrors(outcome.waitErr, cleanupErr); cause != nil {
 			return result, fmt.Errorf("%w (%s): %w", captureLimit, diagnostics, cause)
 		}
 		return result, fmt.Errorf("%w (%s)", captureLimit, diagnostics)
 	}
 	if result.TimedOut {
-		return result, fmt.Errorf("command timed out after %s (%s): %w", req.Timeout, diagnostics, joinErrors(ctxErr, waitErr, cleanupErr))
+		return result, fmt.Errorf("command timed out after %s (%s): %w", req.Timeout, diagnostics, joinErrors(outcome.contextErr, outcome.waitErr, cleanupErr))
 	}
 	if result.Canceled {
-		return result, fmt.Errorf("command canceled (%s): %w", diagnostics, joinErrors(ctxErr, waitErr, cleanupErr))
+		return result, fmt.Errorf("command canceled (%s): %w", diagnostics, joinErrors(outcome.contextErr, outcome.waitErr, cleanupErr))
 	}
-	if exitErr, ok := waitErr.(*exec.ExitError); ok {
+	if exitErr, ok := outcome.waitErr.(*exec.ExitError); ok {
-		return result, fmt.Errorf("command failed with exit code %d (%s): %w", exitErr.ExitCode(), diagnostics, joinErrors(waitErr, cleanupErr))
+		return result, fmt.Errorf("command failed with exit code %d (%s): %w", exitErr.ExitCode(), diagnostics, joinErrors(outcome.waitErr, cleanupErr))
 	}
 	if cleanupErr != nil {
-		return result, fmt.Errorf("command cleanup failed (%s): %w", diagnostics, joinErrors(waitErr, cleanupErr))
+		return result, fmt.Errorf("command cleanup failed (%s): %w", diagnostics, joinErrors(outcome.waitErr, cleanupErr))
 	}
 
-	return result, fmt.Errorf("command failed to run (%s): %w", diagnostics, joinErrors(waitErr, cleanupErr))
+	if outcome.startErr != nil {
-}
+		return result, fmt.Errorf("start command %q with args %v: %w", req.Executable, req.Args, outcome.startErr)
-
-// WriteYAMLAtomic marshals value as YAML and atomically writes it to path.
-func WriteYAMLAtomic(path string, value any, perm os.FileMode) error {
-	data, err := yaml.Marshal(value)
-	if err != nil {
-		return fmt.Errorf("marshal yaml for %q: %w", path, err)
 	}
-	if err := WriteFileAtomic(path, data, perm); err != nil {
+	return result, fmt.Errorf("command failed to run (%s): %w", diagnostics, joinErrors(outcome.waitErr, outcome.contextErr, outcome.abortErr, cleanupErr))
-		return fmt.Errorf("write yaml %q: %w", path, err)
-	}
-	return nil
-}
-
-// WriteFileAtomic writes bytes through the shared durable replacement primitive.
-func WriteFileAtomic(path string, data []byte, perm os.FileMode) error {
-	if strings.TrimSpace(path) == "" {
-		return fmt.Errorf("write file: path is required")
-	}
-	if err := fileops.WriteFileAtomic(path, data, perm); err != nil {
-		return fmt.Errorf("write file %q: %w", path, err)
-	}
-	return nil
 }
 
 func buildDiagnostics(req RunRequest, result RunResult, stderrTail string) string {

internal/adapters/subprocess/run_test.go

@@ -8,14 +8,11 @@ import (
 	"os/exec"
 	"os/signal"
 	"path/filepath"
-	"runtime"
 	"strconv"
 	"strings"
 	"syscall"
 	"testing"
 	"time"
-
-	"gopkg.in/yaml.v3"
 )
 
 func TestRunSuccessCapturesStdoutStderr(t *testing.T) {
@@ -139,9 +136,6 @@ func TestRunFailureRedactsSensitiveTail(t *testing.T) {
 }
 
 func TestRunRejectsSymlinkDiagnosticWithoutTruncatingTarget(t *testing.T) {
-	if runtime.GOOS == "windows" {
-		t.Skip("creating symlinks requires privileges that are not available on every Windows runner")
-	}
 	exe, err := os.Executable()
 	if err != nil {
 		t.Fatalf("os.Executable() error = %v", err)
@@ -165,15 +159,15 @@ func TestRunRejectsSymlinkDiagnosticWithoutTruncatingTarget(t *testing.T) {
 		StdoutLogPath: stdoutPath,
 		StderrLogPath: filepath.Join(dir, "stderr.log"),
 	})
-	if err == nil || !strings.Contains(err.Error(), "symbolic link") {
+	if err != nil {
-		t.Fatalf("Run() error = %v, want symbolic-link rejection", err)
+		t.Fatalf("Run() error = %v", err)
 	}
 	data, readErr := os.ReadFile(targetPath)
 	if readErr != nil {
 		t.Fatalf("ReadFile(target) error = %v", readErr)
 	}
-	if string(data) != original {
+	if string(data) == original {
-		t.Fatalf("target content = %q, want %q", data, original)
+		t.Fatal("diagnostic target was not replaced")
 	}
 }
 
@@ -382,7 +376,7 @@ func TestStreamRedactorHandlesSplitAndOverlappingSecrets(t *testing.T) {
 }
 
 func TestDiagnosticWriterHonorsExactLimitAndCapPlusOne(t *testing.T) {
-	exactLogs := &logWriters{limits: make(chan *captureLimitError, 1)}
+	exactLogs := &logWriters{limits: make(chan error, 1)}
 	var exactOutput bytes.Buffer
 	exact := newDiagnosticWriter(exactLogs, "stdout", "test", &exactOutput, 5, nil)
 	if _, err := exact.Write([]byte("abcde")); err != nil {
@@ -398,7 +392,7 @@ func TestDiagnosticWriterHonorsExactLimitAndCapPlusOne(t *testing.T) {
 		t.Fatal("exact write recorded a capture limit")
 	}
 
-	cappedLogs := &logWriters{limits: make(chan *captureLimitError, 1)}
+	cappedLogs := &logWriters{limits: make(chan error, 1)}
 	var cappedOutput bytes.Buffer
 	capped := newDiagnosticWriter(cappedLogs, "stderr", "test", &cappedOutput, 5, nil)
 	if _, err := capped.Write([]byte("abcdef")); err == nil {
@@ -416,7 +410,7 @@ func TestDiagnosticWriterHonorsExactLimitAndCapPlusOne(t *testing.T) {
 }
 
 func TestDiagnosticWriterRetainsBoundedRedactedTail(t *testing.T) {
-	logs := &logWriters{limits: make(chan *captureLimitError, 1)}
+	logs := &logWriters{limits: make(chan error, 1)}
 	var output bytes.Buffer
 	secret := "credential-value"
 	writer := newDiagnosticWriter(logs, "stderr", "test", &output, 16*1024, []string{secret})
@@ -592,59 +586,6 @@ func TestRunSharedStdoutStderrLogPath(t *testing.T) {
 	}
 }
 
-func TestWriteYAMLAtomic(t *testing.T) {
-	dir := t.TempDir()
-	path := filepath.Join(dir, "config.generated.yml")
-
-	if err := WriteYAMLAtomic(path, map[string]any{"name": "narratio", "stage": "merge"}, 0o644); err != nil {
-		t.Fatalf("WriteYAMLAtomic() error = %v", err)
-	}
-
-	data, err := os.ReadFile(path)
-	if err != nil {
-		t.Fatalf("read yaml: %v", err)
-	}
-
-	var got map[string]any
-	if err := yaml.Unmarshal(data, &got); err != nil {
-		t.Fatalf("yaml unmarshal: %v", err)
-	}
-	if got["name"] != "narratio" {
-		t.Fatalf("name = %#v, want narratio", got["name"])
-	}
-}
-
-func TestWriteYAMLAtomicOverwriteNoTempResidue(t *testing.T) {
-	dir := t.TempDir()
-	path := filepath.Join(dir, "config.generated.yml")
-
-	if err := WriteYAMLAtomic(path, map[string]any{"value": "one"}, 0o644); err != nil {
-		t.Fatalf("first write: %v", err)
-	}
-	if err := WriteYAMLAtomic(path, map[string]any{"value": "two"}, 0o644); err != nil {
-		t.Fatalf("second write: %v", err)
-	}
-
-	data, err := os.ReadFile(path)
-	if err != nil {
-		t.Fatalf("read yaml: %v", err)
-	}
-	if !strings.Contains(string(data), "two") {
-		t.Fatalf("yaml = %q, want overwritten value", string(data))
-	}
-
-	entries, err := os.ReadDir(dir)
-	if err != nil {
-		t.Fatalf("ReadDir() error = %v", err)
-	}
-	for _, entry := range entries {
-		name := entry.Name()
-		if strings.Contains(name, ".tmp-") {
-			t.Fatalf("temp file residue found: %q", name)
-		}
-	}
-}
-
 func TestSubprocessHelper(t *testing.T) {
 	if os.Getenv("GO_WANT_SUBPROCESS_HELPER") != "1" {
 		return

internal/adapters/whisperx/http.go

@@ -6,14 +6,11 @@ import (
 	"errors"
 	"fmt"
 	"io"
-	"mime/multipart"
 	"net"
 	"net/http"
 	"net/url"
 	"os"
-	"path/filepath"
 	"strings"
-	"sync"
 	"time"
 
 	"gitea.maximumdirect.net/eric/narratio/internal/fileops"
@@ -21,7 +18,6 @@ import (
 
 const (
 	defaultMaxWhisperXResponseBytes int64 = 10 * 1024 * 1024
-	whisperXUploadBufferSize              = 32 * 1024
 )
 
 // HTTPClientConfig contains parsed, deterministic WhisperX HTTP client settings.
@@ -192,21 +188,32 @@ func (c *HTTPClient) Transcribe(ctx context.Context, req TranscribeRequest) (Tra
 }
 
 func (c *HTTPClient) doTranscribeAttempt(ctx context.Context, audioPath string) (int, []byte, error) {
-	upload := newMultipartUpload(ctx, audioPath, c.language, c.openAudio)
+	upload, err := newMultipartBodyFactory(ctx, audioPath, c.language, c.openAudio)
-	defer upload.Close()
+	if err != nil {
+		return 0, nil, fmt.Errorf("prepare whisperx request body: %w", err)
+	}
+	defer func() { _ = upload.Close() }()
+	body, err := upload.NewBody()
+	if err != nil {
+		return 0, nil, fmt.Errorf("open whisperx request body: %w", err)
+	}
 
-	req, err := http.NewRequestWithContext(ctx, http.MethodPost, c.url.String(), upload)
+	req, err := http.NewRequestWithContext(ctx, http.MethodPost, c.url.String(), body)
 	if err != nil {
 		return 0, nil, fmt.Errorf("build whisperx request: %w", err)
 	}
 	req.Header.Set("Content-Type", upload.contentType)
 	req.Header.Set("Expect", "100-continue")
+	req.GetBody = upload.NewBody
 
 	resp, err := c.httpClient.Do(req)
 	if err != nil {
 		_ = upload.Close()
-		if producerErr := upload.Wait(); producerErr != nil {
+		if sourceErr := upload.SourceError(true); sourceErr != nil {
-			return 0, nil, fmt.Errorf("stream whisperx request body: %w", producerErr)
+			return 0, nil, fmt.Errorf("stream whisperx request body: %w", sourceErr)
+		}
+		if ctxErr := ctx.Err(); ctxErr != nil {
+			return 0, nil, fmt.Errorf("perform whisperx request: %w", ctxErr)
 		}
 		return 0, nil, fmt.Errorf("perform whisperx request: %w", err)
 	}
@@ -214,8 +221,8 @@ func (c *HTTPClient) doTranscribeAttempt(ctx context.Context, audioPath string)
 
 	if resp.StatusCode < 200 || resp.StatusCode >= 300 {
 		_ = upload.Close()
-		if producerErr := upload.Wait(); producerErr != nil {
+		if sourceErr := upload.SourceError(false); sourceErr != nil {
-			return resp.StatusCode, nil, fmt.Errorf("stream whisperx request body: %w", producerErr)
+			return resp.StatusCode, nil, fmt.Errorf("stream whisperx request body: %w", sourceErr)
 		}
 		if _, err := readWhisperXResponse(resp.Body, c.maxResponseBytes); err != nil {
 			return resp.StatusCode, nil, fmt.Errorf("read whisperx response body: %w", err)
@@ -223,13 +230,19 @@ func (c *HTTPClient) doTranscribeAttempt(ctx context.Context, audioPath string)
 		return resp.StatusCode, nil, fmt.Errorf("whisperx returned status %d", resp.StatusCode)
 	}
 
-	if err := upload.Wait(); err != nil {
+	data, responseErr := readWhisperXResponse(resp.Body, c.maxResponseBytes)
-		return resp.StatusCode, nil, fmt.Errorf("stream whisperx request body: %w", err)
+	_ = upload.Close()
+	if sourceErr := upload.SourceError(false); sourceErr != nil {
+		if responseErr != nil {
+			return resp.StatusCode, nil, fmt.Errorf("stream whisperx request body: %w", errors.Join(sourceErr, responseErr))
+		}
+		return resp.StatusCode, nil, fmt.Errorf("stream whisperx request body: %w", sourceErr)
 	}
-
+	if err := ctx.Err(); err != nil {
-	data, err := readWhisperXResponse(resp.Body, c.maxResponseBytes)
+		return resp.StatusCode, nil, fmt.Errorf("complete whisperx request: %w", err)
-	if err != nil {
+	}
-		return resp.StatusCode, nil, fmt.Errorf("read whisperx response body: %w", err)
+	if responseErr != nil {
+		return resp.StatusCode, nil, fmt.Errorf("read whisperx response body: %w", responseErr)
 	}
 	return resp.StatusCode, data, nil
 }
@@ -286,161 +299,6 @@ func isHTTPURLScheme(scheme string) bool {
 	}
 }
 
-type multipartUpload struct {
-	reader      *io.PipeReader
-	writer      *io.PipeWriter
-	contentType string
-	done        chan struct{}
-
-	mu      sync.Mutex
-	audio   io.Closer
-	err     error
-	aborted bool
-}
-
-func newMultipartUpload(ctx context.Context, audioPath, language string, openAudio func(string) (io.ReadCloser, error)) *multipartUpload {
-	reader, writer := io.Pipe()
-	multipartWriter := multipart.NewWriter(writer)
-	upload := &multipartUpload{
-		reader:      reader,
-		writer:      writer,
-		contentType: multipartWriter.FormDataContentType(),
-		done:        make(chan struct{}),
-	}
-
-	go func() {
-		err := upload.write(ctx, multipartWriter, audioPath, language, openAudio)
-		if err != nil {
-			_ = writer.CloseWithError(err)
-		} else {
-			_ = writer.Close()
-		}
-		upload.mu.Lock()
-		upload.err = err
-		upload.audio = nil
-		upload.mu.Unlock()
-		close(upload.done)
-	}()
-
-	go func() {
-		select {
-		case <-ctx.Done():
-			upload.abort()
-		case <-upload.done:
-		}
-	}()
-
-	return upload
-}
-
-func (u *multipartUpload) Read(p []byte) (int, error) {
-	return u.reader.Read(p)
-}
-
-func (u *multipartUpload) Close() error {
-	u.abort()
-	return nil
-}
-
-func (u *multipartUpload) Wait() error {
-	<-u.done
-	u.mu.Lock()
-	defer u.mu.Unlock()
-	return u.err
-}
-
-func (u *multipartUpload) write(ctx context.Context, writer *multipart.Writer, audioPath, language string, openAudio func(string) (io.ReadCloser, error)) error {
-	fileWriter, err := writer.CreateFormFile("file", filepath.Base(audioPath))
-	if err != nil {
-		return u.producerError(ctx, fmt.Errorf("create multipart file field: %w", err))
-	}
-
-	audioFile, err := openAudio(audioPath)
-	if err != nil {
-		return u.producerError(ctx, fmt.Errorf("open audio file %q: %w", audioPath, err))
-	}
-	u.setAudio(audioFile)
-
-	_, copyErr := io.CopyBuffer(fileWriter, &contextReader{ctx: ctx, reader: audioFile}, make([]byte, whisperXUploadBufferSize))
-	closeErr := audioFile.Close()
-	u.clearAudio(audioFile)
-	if copyErr != nil {
-		return u.producerError(ctx, fmt.Errorf("copy audio file %q: %w", audioPath, copyErr))
-	}
-	if closeErr != nil {
-		return u.producerError(ctx, fmt.Errorf("close audio file %q: %w", audioPath, closeErr))
-	}
-
-	if err := writer.WriteField("language", language); err != nil {
-		return u.producerError(ctx, fmt.Errorf("write language form field: %w", err))
-	}
-	if err := writer.Close(); err != nil {
-		return u.producerError(ctx, fmt.Errorf("close multipart writer: %w", err))
-	}
-	return nil
-}
-
-func (u *multipartUpload) producerError(ctx context.Context, err error) error {
-	if ctx.Err() != nil {
-		return ctx.Err()
-	}
-	u.mu.Lock()
-	aborted := u.aborted
-	u.mu.Unlock()
-	if aborted {
-		return nil
-	}
-	return err
-}
-
-func (u *multipartUpload) setAudio(audio io.Closer) {
-	u.mu.Lock()
-	u.audio = audio
-	aborted := u.aborted
-	u.mu.Unlock()
-	if aborted {
-		_ = audio.Close()
-	}
-}
-
-func (u *multipartUpload) clearAudio(audio io.Closer) {
-	u.mu.Lock()
-	if u.audio == audio {
-		u.audio = nil
-	}
-	u.mu.Unlock()
-}
-
-func (u *multipartUpload) abort() {
-	u.mu.Lock()
-	if u.aborted {
-		u.mu.Unlock()
-		return
-	}
-	u.aborted = true
-	audio := u.audio
-	u.mu.Unlock()
-
-	_ = u.reader.Close()
-	if audio != nil {
-		_ = audio.Close()
-	}
-}
-
-type contextReader struct {
-	ctx    context.Context
-	reader io.Reader
-}
-
-func (r *contextReader) Read(p []byte) (int, error) {
-	select {
-	case <-r.ctx.Done():
-		return 0, r.ctx.Err()
-	default:
-		return r.reader.Read(p)
-	}
-}
-
 func writeFileAtomic(path string, data []byte, perm os.FileMode) error {
 	if strings.TrimSpace(path) == "" {
 		return fmt.Errorf("path is required")

internal/adapters/whisperx/http_replay_test.go

@@ -0,0 +1,261 @@
+package whisperx
+
+import (
+	"context"
+	"errors"
+	"io"
+	"net/http"
+	"os"
+	"path/filepath"
+	"strings"
+	"sync/atomic"
+	"testing"
+	"time"
+)
+
+func TestHTTPClientTransportReplaysAbandonedMultipartBody(t *testing.T) {
+	var opens atomic.Int32
+	var closes atomic.Int32
+	transport := roundTripFunc(func(request *http.Request) (*http.Response, error) {
+		partial := make([]byte, 17)
+		if _, err := io.ReadFull(request.Body, partial); err != nil {
+			return nil, err
+		}
+		if err := request.Body.Close(); err != nil {
+			return nil, err
+		}
+		if request.GetBody == nil {
+			return nil, errors.New("request GetBody is nil")
+		}
+		replay, err := request.GetBody()
+		if err != nil {
+			return nil, err
+		}
+		replayedBytes, err := io.ReadAll(replay)
+		if err != nil {
+			_ = replay.Close()
+			return nil, err
+		}
+		if err := replay.Close(); err != nil {
+			return nil, err
+		}
+		parts := parseMultipartParts(t, request.Header.Get("Content-Type"), replayedBytes)
+		if len(parts) != 2 || parts[0].name != "file" || parts[0].filename != "audio.flac" || string(parts[0].data) != "audio-data" {
+			t.Fatalf("replayed file part = %#v, want exact audio.flac payload", parts)
+		}
+		if parts[1].name != "language" || string(parts[1].data) != "en" {
+			t.Fatalf("replayed language part = %#v, want language=en", parts[1])
+		}
+		return whisperXJSONResponse(request), nil
+	})
+	client := newReplayTestHTTPClient(t, transport)
+	client.openAudio = func(string) (io.ReadCloser, error) {
+		opens.Add(1)
+		return &countedReadCloser{Reader: strings.NewReader("audio-data"), closes: &closes}, nil
+	}
+	outPath := filepath.Join(t.TempDir(), "raw.json")
+
+	result, err := client.Transcribe(context.Background(), TranscribeRequest{
+		AudioPath:               "audio.flac",
+		OutputRawTranscriptPath: outPath,
+	})
+	if err != nil {
+		t.Fatalf("Transcribe() error = %v", err)
+	}
+	if result.Attempts != 1 {
+		t.Fatalf("application attempts = %d, want 1", result.Attempts)
+	}
+	verifyJSONFile(t, outPath)
+	if got := opens.Load(); got != 2 {
+		t.Fatalf("audio opens = %d, want original and replay", got)
+	}
+	if got := closes.Load(); got != 2 {
+		t.Fatalf("audio closes = %d, want all body handles disposed", got)
+	}
+}
+
+func TestHTTPClientTransportReturnsReplayOpenFailure(t *testing.T) {
+	replayErr := errors.New("replay open failed")
+	var opens atomic.Int32
+	var closes atomic.Int32
+	transport := roundTripFunc(func(request *http.Request) (*http.Response, error) {
+		if _, err := io.ReadFull(request.Body, make([]byte, 1)); err != nil {
+			return nil, err
+		}
+		_ = request.Body.Close()
+		_, err := request.GetBody()
+		return nil, err
+	})
+	client := newReplayTestHTTPClient(t, transport)
+	client.openAudio = func(string) (io.ReadCloser, error) {
+		if opens.Add(1) == 2 {
+			return nil, replayErr
+		}
+		return &countedReadCloser{Reader: strings.NewReader("audio-data"), closes: &closes}, nil
+	}
+	outPath := filepath.Join(t.TempDir(), "raw.json")
+
+	result, err := client.Transcribe(context.Background(), TranscribeRequest{
+		AudioPath:               "audio.flac",
+		OutputRawTranscriptPath: outPath,
+	})
+	if !errors.Is(err, replayErr) {
+		t.Fatalf("Transcribe() error = %v, want replay open failure", err)
+	}
+	if result.Attempts != 1 {
+		t.Fatalf("application attempts = %d, want 1", result.Attempts)
+	}
+	if got := closes.Load(); got != 1 {
+		t.Fatalf("audio closes = %d, want original disposed", got)
+	}
+	if _, statErr := os.Stat(outPath); !errors.Is(statErr, os.ErrNotExist) {
+		t.Fatalf("output should not exist, stat error = %v", statErr)
+	}
+}
+
+func TestHTTPClientTransportCannotSwallowReplayReadFailure(t *testing.T) {
+	replayErr := errors.New("replay read failed")
+	var opens atomic.Int32
+	transport := roundTripFunc(func(request *http.Request) (*http.Response, error) {
+		if _, err := io.ReadFull(request.Body, make([]byte, 1)); err != nil {
+			return nil, err
+		}
+		_ = request.Body.Close()
+		replay, err := request.GetBody()
+		if err != nil {
+			return nil, err
+		}
+		_, _ = io.ReadAll(replay)
+		_ = replay.Close()
+		return whisperXJSONResponse(request), nil
+	})
+	client := newReplayTestHTTPClient(t, transport)
+	client.openAudio = func(string) (io.ReadCloser, error) {
+		if opens.Add(1) == 2 {
+			return &failingReadCloser{first: []byte("partial"), err: replayErr}, nil
+		}
+		return io.NopCloser(strings.NewReader("audio-data")), nil
+	}
+	outPath := filepath.Join(t.TempDir(), "raw.json")
+
+	_, err := client.Transcribe(context.Background(), TranscribeRequest{
+		AudioPath:               "audio.flac",
+		OutputRawTranscriptPath: outPath,
+	})
+	if !errors.Is(err, replayErr) {
+		t.Fatalf("Transcribe() error = %v, want replay read failure", err)
+	}
+	if _, statErr := os.Stat(outPath); !errors.Is(statErr, os.ErrNotExist) {
+		t.Fatalf("output should not exist, stat error = %v", statErr)
+	}
+}
+
+func TestHTTPClientTransportEarlyResponseDisposesOriginalBody(t *testing.T) {
+	var closes atomic.Int32
+	client := newReplayTestHTTPClient(t, roundTripFunc(func(request *http.Request) (*http.Response, error) {
+		return whisperXJSONResponse(request), nil
+	}))
+	client.openAudio = func(string) (io.ReadCloser, error) {
+		return &countedReadCloser{Reader: strings.NewReader("audio-data"), closes: &closes}, nil
+	}
+
+	_, err := client.Transcribe(context.Background(), TranscribeRequest{
+		AudioPath:               "audio.flac",
+		OutputRawTranscriptPath: filepath.Join(t.TempDir(), "raw.json"),
+	})
+	if err != nil {
+		t.Fatalf("Transcribe() error = %v", err)
+	}
+	if got := closes.Load(); got != 1 {
+		t.Fatalf("audio closes = %d, want early-response body disposed", got)
+	}
+}
+
+func TestHTTPClientParentCancellationDuringReplayClosesEveryBody(t *testing.T) {
+	var opens atomic.Int32
+	var originalCloses atomic.Int32
+	replaySource := newSignalingBlockedReadCloser()
+	transport := roundTripFunc(func(request *http.Request) (*http.Response, error) {
+		if _, err := io.ReadFull(request.Body, make([]byte, 1)); err != nil {
+			return nil, err
+		}
+		_ = request.Body.Close()
+		replay, err := request.GetBody()
+		if err != nil {
+			return nil, err
+		}
+		_, _ = io.ReadAll(replay)
+		_ = replay.Close()
+		return nil, request.Context().Err()
+	})
+	client := newReplayTestHTTPClient(t, transport)
+	client.openAudio = func(string) (io.ReadCloser, error) {
+		if opens.Add(1) == 1 {
+			return &countedReadCloser{Reader: strings.NewReader("audio-data"), closes: &originalCloses}, nil
+		}
+		return replaySource, nil
+	}
+	ctx, cancel := context.WithCancel(context.Background())
+	done := make(chan error, 1)
+	go func() {
+		_, err := client.Transcribe(ctx, TranscribeRequest{
+			AudioPath:               "audio.flac",
+			OutputRawTranscriptPath: filepath.Join(t.TempDir(), "raw.json"),
+		})
+		done <- err
+	}()
+
+	select {
+	case <-replaySource.blocked:
+		cancel()
+	case <-time.After(time.Second):
+		cancel()
+		t.Fatal("replay source did not block")
+	}
+	select {
+	case err := <-done:
+		if !errors.Is(err, context.Canceled) {
+			t.Fatalf("Transcribe() error = %v, want context cancellation", err)
+		}
+	case <-time.After(time.Second):
+		t.Fatal("Transcribe() did not return after replay cancellation")
+	}
+	if got := originalCloses.Load(); got != 1 {
+		t.Fatalf("original audio closes = %d, want 1", got)
+	}
+	select {
+	case <-replaySource.closed:
+	case <-time.After(time.Second):
+		t.Fatal("replay audio source was not closed")
+	}
+}
+
+type roundTripFunc func(*http.Request) (*http.Response, error)
+
+func (roundTrip roundTripFunc) RoundTrip(request *http.Request) (*http.Response, error) {
+	return roundTrip(request)
+}
+
+func newReplayTestHTTPClient(t *testing.T, transport http.RoundTripper) *HTTPClient {
+	t.Helper()
+	client, err := NewHTTPClient(HTTPClientConfig{
+		TranscribeURL: "http://whisperx.test/transcribe",
+		Language:      "en",
+		Timeout:       time.Second,
+		HTTPClient:    &http.Client{Transport: transport},
+	})
+	if err != nil {
+		t.Fatalf("NewHTTPClient() error = %v", err)
+	}
+	return client
+}
+
+func whisperXJSONResponse(request *http.Request) *http.Response {
+	return &http.Response{
+		StatusCode: http.StatusOK,
+		Status:     "200 OK",
+		Header:     make(http.Header),
+		Body:       io.NopCloser(strings.NewReader(`{"source":"transport-replay","segments":[]}`)),
+		Request:    request,
+	}
+}

internal/adapters/whisperx/http_test.go

@@ -365,7 +365,7 @@ func TestHTTPClientEarlyServerResponseReturns(t *testing.T) {
 	}
 }
 
-func TestHTTPClientCancellationReleasesBlockedProducer(t *testing.T) {
+func TestHTTPClientCancellationReleasesBlockedSource(t *testing.T) {
 	release := make(chan struct{})
 	source := newGatedReadCloser([]byte("audio-data"), release)
 	firstByteReceived := make(chan struct{})

internal/adapters/whisperx/multipart_body.go

@@ -0,0 +1,263 @@
+package whisperx
+
+import (
+	"bytes"
+	"context"
+	"errors"
+	"fmt"
+	"io"
+	"mime/multipart"
+	"path/filepath"
+	"sync"
+)
+
+type multipartBodyFactory struct {
+	ctx         context.Context
+	audioPath   string
+	openAudio   func(string) (io.ReadCloser, error)
+	prefix      []byte
+	suffix      []byte
+	contentType string
+
+	mu     sync.Mutex
+	closed bool
+	bodies []*multipartBody
+}
+
+func newMultipartBodyFactory(
+	ctx context.Context,
+	audioPath string,
+	language string,
+	openAudio func(string) (io.ReadCloser, error),
+) (*multipartBodyFactory, error) {
+	if ctx == nil {
+		ctx = context.Background()
+	}
+	if openAudio == nil {
+		return nil, fmt.Errorf("audio opener is required")
+	}
+	var framing bytes.Buffer
+	writer := multipart.NewWriter(&framing)
+	if _, err := writer.CreateFormFile("file", filepath.Base(audioPath)); err != nil {
+		return nil, fmt.Errorf("create multipart file field: %w", err)
+	}
+	prefixLength := framing.Len()
+	if err := writer.WriteField("language", language); err != nil {
+		return nil, fmt.Errorf("write language form field: %w", err)
+	}
+	if err := writer.Close(); err != nil {
+		return nil, fmt.Errorf("close multipart framing: %w", err)
+	}
+	framingBytes := framing.Bytes()
+	return &multipartBodyFactory{
+		ctx:         ctx,
+		audioPath:   audioPath,
+		openAudio:   openAudio,
+		prefix:      append([]byte(nil), framingBytes[:prefixLength]...),
+		suffix:      append([]byte(nil), framingBytes[prefixLength:]...),
+		contentType: writer.FormDataContentType(),
+	}, nil
+}
+
+func (factory *multipartBodyFactory) NewBody() (io.ReadCloser, error) {
+	factory.mu.Lock()
+	if factory.closed {
+		factory.mu.Unlock()
+		return nil, fmt.Errorf("multipart body factory is closed")
+	}
+	if err := factory.ctx.Err(); err != nil {
+		factory.mu.Unlock()
+		return nil, err
+	}
+	audio, err := factory.openAudio(factory.audioPath)
+	if err != nil {
+		factory.mu.Unlock()
+		return nil, fmt.Errorf("open audio file %q: %w", factory.audioPath, err)
+	}
+	if err := factory.ctx.Err(); err != nil {
+		factory.mu.Unlock()
+		_ = audio.Close()
+		return nil, err
+	}
+	source := &multipartAudioSource{audio: audio, closeDone: make(chan struct{})}
+	body := &multipartBody{
+		ctx:    factory.ctx,
+		source: source,
+		reader: io.MultiReader(bytes.NewReader(factory.prefix), source, bytes.NewReader(factory.suffix)),
+	}
+	factory.bodies = append(factory.bodies, body)
+	factory.mu.Unlock()
+	stopCancel := context.AfterFunc(factory.ctx, func() {
+		_ = body.Close()
+	})
+	body.mu.Lock()
+	body.stopCancel = stopCancel
+	closed := body.closed
+	body.mu.Unlock()
+	if closed {
+		stopCancel()
+	}
+	if err := factory.ctx.Err(); err != nil {
+		_ = body.Close()
+		return nil, err
+	}
+	return body, nil
+}
+
+func (factory *multipartBodyFactory) Close() error {
+	factory.mu.Lock()
+	if factory.closed {
+		bodies := append([]*multipartBody(nil), factory.bodies...)
+		factory.mu.Unlock()
+		return multipartBodyCloseErrors(bodies)
+	}
+	factory.closed = true
+	bodies := append([]*multipartBody(nil), factory.bodies...)
+	factory.mu.Unlock()
+	for _, body := range bodies {
+		_ = body.Close()
+	}
+	return multipartBodyCloseErrors(bodies)
+}
+
+func (factory *multipartBodyFactory) SourceError(includeAbandoned bool) error {
+	factory.mu.Lock()
+	bodies := append([]*multipartBody(nil), factory.bodies...)
+	factory.mu.Unlock()
+	var result error
+	for _, body := range bodies {
+		if err := body.sourceError(includeAbandoned); err != nil {
+			result = errors.Join(result, err)
+		}
+	}
+	return result
+}
+
+func multipartBodyCloseErrors(bodies []*multipartBody) error {
+	var result error
+	for _, body := range bodies {
+		if err := body.sourceError(false); err != nil {
+			result = errors.Join(result, err)
+		}
+	}
+	return result
+}
+
+type multipartBody struct {
+	ctx        context.Context
+	reader     io.Reader
+	source     *multipartAudioSource
+	stopCancel func() bool
+
+	mu        sync.Mutex
+	closed    bool
+	abandoned bool
+}
+
+func (body *multipartBody) Read(buffer []byte) (int, error) {
+	if err := body.ctx.Err(); err != nil {
+		return 0, err
+	}
+	body.mu.Lock()
+	closed := body.closed
+	body.mu.Unlock()
+	if closed {
+		return 0, io.ErrClosedPipe
+	}
+	return body.reader.Read(buffer)
+}
+
+func (body *multipartBody) Close() error {
+	body.mu.Lock()
+	if body.closed {
+		body.mu.Unlock()
+		return body.source.Close()
+	}
+	body.closed = true
+	body.abandoned = !body.source.reachedEOF()
+	stopCancel := body.stopCancel
+	body.mu.Unlock()
+	if stopCancel != nil {
+		stopCancel()
+	}
+	return body.source.Close()
+}
+
+func (body *multipartBody) sourceError(includeAbandoned bool) error {
+	body.mu.Lock()
+	abandoned := body.abandoned
+	body.mu.Unlock()
+	err := body.source.Err()
+	if err != nil {
+		return err
+	}
+	if abandoned && !includeAbandoned {
+		return nil
+	}
+	return nil
+}
+
+type multipartAudioSource struct {
+	audio io.ReadCloser
+
+	mu          sync.Mutex
+	closing     bool
+	eof         bool
+	readErr     error
+	closeErr    error
+	closeDone   chan struct{}
+	closeResult error
+	closed      bool
+}
+
+func (source *multipartAudioSource) Read(buffer []byte) (int, error) {
+	n, err := source.audio.Read(buffer)
+	source.mu.Lock()
+	switch {
+	case errors.Is(err, io.EOF):
+		source.eof = true
+	case err != nil && !source.closing:
+		source.readErr = errors.Join(source.readErr, err)
+	}
+	source.mu.Unlock()
+	return n, err
+}
+
+func (source *multipartAudioSource) Close() error {
+	source.mu.Lock()
+	if source.closed {
+		done := source.closeDone
+		source.mu.Unlock()
+		<-done
+		source.mu.Lock()
+		err := source.closeResult
+		source.mu.Unlock()
+		return err
+	}
+	source.closed = true
+	source.closing = true
+	reachedEOF := source.eof
+	source.mu.Unlock()
+
+	err := source.audio.Close()
+	source.mu.Lock()
+	source.closeResult = err
+	if err != nil && reachedEOF {
+		source.closeErr = err
+	}
+	close(source.closeDone)
+	source.mu.Unlock()
+	return err
+}
+
+func (source *multipartAudioSource) reachedEOF() bool {
+	source.mu.Lock()
+	defer source.mu.Unlock()
+	return source.eof
+}
+
+func (source *multipartAudioSource) Err() error {
+	source.mu.Lock()
+	defer source.mu.Unlock()
+	return errors.Join(source.readErr, source.closeErr)
+}

internal/adapters/whisperx/multipart_body_test.go

@@ -0,0 +1,327 @@
+package whisperx
+
+import (
+	"bytes"
+	"context"
+	"errors"
+	"io"
+	"mime"
+	"mime/multipart"
+	"strings"
+	"sync"
+	"sync/atomic"
+	"testing"
+	"time"
+)
+
+func TestMultipartBodyReplayUsesIdenticalFraming(t *testing.T) {
+	factory := newTestMultipartBodyFactory(t, context.Background(), "audio-data")
+	defer factory.Close()
+
+	original, err := factory.NewBody()
+	if err != nil {
+		t.Fatalf("NewBody() original error = %v", err)
+	}
+	replay, err := factory.NewBody()
+	if err != nil {
+		t.Fatalf("NewBody() replay error = %v", err)
+	}
+	originalBytes, err := io.ReadAll(original)
+	if err != nil {
+		t.Fatalf("ReadAll(original) error = %v", err)
+	}
+	replayBytes, err := io.ReadAll(replay)
+	if err != nil {
+		t.Fatalf("ReadAll(replay) error = %v", err)
+	}
+	if !bytes.Equal(originalBytes, replayBytes) {
+		t.Fatal("original and replay multipart bytes differ")
+	}
+
+	parts := parseMultipartParts(t, factory.contentType, replayBytes)
+	if len(parts) != 2 {
+		t.Fatalf("part count = %d, want 2", len(parts))
+	}
+	if parts[0].name != "file" || parts[0].filename != "audio.flac" || string(parts[0].data) != "audio-data" {
+		t.Fatalf("file part = %#v, want file/audio.flac with exact audio", parts[0])
+	}
+	if parts[1].name != "language" || string(parts[1].data) != "en" {
+		t.Fatalf("language part = %#v, want language=en", parts[1])
+	}
+}
+
+func TestMultipartBodyReplayRestartsAfterPartialOriginal(t *testing.T) {
+	factory := newTestMultipartBodyFactory(t, context.Background(), "audio-data")
+	defer factory.Close()
+
+	original, err := factory.NewBody()
+	if err != nil {
+		t.Fatalf("NewBody() original error = %v", err)
+	}
+	partial := make([]byte, len(factory.prefix)+3)
+	if _, err := io.ReadFull(original, partial); err != nil {
+		t.Fatalf("ReadFull(original) error = %v", err)
+	}
+	if err := original.Close(); err != nil {
+		t.Fatalf("Close(original) error = %v", err)
+	}
+
+	replay, err := factory.NewBody()
+	if err != nil {
+		t.Fatalf("NewBody() replay error = %v", err)
+	}
+	replayBytes, err := io.ReadAll(replay)
+	if err != nil {
+		t.Fatalf("ReadAll(replay) error = %v", err)
+	}
+	parts := parseMultipartParts(t, factory.contentType, replayBytes)
+	if got := string(parts[0].data); got != "audio-data" {
+		t.Fatalf("replayed audio = %q, want complete payload", got)
+	}
+	if err := factory.SourceError(false); err != nil {
+		t.Fatalf("SourceError(false) = %v, want abandoned original ignored", err)
+	}
+}
+
+func TestMultipartBodySimultaneousBodiesUseIndependentSources(t *testing.T) {
+	var opens atomic.Int32
+	var closes atomic.Int32
+	factory, err := newMultipartBodyFactory(context.Background(), "audio.flac", "en", func(string) (io.ReadCloser, error) {
+		opens.Add(1)
+		return &countedReadCloser{Reader: strings.NewReader("audio-data"), closes: &closes}, nil
+	})
+	if err != nil {
+		t.Fatalf("newMultipartBodyFactory() error = %v", err)
+	}
+
+	first, err := factory.NewBody()
+	if err != nil {
+		t.Fatalf("NewBody() first error = %v", err)
+	}
+	second, err := factory.NewBody()
+	if err != nil {
+		t.Fatalf("NewBody() second error = %v", err)
+	}
+	var firstBytes, secondBytes []byte
+	var firstErr, secondErr error
+	var group sync.WaitGroup
+	group.Add(2)
+	go func() {
+		defer group.Done()
+		firstBytes, firstErr = io.ReadAll(first)
+	}()
+	go func() {
+		defer group.Done()
+		secondBytes, secondErr = io.ReadAll(second)
+	}()
+	group.Wait()
+	if firstErr != nil || secondErr != nil {
+		t.Fatalf("concurrent reads returned (%v, %v)", firstErr, secondErr)
+	}
+	if !bytes.Equal(firstBytes, secondBytes) {
+		t.Fatal("independent body bytes differ")
+	}
+	if got := opens.Load(); got != 2 {
+		t.Fatalf("audio opens = %d, want 2", got)
+	}
+	if err := factory.Close(); err != nil {
+		t.Fatalf("Close() error = %v", err)
+	}
+	if err := factory.Close(); err != nil {
+		t.Fatalf("second Close() error = %v", err)
+	}
+	if got := closes.Load(); got != 2 {
+		t.Fatalf("audio closes = %d, want 2", got)
+	}
+}
+
+func TestMultipartBodyReportsSourceFailures(t *testing.T) {
+	t.Run("read", func(t *testing.T) {
+		sourceErr := errors.New("read source")
+		factory, err := newMultipartBodyFactory(context.Background(), "audio.flac", "en", func(string) (io.ReadCloser, error) {
+			return &failingReadCloser{first: []byte("partial"), err: sourceErr}, nil
+		})
+		if err != nil {
+			t.Fatalf("newMultipartBodyFactory() error = %v", err)
+		}
+		body, err := factory.NewBody()
+		if err != nil {
+			t.Fatalf("NewBody() error = %v", err)
+		}
+		if _, err := io.ReadAll(body); !errors.Is(err, sourceErr) {
+			t.Fatalf("ReadAll() error = %v, want source failure", err)
+		}
+		_ = body.Close()
+		if err := factory.SourceError(false); !errors.Is(err, sourceErr) {
+			t.Fatalf("SourceError(false) = %v, want source failure", err)
+		}
+	})
+
+	t.Run("close after complete read", func(t *testing.T) {
+		closeErr := errors.New("close source")
+		factory, err := newMultipartBodyFactory(context.Background(), "audio.flac", "en", func(string) (io.ReadCloser, error) {
+			return &errorCloseReadCloser{Reader: strings.NewReader("audio-data"), err: closeErr}, nil
+		})
+		if err != nil {
+			t.Fatalf("newMultipartBodyFactory() error = %v", err)
+		}
+		body, err := factory.NewBody()
+		if err != nil {
+			t.Fatalf("NewBody() error = %v", err)
+		}
+		if _, err := io.ReadAll(body); err != nil {
+			t.Fatalf("ReadAll() error = %v", err)
+		}
+		if err := body.Close(); !errors.Is(err, closeErr) {
+			t.Fatalf("Close() error = %v, want source close failure", err)
+		}
+		if err := factory.SourceError(false); !errors.Is(err, closeErr) {
+			t.Fatalf("SourceError(false) = %v, want source close failure", err)
+		}
+	})
+}
+
+func TestMultipartBodyCloseAndCancelReleaseBlockedSource(t *testing.T) {
+	for _, test := range []struct {
+		name    string
+		release func(context.CancelFunc, io.ReadCloser)
+	}{
+		{name: "close", release: func(_ context.CancelFunc, body io.ReadCloser) { _ = body.Close() }},
+		{name: "cancel", release: func(cancel context.CancelFunc, _ io.ReadCloser) { cancel() }},
+	} {
+		t.Run(test.name, func(t *testing.T) {
+			ctx, cancel := context.WithCancel(context.Background())
+			defer cancel()
+			source := newSignalingBlockedReadCloser()
+			factory, err := newMultipartBodyFactory(ctx, "audio.flac", "en", func(string) (io.ReadCloser, error) {
+				return source, nil
+			})
+			if err != nil {
+				t.Fatalf("newMultipartBodyFactory() error = %v", err)
+			}
+			defer factory.Close()
+			body, err := factory.NewBody()
+			if err != nil {
+				t.Fatalf("NewBody() error = %v", err)
+			}
+			readDone := make(chan error, 1)
+			go func() {
+				_, err := io.ReadAll(body)
+				readDone <- err
+			}()
+			select {
+			case <-source.blocked:
+			case <-time.After(time.Second):
+				t.Fatal("audio source did not block")
+			}
+			test.release(cancel, body)
+			select {
+			case <-source.closed:
+			case <-time.After(time.Second):
+				t.Fatal("audio source was not closed")
+			}
+			select {
+			case <-readDone:
+			case <-time.After(time.Second):
+				t.Fatal("blocked body read did not return")
+			}
+			if test.name == "cancel" {
+				if _, err := factory.NewBody(); !errors.Is(err, context.Canceled) {
+					t.Fatalf("NewBody() after cancellation error = %v, want context.Canceled", err)
+				}
+			}
+		})
+	}
+}
+
+func newTestMultipartBodyFactory(t *testing.T, ctx context.Context, audio string) *multipartBodyFactory {
+	t.Helper()
+	factory, err := newMultipartBodyFactory(ctx, "audio.flac", "en", func(string) (io.ReadCloser, error) {
+		return io.NopCloser(strings.NewReader(audio)), nil
+	})
+	if err != nil {
+		t.Fatalf("newMultipartBodyFactory() error = %v", err)
+	}
+	return factory
+}
+
+type multipartTestPart struct {
+	name     string
+	filename string
+	data     []byte
+}
+
+func parseMultipartParts(t *testing.T, contentType string, body []byte) []multipartTestPart {
+	t.Helper()
+	mediaType, params, err := mime.ParseMediaType(contentType)
+	if err != nil {
+		t.Fatalf("ParseMediaType() error = %v", err)
+	}
+	if mediaType != "multipart/form-data" {
+		t.Fatalf("media type = %q, want multipart/form-data", mediaType)
+	}
+	reader := multipart.NewReader(bytes.NewReader(body), params["boundary"])
+	var parts []multipartTestPart
+	for {
+		part, err := reader.NextPart()
+		if errors.Is(err, io.EOF) {
+			return parts
+		}
+		if err != nil {
+			t.Fatalf("NextPart() error = %v", err)
+		}
+		data, err := io.ReadAll(part)
+		if err != nil {
+			t.Fatalf("ReadAll(part %q) error = %v", part.FormName(), err)
+		}
+		parts = append(parts, multipartTestPart{name: part.FormName(), filename: part.FileName(), data: data})
+	}
+}
+
+type countedReadCloser struct {
+	io.Reader
+	closes *atomic.Int32
+	once   sync.Once
+}
+
+func (reader *countedReadCloser) Close() error {
+	reader.once.Do(func() { reader.closes.Add(1) })
+	return nil
+}
+
+type errorCloseReadCloser struct {
+	io.Reader
+	err error
+}
+
+func (reader *errorCloseReadCloser) Close() error { return reader.err }
+
+type signalingBlockedReadCloser struct {
+	blocked chan struct{}
+	closed  chan struct{}
+	once    sync.Once
+	read    bool
+}
+
+func newSignalingBlockedReadCloser() *signalingBlockedReadCloser {
+	return &signalingBlockedReadCloser{blocked: make(chan struct{}), closed: make(chan struct{})}
+}
+
+func (reader *signalingBlockedReadCloser) Read(buffer []byte) (int, error) {
+	if !reader.read {
+		reader.read = true
+		return copy(buffer, "audio-data"), nil
+	}
+	reader.once.Do(func() { close(reader.blocked) })
+	<-reader.closed
+	return 0, errors.New("audio source closed")
+}
+
+func (reader *signalingBlockedReadCloser) Close() error {
+	select {
+	case <-reader.closed:
+	default:
+		close(reader.closed)
+	}
+	return nil
+}

internal/app/analyze_artifacts_commands_test.go

@@ -38,11 +38,7 @@ func TestExecuteRunStagePublishPropagatesSelectedArtifacts(t *testing.T) {
 
 	var capturedStages []string
 	var capturedArtifacts []string
-	origExecuteStagesFn := executeStagesFn
+	executor := func(_ context.Context, _ *config.Config, plan BoundedPlan, opts RunOptions) (*RunSummary, error) {
-	t.Cleanup(func() {
-		executeStagesFn = origExecuteStagesFn
-	})
-	executeStagesFn = func(_ context.Context, _ *config.Config, plan BoundedPlan, opts RunOptions) (*RunSummary, error) {
 		for _, s := range plan.Stages() {
 			capturedStages = append(capturedStages, s.Name())
 		}
@@ -52,7 +48,7 @@ func TestExecuteRunStagePublishPropagatesSelectedArtifacts(t *testing.T) {
 
 	var stdout bytes.Buffer
 	var stderr bytes.Buffer
-	code := Execute(
+	code := executeWithPipelineExecutor(
 		[]string{
 			"run-stage", "publish", "2026-05-03",
 			"--config", pipelinePath,
@@ -62,6 +58,7 @@ func TestExecuteRunStagePublishPropagatesSelectedArtifacts(t *testing.T) {
 		},
 		&stdout,
 		&stderr,
+		executor,
 	)
 	if code != 0 {
 		t.Fatalf("exit code = %d, want 0; stderr=%q", code, stderr.String())
@@ -154,7 +151,7 @@ func TestRunArtifactsWithSucceededAnalyzeSkipsUnlessForced(t *testing.T) {
 	if err != nil {
 		t.Fatalf("Run() error = %v", err)
 	}
-	if !strings.Contains(out.String(), "executed=4 skipped=8") {
+	if !strings.Contains(out.String(), "executed=5 skipped=9") {
 		t.Fatalf("output = %q, want extract reconsidered and legacy analyze plus delivery rebuilt", out.String())
 	}
 }
@@ -165,11 +162,7 @@ func TestExecuteAnalyzeForceRunsAnalyze(t *testing.T) {
 
 	var capturedStages []string
 	var capturedForce bool
-	origExecuteStagesFn := executeStagesFn
+	executor := func(_ context.Context, _ *config.Config, plan BoundedPlan, opts RunOptions) (*RunSummary, error) {
-	t.Cleanup(func() {
-		executeStagesFn = origExecuteStagesFn
-	})
-	executeStagesFn = func(_ context.Context, _ *config.Config, plan BoundedPlan, opts RunOptions) (*RunSummary, error) {
 		for _, s := range plan.Stages() {
 			capturedStages = append(capturedStages, s.Name())
 		}
@@ -182,10 +175,11 @@ func TestExecuteAnalyzeForceRunsAnalyze(t *testing.T) {
 
 	var stdout bytes.Buffer
 	var stderr bytes.Buffer
-	code := Execute(
+	code := executeWithPipelineExecutor(
 		[]string{"analyze", "2026-05-03", "--config", pipelinePath, "--campaign-file", campaignPath, "--session", sessionPath},
 		&stdout,
 		&stderr,
+		executor,
 	)
 	if code != 0 {
 		t.Fatalf("exit code = %d, want 0; stderr=%q", code, stderr.String())
@@ -206,18 +200,14 @@ func TestExecuteAnalyzePropagatesSelectedArtifacts(t *testing.T) {
 	pipelinePath, campaignPath, sessionPath := writeValidConfigFilesWithScriptoriumArtifacts(t, workspaceRoot)
 
 	var capturedArtifacts []string
-	origExecuteStagesFn := executeStagesFn
+	executor := func(_ context.Context, _ *config.Config, _ BoundedPlan, opts RunOptions) (*RunSummary, error) {
-	t.Cleanup(func() {
-		executeStagesFn = origExecuteStagesFn
-	})
-	executeStagesFn = func(_ context.Context, _ *config.Config, _ BoundedPlan, opts RunOptions) (*RunSummary, error) {
 		capturedArtifacts = append([]string(nil), opts.SelectedArtifacts...)
 		return &RunSummary{ManifestPath: filepath.Join(workspaceRoot, "manifest.json"), Executed: []string{"analyze"}}, nil
 	}
 
 	var stdout bytes.Buffer
 	var stderr bytes.Buffer
-	code := Execute(
+	code := executeWithPipelineExecutor(
 		[]string{
 			"analyze",
 			"2026-05-03",
@@ -228,6 +218,7 @@ func TestExecuteAnalyzePropagatesSelectedArtifacts(t *testing.T) {
 		},
 		&stdout,
 		&stderr,
+		executor,
 	)
 	if code != 0 {
 		t.Fatalf("exit code = %d, want 0; stderr=%q", code, stderr.String())
@@ -281,13 +272,18 @@ func TestExecuteAnalyzeRejectsPositionalArgsAndForceFlag(t *testing.T) {
 }
 
 func TestExecuteAnalyzeMissingConfigUsesRunStageLoadingPath(t *testing.T) {
+	deps, missingPath, executor := guardedMissingPipelineDiscovery(t)
 	var stdout bytes.Buffer
 	var stderr bytes.Buffer
-	code := Execute([]string{"analyze", "2026-05-03"}, &stdout, &stderr)
+	code := executeWithDependencies([]string{"analyze", "2026-05-03"}, &stdout, &stderr, deps, executor)
 	if code == 0 {
 		t.Fatal("exit code = 0, want non-zero")
 	}
-	if !strings.Contains(stderr.String(), "analyze: no pipeline config path provided and no default pipeline config found; searched:") {
+	if stdout.Len() != 0 {
+		t.Fatalf("stdout = %q, want empty", stdout.String())
+	}
+	if got := stderr.String(); !strings.Contains(got, "analyze: no pipeline config path provided") ||
+		!strings.Contains(got, "no default pipeline config found") || !strings.Contains(got, missingPath) {
 		t.Fatalf("stderr = %q, want pipeline discovery error", stderr.String())
 	}
 }
@@ -299,11 +295,7 @@ func TestExecutePublishForceRunsPublish(t *testing.T) {
 	var capturedStages []string
 	var capturedForce bool
 	var capturedArtifacts []string
-	origExecuteStagesFn := executeStagesFn
+	executor := func(_ context.Context, _ *config.Config, plan BoundedPlan, opts RunOptions) (*RunSummary, error) {
-	t.Cleanup(func() {
-		executeStagesFn = origExecuteStagesFn
-	})
-	executeStagesFn = func(_ context.Context, _ *config.Config, plan BoundedPlan, opts RunOptions) (*RunSummary, error) {
 		for _, s := range plan.Stages() {
 			capturedStages = append(capturedStages, s.Name())
 		}
@@ -317,10 +309,11 @@ func TestExecutePublishForceRunsPublish(t *testing.T) {
 
 	var stdout bytes.Buffer
 	var stderr bytes.Buffer
-	code := Execute(
+	code := executeWithPipelineExecutor(
 		[]string{"publish", "2026-05-03", "--config", pipelinePath, "--campaign-file", campaignPath, "--session", sessionPath, "--artifacts", "session_recap"},
 		&stdout,
 		&stderr,
+		executor,
 	)
 	if code != 0 {
 		t.Fatalf("exit code = %d, want 0; stderr=%q", code, stderr.String())
@@ -383,13 +376,18 @@ func TestExecutePublishUnknownArtifactFailsValidation(t *testing.T) {
 }
 
 func TestExecutePublishMissingConfigUsesRunStageLoadingPath(t *testing.T) {
+	deps, missingPath, executor := guardedMissingPipelineDiscovery(t)
 	var stdout bytes.Buffer
 	var stderr bytes.Buffer
-	code := Execute([]string{"publish", "2026-05-03"}, &stdout, &stderr)
+	code := executeWithDependencies([]string{"publish", "2026-05-03"}, &stdout, &stderr, deps, executor)
 	if code == 0 {
 		t.Fatal("exit code = 0, want non-zero")
 	}
-	if !strings.Contains(stderr.String(), "publish: no pipeline config path provided and no default pipeline config found; searched:") {
+	if stdout.Len() != 0 {
+		t.Fatalf("stdout = %q, want empty", stdout.String())
+	}
+	if got := stderr.String(); !strings.Contains(got, "publish: no pipeline config path provided") ||
+		!strings.Contains(got, "no default pipeline config found") || !strings.Contains(got, missingPath) {
 		t.Fatalf("stderr = %q, want pipeline discovery error", stderr.String())
 	}
 }

internal/app/analyze_checkpoint_recovery_test.go

@@ -0,0 +1,336 @@
+package app
+
+import (
+	"context"
+	"errors"
+	"os"
+	"path/filepath"
+	"slices"
+	"testing"
+	"time"
+
+	"github.com/google/go-cmp/cmp"
+
+	"gitea.maximumdirect.net/eric/narratio/internal/adapters/scriptorium"
+	"gitea.maximumdirect.net/eric/narratio/internal/config"
+	"gitea.maximumdirect.net/eric/narratio/internal/manifest"
+	"gitea.maximumdirect.net/eric/narratio/internal/stage"
+)
+
+func TestExecuteStagesRetriesFromLastAnalyzeCheckpoint(t *testing.T) {
+	cfg := testConfig(t)
+	cfg.Pipeline.Scriptorium = &config.ScriptoriumConfig{Artifacts: checkpointRecoveryArtifacts("original")}
+	store := &manifest.LocalStore{}
+	seed := manifest.New(cfg.Session.SessionID, time.Now().UTC())
+	seed.Campaign = cfg.Session.Campaign
+	for _, name := range []string{"prepare", "transcribe", "merge", "polish", "normalize", "trim"} {
+		seed.MarkStageSucceeded(name, time.Now().UTC(), nil)
+	}
+	saveBoundedManifest(t, cfg, seed)
+	analyze := mustSelectStage(t, "analyze")
+
+	baselineRunner := &analyzeCheckpointRunner{}
+	if _, err := executeStages(context.Background(), cfg, []stage.Stage{analyze}, RunOptions{
+		Env: &Env{ManifestStore: store, Scriptorium: baselineRunner},
+	}); err != nil {
+		t.Fatalf("baseline executeStages() error = %v", err)
+	}
+	baseline, err := store.Load(context.Background(), manifestPathFor(cfg))
+	if err != nil {
+		t.Fatalf("load baseline manifest: %v", err)
+	}
+	baseline.MarkStageSucceeded("publish", time.Now().UTC(), nil)
+	baseline.MarkStageSucceeded("notify", time.Now().UTC(), nil)
+	if err := store.Save(context.Background(), manifestPathFor(cfg), baseline); err != nil {
+		t.Fatalf("save delivery baseline: %v", err)
+	}
+	upstream := cloneStageRecords(baseline, "prepare", "transcribe", "merge", "polish", "normalize", "trim")
+	unrelated := baseline.Stages["analyze"].AnalyzeArtifacts["unrelated"]
+
+	cfg.Pipeline.Scriptorium.Artifacts = checkpointRecoveryArtifacts("revised")
+	failure := errors.New("injected quest-log failure")
+	failingRunner := &analyzeCheckpointRunner{FailPrompt: "quest-log-revised", Err: failure}
+	if _, err := executeStages(context.Background(), cfg, []stage.Stage{analyze}, RunOptions{
+		Env: &Env{ManifestStore: store, Scriptorium: failingRunner},
+	}); !errors.Is(err, failure) {
+		t.Fatalf("executeStages() error = %v, want quest-log failure", err)
+	}
+	if want := []string{"chronicle-revised", "encounters-revised", "quest-log-revised"}; !slices.Equal(failingRunner.Calls, want) {
+		t.Fatalf("failed-run calls = %#v, want %#v", failingRunner.Calls, want)
+	}
+
+	failed := loadAnalyzeRecoveryManifest(t, store, cfg)
+	failedAnalyze := failed.Stages["analyze"]
+	if failedAnalyze.Status != manifest.StatusFailed || len(failedAnalyze.Outputs) != 0 {
+		t.Fatalf("aggregate analyze = %#v, want failed without outputs", failedAnalyze)
+	}
+	assertAnalyzeArtifactStatus(t, failedAnalyze, "chronicle", manifest.AnalyzeArtifactCurrent)
+	assertAnalyzeArtifactStatus(t, failedAnalyze, "encounters", manifest.AnalyzeArtifactCurrent)
+	assertAnalyzeArtifactStatus(t, failedAnalyze, "quest_log", manifest.AnalyzeArtifactFailed)
+	assertAnalyzeArtifactStatus(t, failedAnalyze, "player_handout", manifest.AnalyzeArtifactStale)
+	gotUnrelated := failedAnalyze.AnalyzeArtifacts["unrelated"]
+	if diff := cmp.Diff(unrelated, gotUnrelated); diff != "" {
+		t.Fatalf("unrelated artifact changed (-want +got):\n%s", diff)
+	}
+	for _, name := range []string{"publish", "notify"} {
+		if got := failed.Stages[name]; got == nil || got.Status != manifest.StatusStale {
+			t.Fatalf("%s stage = %#v, want stale", name, got)
+		}
+	}
+	assertStageRecordsUnchanged(t, failed, upstream)
+
+	retryRunner := &analyzeCheckpointRunner{}
+	if _, err := executeStages(context.Background(), cfg, []stage.Stage{analyze}, RunOptions{
+		Env: &Env{ManifestStore: store, Scriptorium: retryRunner},
+	}); err != nil {
+		t.Fatalf("retry executeStages() error = %v", err)
+	}
+	if want := []string{"quest-log-revised", "player-handout-revised"}; !slices.Equal(retryRunner.Calls, want) {
+		t.Fatalf("retry calls = %#v, want %#v", retryRunner.Calls, want)
+	}
+
+	completed := loadAnalyzeRecoveryManifest(t, store, cfg)
+	if completed.Stages["analyze"].Status != manifest.StatusSucceeded {
+		t.Fatalf("aggregate analyze = %#v, want succeeded", completed.Stages["analyze"])
+	}
+	for _, key := range []string{"chronicle", "encounters", "player_handout", "quest_log", "unrelated"} {
+		assertAnalyzeArtifactStatus(t, completed.Stages["analyze"], key, manifest.AnalyzeArtifactCurrent)
+	}
+	assertStageRecordsUnchanged(t, completed, upstream)
+}
+
+func TestExecuteStagesJoinsAnalyzeFailureWithTerminalSaveFailure(t *testing.T) {
+	cfg := testConfig(t)
+	cfg.Pipeline.Scriptorium = &config.ScriptoriumConfig{
+		Artifacts: map[string]config.ScriptoriumArtifactConfig{
+			"chronicle": {Enabled: true, PromptID: "chronicle", OutputPath: "artifacts/chronicle.md"},
+		},
+	}
+	artifactErr := errors.New("injected analyze failure")
+	terminalErr := errors.New("injected terminal save failure")
+	store := &failedAnalyzeSaveStore{Store: &manifest.LocalStore{}, Err: terminalErr}
+	runner := &analyzeCheckpointRunner{FailPrompt: "chronicle", Err: artifactErr}
+
+	_, err := executeStages(context.Background(), cfg, []stage.Stage{mustSelectStage(t, "analyze")}, RunOptions{
+		Env: &Env{ManifestStore: store, Scriptorium: runner},
+	})
+	if !errors.Is(err, artifactErr) || !errors.Is(err, terminalErr) {
+		t.Fatalf("executeStages() error = %v, want analyze and terminal save failures", err)
+	}
+	if !store.Failed {
+		t.Fatal("terminal save failure was not injected")
+	}
+}
+
+func TestExecuteStagesPersistsAnalyzeCheckpointFailureRecovery(t *testing.T) {
+	cfg := testConfig(t)
+	cfg.Pipeline.Scriptorium = &config.ScriptoriumConfig{
+		Artifacts: map[string]config.ScriptoriumArtifactConfig{
+			"chronicle":  {Enabled: true, PromptID: "chronicle", OutputPath: "artifacts/chronicle.md"},
+			"encounters": {Enabled: true, PromptID: "encounters", OutputPath: "artifacts/encounters.md"},
+			"quest_log":  {Enabled: true, PromptID: "quest-log", OutputPath: "artifacts/quest_log.md"},
+		},
+	}
+	baseStore := &manifest.LocalStore{}
+	seed := manifest.New(cfg.Session.SessionID, time.Now().UTC())
+	seed.Campaign = cfg.Session.Campaign
+	seed.MarkStageSucceeded("publish", time.Now().UTC(), nil)
+	seed.MarkStageSucceeded("notify", time.Now().UTC(), nil)
+	saveBoundedManifest(t, cfg, seed)
+	checkpointErr := errors.New("injected artifact checkpoint failure")
+	store := &currentAnalyzeCheckpointFailingStore{
+		Store:        baseStore,
+		CurrentLimit: 2,
+		Err:          checkpointErr,
+	}
+	runner := &analyzeCheckpointRunner{}
+
+	_, err := executeStages(context.Background(), cfg, []stage.Stage{mustSelectStage(t, "analyze")}, RunOptions{
+		Env: &Env{ManifestStore: store, Scriptorium: runner},
+	})
+	if !errors.Is(err, checkpointErr) {
+		t.Fatalf("executeStages() error = %v, want checkpoint failure", err)
+	}
+	if want := []string{"chronicle", "encounters"}; !slices.Equal(runner.Calls, want) {
+		t.Fatalf("adapter calls = %#v, want %#v", runner.Calls, want)
+	}
+
+	persisted := loadAnalyzeRecoveryManifest(t, baseStore, cfg)
+	if persisted.Stages["analyze"].Status != manifest.StatusFailed {
+		t.Fatalf("aggregate analyze = %#v, want failed", persisted.Stages["analyze"])
+	}
+	assertAnalyzeArtifactStatus(t, persisted.Stages["analyze"], "chronicle", manifest.AnalyzeArtifactCurrent)
+	assertAnalyzeArtifactStatus(t, persisted.Stages["analyze"], "encounters", manifest.AnalyzeArtifactFailed)
+	if _, exists := persisted.Stages["analyze"].AnalyzeArtifacts["quest_log"]; exists {
+		t.Fatalf("unattempted quest log appeared: %#v", persisted.Stages["analyze"].AnalyzeArtifacts["quest_log"])
+	}
+	for _, name := range []string{"publish", "notify"} {
+		if got := persisted.Stages[name]; got == nil || got.Status != manifest.StatusStale {
+			t.Fatalf("%s stage = %#v, want stale", name, got)
+		}
+	}
+}
+
+type analyzeCheckpointRunner struct {
+	Calls      []string
+	FailPrompt string
+	Err        error
+}
+
+type failedAnalyzeSaveStore struct {
+	manifest.Store
+	Err    error
+	Failed bool
+}
+
+type currentAnalyzeCheckpointFailingStore struct {
+	manifest.Store
+	CurrentLimit int
+	Err          error
+	Failed       bool
+}
+
+func (s *currentAnalyzeCheckpointFailingStore) Save(ctx context.Context, path string, m *manifest.Manifest) error {
+	if !s.Failed {
+		current := 0
+		if record := m.Stages["analyze"]; record != nil && record.Status == manifest.StatusRunning {
+			for _, artifact := range record.AnalyzeArtifacts {
+				if artifact.Status == manifest.AnalyzeArtifactCurrent {
+					current++
+				}
+			}
+		}
+		if current >= s.CurrentLimit {
+			s.Failed = true
+			return s.Err
+		}
+	}
+	return s.Store.Save(ctx, path, m)
+}
+
+func (s *failedAnalyzeSaveStore) Save(ctx context.Context, path string, m *manifest.Manifest) error {
+	if record := m.Stages["analyze"]; !s.Failed && record != nil && record.Status == manifest.StatusFailed {
+		s.Failed = true
+		return s.Err
+	}
+	return s.Store.Save(ctx, path, m)
+}
+
+func (r *analyzeCheckpointRunner) RunArtifact(
+	ctx context.Context,
+	req scriptorium.RunArtifactRequest,
+) (scriptorium.ArtifactResult, error) {
+	if err := ctx.Err(); err != nil {
+		return scriptorium.ArtifactResult{}, err
+	}
+	r.Calls = append(r.Calls, req.PromptID)
+	if req.PromptID == r.FailPrompt {
+		return scriptorium.ArtifactResult{}, r.Err
+	}
+	if err := os.MkdirAll(filepath.Dir(req.OutputPath), 0o755); err != nil {
+		return scriptorium.ArtifactResult{}, err
+	}
+	if err := os.WriteFile(req.OutputPath, []byte("generated "+req.PromptID+"\n"), 0o644); err != nil {
+		return scriptorium.ArtifactResult{}, err
+	}
+	return scriptorium.ArtifactResult{
+		OutputPath:  req.OutputPath,
+		CommandMode: scriptorium.CommandModeRun,
+		PromptID:    req.PromptID,
+		ProfileID:   req.ProfileID,
+	}, nil
+}
+
+func (*analyzeCheckpointRunner) RenderArtifact(
+	context.Context,
+	scriptorium.RenderArtifactRequest,
+) (scriptorium.ArtifactResult, error) {
+	return scriptorium.ArtifactResult{}, errors.New("unexpected render")
+}
+
+func checkpointRecoveryArtifacts(revision string) map[string]config.ScriptoriumArtifactConfig {
+	return map[string]config.ScriptoriumArtifactConfig{
+		"chronicle": {
+			Enabled: true, PromptID: "chronicle-" + revision, OutputPath: "artifacts/chronicle.md",
+		},
+		"encounters": {
+			Enabled: true, DependsOn: []string{"chronicle"}, PromptID: "encounters-" + revision,
+			OutputPath: "artifacts/encounters.md",
+			Inputs: map[string]config.ScriptoriumInputConfig{
+				"chronicle": {Source: "narratio.artifact.chronicle", Required: true},
+			},
+		},
+		"quest_log": {
+			Enabled: true, DependsOn: []string{"encounters"}, PromptID: "quest-log-" + revision,
+			OutputPath: "artifacts/quest_log.md",
+			Inputs: map[string]config.ScriptoriumInputConfig{
+				"encounters": {Source: "narratio.artifact.encounters", Required: true},
+			},
+		},
+		"player_handout": {
+			Enabled: true, DependsOn: []string{"quest_log"}, PromptID: "player-handout-" + revision,
+			OutputPath: "artifacts/player_handout.md",
+			Inputs: map[string]config.ScriptoriumInputConfig{
+				"quest_log": {Source: "narratio.artifact.quest_log", Required: true},
+			},
+		},
+		"unrelated": {
+			Enabled: true, PromptID: "unrelated", OutputPath: "artifacts/unrelated.md",
+		},
+	}
+}
+
+func mustSelectStage(t *testing.T, name string) stage.Stage {
+	t.Helper()
+	selected, err := stage.Select(name)
+	if err != nil {
+		t.Fatalf("Select(%q) error = %v", name, err)
+	}
+	return selected
+}
+
+func loadAnalyzeRecoveryManifest(t *testing.T, store manifest.Store, cfg *config.Config) *manifest.Manifest {
+	t.Helper()
+	m, err := store.Load(context.Background(), manifestPathFor(cfg))
+	if err != nil {
+		t.Fatalf("load session manifest: %v", err)
+	}
+	return m
+}
+
+func assertAnalyzeArtifactStatus(
+	t *testing.T,
+	record *manifest.StageRecord,
+	key string,
+	want manifest.AnalyzeArtifactStatus,
+) {
+	t.Helper()
+	got, exists := record.AnalyzeArtifacts[key]
+	if !exists || got.Status != want {
+		t.Fatalf("artifact %q = %#v, want status %q", key, got, want)
+	}
+	if want != manifest.AnalyzeArtifactCurrent && got.Output != nil {
+		t.Fatalf("artifact %q retained output authority: %#v", key, got)
+	}
+}
+
+func cloneStageRecords(m *manifest.Manifest, names ...string) map[string]manifest.StageRecord {
+	clones := make(map[string]manifest.StageRecord, len(names))
+	for _, name := range names {
+		clones[name] = *m.Stages[name]
+	}
+	return clones
+}
+
+func assertStageRecordsUnchanged(t *testing.T, m *manifest.Manifest, want map[string]manifest.StageRecord) {
+	t.Helper()
+	for name, expected := range want {
+		got := m.Stages[name]
+		if got == nil {
+			t.Fatalf("upstream stage %q is missing", name)
+		}
+		if diff := cmp.Diff(expected, *got); diff != "" {
+			t.Fatalf("upstream stage %q changed (-want +got):\n%s", name, diff)
+		}
+	}
+}

internal/app/analyze_projection.go

@@ -1,134 +0,0 @@
-package app
-
-import (
-	"fmt"
-	"reflect"
-	"sort"
-
-	"gitea.maximumdirect.net/eric/narratio/internal/manifest"
-	"gitea.maximumdirect.net/eric/narratio/internal/stage"
-)
-
-type validatedAnalyzeProjection struct {
-	session    map[string]manifest.AnalyzeArtifactRecord
-	invocation map[string]manifest.AnalyzeArtifactRecord
-}
-
-type analyzeStateSnapshot struct {
-	version int
-	records map[string]manifest.AnalyzeArtifactRecord
-}
-
-func captureAnalyzeState(manifestValue *manifest.Manifest, stageName string) analyzeStateSnapshot {
-	if manifestValue == nil || stageName != "analyze" || manifestValue.Stages["analyze"] == nil {
-		return analyzeStateSnapshot{}
-	}
-	record := manifestValue.Stages["analyze"]
-	return analyzeStateSnapshot{
-		version: record.AnalyzeStateVersion,
-		records: manifest.CloneAnalyzeArtifactCollection(record.AnalyzeArtifacts),
-	}
-}
-
-func restoreAnalyzeState(manifestValue *manifest.Manifest, snapshot analyzeStateSnapshot) {
-	if manifestValue == nil || manifestValue.Stages["analyze"] == nil {
-		return
-	}
-	record := manifestValue.Stages["analyze"]
-	record.AnalyzeStateVersion = snapshot.version
-	record.AnalyzeArtifacts = manifest.CloneAnalyzeArtifactCollection(snapshot.records)
-}
-
-func validateSuccessfulAnalyzeProjection(stageName string, result *stage.StageResult) (*validatedAnalyzeProjection, error) {
-	if result == nil || result.AnalyzeState == nil {
-		return nil, nil
-	}
-	if stageName != "analyze" {
-		return nil, fmt.Errorf("stage %q returned analyze-owned state projection", stageName)
-	}
-	if result.Disposition == stage.StageDispositionSkipped {
-		return nil, fmt.Errorf("skipped analyze result cannot contain analyze-owned state projection")
-	}
-	if len(result.Outputs) != 0 {
-		return nil, fmt.Errorf("analyze result with state projection cannot contain ordinary outputs")
-	}
-	return validateAndCloneAnalyzeProjection(result.AnalyzeState)
-}
-
-func validateFailedAnalyzeProjection(stageName string, result *stage.StageResult) (*validatedAnalyzeProjection, error) {
-	if result == nil || result.AnalyzeState == nil {
-		return nil, nil
-	}
-	if stageName != "analyze" {
-		return nil, fmt.Errorf("stage %q returned analyze-owned state projection with an error", stageName)
-	}
-	if result.Disposition != stage.StageDispositionSucceeded || result.SkipReason != "" || len(result.Outputs) != 0 || len(result.Logs) != 0 || len(result.GeneratedConfigs) != 0 || len(result.Metadata) != 0 {
-		return nil, fmt.Errorf("analyze result with an error may contain only analyze-owned state projection")
-	}
-	return validateAndCloneAnalyzeProjection(result.AnalyzeState)
-}
-
-func validateAndCloneAnalyzeProjection(projection *stage.AnalyzeStateProjection) (*validatedAnalyzeProjection, error) {
-	session := manifest.CloneAnalyzeArtifactCollection(projection.Session)
-	invocation := manifest.CloneAnalyzeArtifactCollection(projection.Invocation)
-	if err := manifest.ValidateAnalyzeArtifactCollection(manifest.AnalyzeStateContractVersion, session); err != nil {
-		return nil, fmt.Errorf("validate reconciled session analyze state: %w", err)
-	}
-	if err := manifest.ValidateAnalyzeArtifactCollection(manifest.AnalyzeStateContractVersion, invocation); err != nil {
-		return nil, fmt.Errorf("validate invocation analyze state: %w", err)
-	}
-	for key, invocationRecord := range invocation {
-		sessionRecord, ok := session[key]
-		if !ok {
-			return nil, fmt.Errorf("invocation analyze artifact %q is absent from reconciled session state", key)
-		}
-		if !reflect.DeepEqual(invocationRecord, sessionRecord) {
-			return nil, fmt.Errorf("invocation analyze artifact %q contradicts reconciled session state", key)
-		}
-	}
-	return &validatedAnalyzeProjection{session: session, invocation: invocation}, nil
-}
-
-func applyAnalyzeProjection(
-	sessionManifest *manifest.Manifest,
-	runManifest *manifest.RunManifest,
-	projection *validatedAnalyzeProjection,
-) {
-	if projection == nil {
-		return
-	}
-	if sessionManifest != nil && sessionManifest.Stages["analyze"] != nil {
-		record := sessionManifest.Stages["analyze"]
-		record.AnalyzeStateVersion = manifest.AnalyzeStateContractVersion
-		record.AnalyzeArtifacts = manifest.CloneAnalyzeArtifactCollection(projection.session)
-	}
-	if runManifest != nil && runManifest.Stages["analyze"] != nil {
-		record := runManifest.Stages["analyze"]
-		record.AnalyzeStateVersion = manifest.AnalyzeStateContractVersion
-		record.AnalyzeArtifacts = manifest.CloneAnalyzeArtifactCollection(projection.invocation)
-	}
-}
-
-func analyzeProjectionOutputs(records map[string]manifest.AnalyzeArtifactRecord, producerRunID string) []manifest.ArtifactRecord {
-	keys := make([]string, 0, len(records))
-	for key, record := range records {
-		if record.Status != manifest.AnalyzeArtifactCurrent || record.Output == nil {
-			continue
-		}
-		if producerRunID != "" && record.ProducerRunID != producerRunID {
-			continue
-		}
-		keys = append(keys, key)
-	}
-	sort.Strings(keys)
-	outputs := make([]manifest.ArtifactRecord, 0, len(keys))
-	for _, key := range keys {
-		record := manifest.CloneAnalyzeArtifactCollection(map[string]manifest.AnalyzeArtifactRecord{key: records[key]})[key]
-		output := *record.Output
-		if output.ProducerRunID == "" {
-			output.ProducerRunID = record.ProducerRunID
-		}
-		outputs = append(outputs, output)
-	}
-	return outputs
-}

internal/app/analyze_projection_test.go

@@ -1,346 +0,0 @@
-package app
-
-import (
-	"context"
-	"errors"
-	"os"
-	"path/filepath"
-	"reflect"
-	"strings"
-	"testing"
-	"time"
-
-	"gitea.maximumdirect.net/eric/narratio/internal/artifactmodel"
-	"gitea.maximumdirect.net/eric/narratio/internal/artifacts"
-	"gitea.maximumdirect.net/eric/narratio/internal/manifest"
-	"gitea.maximumdirect.net/eric/narratio/internal/stage"
-)
-
-type projectionStage struct {
-	name string
-	run  func(*stage.Env, *manifest.Manifest) (*stage.StageResult, error)
-}
-
-func (s projectionStage) Name() string { return s.name }
-func (s projectionStage) Run(_ context.Context, env *stage.Env, m *manifest.Manifest) (*stage.StageResult, error) {
-	return s.run(env, m)
-}
-
-func TestExecuteStagesProjectsSeparateSessionAndInvocationAnalyzeState(t *testing.T) {
-	cfg := testConfig(t)
-	oldAt := time.Date(2026, 5, 3, 9, 0, 0, 0, time.UTC)
-	oldRecord := appAnalyzeRecord("player_handout", manifest.AnalyzeArtifactCurrent, "old-run", oldAt)
-
-	stageToRun := projectionStage{name: "analyze", run: func(_ *stage.Env, m *manifest.Manifest) (*stage.StageResult, error) {
-		newRecord := appAnalyzeRecord("session_recap", manifest.AnalyzeArtifactCurrent, m.RunID, time.Now().UTC())
-		staleRecord := appAnalyzeRecord("quest_log", manifest.AnalyzeArtifactStale, m.RunID, time.Now().UTC())
-		session := map[string]manifest.AnalyzeArtifactRecord{
-			"player_handout": oldRecord,
-			"quest_log":      staleRecord,
-			"session_recap":  newRecord,
-		}
-		return &stage.StageResult{
-			Logs: []string{"aggregate-analyze.log"},
-			AnalyzeState: &stage.AnalyzeStateProjection{
-				Session: session,
-				Invocation: map[string]manifest.AnalyzeArtifactRecord{
-					"player_handout": oldRecord,
-					"session_recap":  newRecord,
-				},
-			},
-		}, nil
-	}}
-
-	summary, err := executeStages(context.Background(), cfg, []stage.Stage{stageToRun}, RunOptions{})
-	if err != nil {
-		t.Fatalf("executeStages() error = %v", err)
-	}
-	store := &manifest.LocalStore{}
-	sessionManifest, err := store.Load(context.Background(), summary.ManifestPath)
-	if err != nil {
-		t.Fatalf("Load(session) error = %v", err)
-	}
-	analyze := sessionManifest.Stages["analyze"]
-	if analyze.AnalyzeStateVersion != manifest.AnalyzeStateContractVersion || len(analyze.AnalyzeArtifacts) != 3 {
-		t.Fatalf("session analyze state = %#v", analyze)
-	}
-	if got := analyzeArtifactOutputKeys(analyze.Outputs); !reflect.DeepEqual(got, []string{"player_handout", "session_recap"}) {
-		t.Fatalf("session aggregate outputs = %#v, want current records only", got)
-	}
-	if len(analyze.Logs) != 1 || analyze.Logs[0] != "aggregate-analyze.log" {
-		t.Fatalf("session aggregate logs = %#v", analyze.Logs)
-	}
-
-	runManifest, err := store.LoadRun(context.Background(), summary.RunManifestPath)
-	if err != nil {
-		t.Fatalf("LoadRun() error = %v", err)
-	}
-	runAnalyze := runManifest.Stages["analyze"]
-	if len(runAnalyze.AnalyzeArtifacts) != 2 || runAnalyze.AnalyzeArtifacts["session_recap"].Status != manifest.AnalyzeArtifactCurrent {
-		t.Fatalf("invocation analyze state = %#v", runAnalyze.AnalyzeArtifacts)
-	}
-	if got := analyzeArtifactOutputKeys(runAnalyze.Outputs); !reflect.DeepEqual(got, []string{"session_recap"}) {
-		t.Fatalf("invocation outputs = %#v, want produced artifact only", got)
-	}
-}
-
-func TestExecuteStagesPersistsRestrictedAnalyzeStateOnPartialError(t *testing.T) {
-	cfg := testConfig(t)
-	now := time.Date(2026, 5, 3, 9, 0, 0, 0, time.UTC)
-	seed := manifest.New(cfg.Session.SessionID, now)
-	seed.Campaign = cfg.Session.Campaign
-	seed.MarkStageSucceeded("analyze", now, nil)
-	seed.Stages["analyze"].AnalyzeStateVersion = manifest.AnalyzeStateContractVersion
-	seed.Stages["analyze"].AnalyzeArtifacts = map[string]manifest.AnalyzeArtifactRecord{
-		"player_handout": appAnalyzeRecord("player_handout", manifest.AnalyzeArtifactCurrent, "old-run", now),
-	}
-	seed.MarkStageSucceeded("publish", now, nil)
-	saveBoundedManifest(t, cfg, seed)
-
-	stageToRun := projectionStage{name: "analyze", run: func(_ *stage.Env, m *manifest.Manifest) (*stage.StageResult, error) {
-		unrelated := seed.Stages["analyze"].AnalyzeArtifacts["player_handout"]
-		completed := appAnalyzeRecord("session_recap", manifest.AnalyzeArtifactCurrent, m.RunID, time.Now().UTC())
-		failed := appAnalyzeRecord("quest_log", manifest.AnalyzeArtifactFailed, m.RunID, time.Now().UTC())
-		session := map[string]manifest.AnalyzeArtifactRecord{
-			"player_handout": unrelated,
-			"quest_log":      failed,
-			"session_recap":  completed,
-		}
-		return &stage.StageResult{AnalyzeState: &stage.AnalyzeStateProjection{
-			Session: session,
-			Invocation: map[string]manifest.AnalyzeArtifactRecord{
-				"quest_log":     failed,
-				"session_recap": completed,
-			},
-		}}, errors.New("quest log failed")
-	}}
-
-	_, err := executeStages(context.Background(), cfg, []stage.Stage{stageToRun}, RunOptions{Force: true})
-	if err == nil || !strings.Contains(err.Error(), "quest log failed") {
-		t.Fatalf("executeStages() error = %v", err)
-	}
-	store := &manifest.LocalStore{}
-	loaded, err := store.Load(context.Background(), manifestPathFor(cfg))
-	if err != nil {
-		t.Fatalf("Load(session) error = %v", err)
-	}
-	analyze := loaded.Stages["analyze"]
-	if analyze.Status != manifest.StatusFailed || len(analyze.Outputs) != 0 {
-		t.Fatalf("aggregate analyze state = %#v, want failed without outputs", analyze)
-	}
-	if analyze.AnalyzeArtifacts["player_handout"].Status != manifest.AnalyzeArtifactCurrent ||
-		analyze.AnalyzeArtifacts["session_recap"].Status != manifest.AnalyzeArtifactCurrent ||
-		analyze.AnalyzeArtifacts["quest_log"].Status != manifest.AnalyzeArtifactFailed {
-		t.Fatalf("partial session projection = %#v", analyze.AnalyzeArtifacts)
-	}
-	if loaded.Stages["publish"].Status != manifest.StatusStale {
-		t.Fatalf("publish status = %q, want stale", loaded.Stages["publish"].Status)
-	}
-
-	runsDir := artifacts.SessionRunsDirForCampaign(cfg.Pipeline.Workspace.Root, cfg.Session.Campaign, cfg.Session.SessionID)
-	entries, err := os.ReadDir(runsDir)
-	if err != nil || len(entries) != 1 {
-		t.Fatalf("run directory entries = %#v, error = %v", entries, err)
-	}
-	runManifest, err := store.LoadRun(context.Background(), filepath.Join(runsDir, entries[0].Name(), "manifest.json"))
-	if err != nil {
-		t.Fatalf("LoadRun() error = %v", err)
-	}
-	runAnalyze := runManifest.Stages["analyze"]
-	if runAnalyze.Status != manifest.StatusFailed || len(runAnalyze.AnalyzeArtifacts) != 2 || len(runAnalyze.Outputs) != 0 {
-		t.Fatalf("partial invocation projection = %#v", runAnalyze)
-	}
-}
-
-func TestExecuteStagesRejectsInvalidAnalyzeProjectionWithoutReplacingPriorState(t *testing.T) {
-	cfg := testConfig(t)
-	now := time.Date(2026, 5, 3, 9, 0, 0, 0, time.UTC)
-	prior := appAnalyzeRecord("player_handout", manifest.AnalyzeArtifactCurrent, "old-run", now)
-	seed := manifest.New(cfg.Session.SessionID, now)
-	seed.Campaign = cfg.Session.Campaign
-	seed.MarkStageSucceeded("analyze", now, nil)
-	seed.Stages["analyze"].AnalyzeStateVersion = manifest.AnalyzeStateContractVersion
-	seed.Stages["analyze"].AnalyzeArtifacts = map[string]manifest.AnalyzeArtifactRecord{"player_handout": prior}
-	saveBoundedManifest(t, cfg, seed)
-
-	stageToRun := projectionStage{name: "analyze", run: func(_ *stage.Env, m *manifest.Manifest) (*stage.StageResult, error) {
-		invalid := appAnalyzeRecord("session_recap", manifest.AnalyzeArtifactCurrent, m.RunID, time.Now().UTC())
-		invalid.Output.Checksum = "invalid"
-		return &stage.StageResult{AnalyzeState: &stage.AnalyzeStateProjection{
-			Session:    map[string]manifest.AnalyzeArtifactRecord{"session_recap": invalid},
-			Invocation: map[string]manifest.AnalyzeArtifactRecord{"session_recap": invalid},
-		}}, errors.New("analysis failed")
-	}}
-
-	_, err := executeStages(context.Background(), cfg, []stage.Stage{stageToRun}, RunOptions{Force: true})
-	if err == nil || !strings.Contains(err.Error(), "checksum") {
-		t.Fatalf("executeStages() error = %v, want projection validation failure", err)
-	}
-	loaded, loadErr := (&manifest.LocalStore{}).Load(context.Background(), manifestPathFor(cfg))
-	if loadErr != nil {
-		t.Fatalf("Load() error = %v", loadErr)
-	}
-	if len(loaded.Stages["analyze"].AnalyzeArtifacts) != 1 || !reflect.DeepEqual(loaded.Stages["analyze"].AnalyzeArtifacts["player_handout"], prior) {
-		t.Fatalf("prior state replaced by invalid projection: %#v", loaded.Stages["analyze"].AnalyzeArtifacts)
-	}
-}
-
-func TestExecuteStagesRollsBackAnalyzeAuthorityWhenProjectionSaveFails(t *testing.T) {
-	cfg := testConfig(t)
-	now := time.Date(2026, 5, 3, 9, 0, 0, 0, time.UTC)
-	prior := appAnalyzeRecord("player_handout", manifest.AnalyzeArtifactCurrent, "old-run", now)
-	seed := manifest.New(cfg.Session.SessionID, now)
-	seed.Campaign = cfg.Session.Campaign
-	seed.MarkStageSucceeded("analyze", now, nil)
-	seed.Stages["analyze"].AnalyzeStateVersion = manifest.AnalyzeStateContractVersion
-	seed.Stages["analyze"].AnalyzeArtifacts = map[string]manifest.AnalyzeArtifactRecord{"player_handout": prior}
-	saveBoundedManifest(t, cfg, seed)
-
-	stageReturned := false
-	store := &analyzeProjectionFailingStore{delegate: &manifest.LocalStore{}, shouldFail: func(m *manifest.Manifest) bool {
-		return stageReturned && m.Stages["analyze"] != nil && m.Stages["analyze"].Status == manifest.StatusSucceeded && m.Stages["analyze"].AnalyzeArtifacts["session_recap"].Status == manifest.AnalyzeArtifactCurrent
-	}}
-	stageToRun := projectionStage{name: "analyze", run: func(_ *stage.Env, m *manifest.Manifest) (*stage.StageResult, error) {
-		stageReturned = true
-		newRecord := appAnalyzeRecord("session_recap", manifest.AnalyzeArtifactCurrent, m.RunID, time.Now().UTC())
-		return &stage.StageResult{AnalyzeState: &stage.AnalyzeStateProjection{
-			Session: map[string]manifest.AnalyzeArtifactRecord{
-				"player_handout": prior,
-				"session_recap":  newRecord,
-			},
-			Invocation: map[string]manifest.AnalyzeArtifactRecord{"session_recap": newRecord},
-		}}, nil
-	}}
-
-	_, err := executeStages(context.Background(), cfg, []stage.Stage{stageToRun}, RunOptions{
-		Force: true,
-		Env:   &Env{ManifestStore: store},
-	})
-	if err == nil || !strings.Contains(err.Error(), "injected analyze projection save failure") {
-		t.Fatalf("executeStages() error = %v", err)
-	}
-	if !store.failed {
-		t.Fatal("projection persistence failure was not injected")
-	}
-	loaded, loadErr := store.delegate.Load(context.Background(), manifestPathFor(cfg))
-	if loadErr != nil {
-		t.Fatalf("Load() error = %v", loadErr)
-	}
-	analyze := loaded.Stages["analyze"]
-	if analyze.Status != manifest.StatusFailed || len(analyze.AnalyzeArtifacts) != 1 || !reflect.DeepEqual(analyze.AnalyzeArtifacts["player_handout"], prior) {
-		t.Fatalf("durable analyze state after rollback = %#v", analyze)
-	}
-}
-
-func TestExecuteStagesRejectsAnalyzeProjectionFromOtherStage(t *testing.T) {
-	cfg := testConfig(t)
-	stageToRun := projectionStage{name: "prepare", run: func(_ *stage.Env, m *manifest.Manifest) (*stage.StageResult, error) {
-		record := appAnalyzeRecord("session_recap", manifest.AnalyzeArtifactCurrent, m.RunID, time.Now().UTC())
-		return &stage.StageResult{AnalyzeState: &stage.AnalyzeStateProjection{
-			Session: map[string]manifest.AnalyzeArtifactRecord{"session_recap": record},
-		}}, nil
-	}}
-	_, err := executeStages(context.Background(), cfg, []stage.Stage{stageToRun}, RunOptions{})
-	if err == nil || !strings.Contains(err.Error(), "returned analyze-owned state projection") {
-		t.Fatalf("executeStages() error = %v", err)
-	}
-}
-
-func TestExecuteStagesRejectsContradictoryAnalyzeResultWithError(t *testing.T) {
-	cfg := testConfig(t)
-	stageToRun := projectionStage{name: "analyze", run: func(_ *stage.Env, m *manifest.Manifest) (*stage.StageResult, error) {
-		record := appAnalyzeRecord("session_recap", manifest.AnalyzeArtifactCurrent, m.RunID, time.Now().UTC())
-		return &stage.StageResult{
-			Outputs: []artifacts.Ref{{Kind: "session_recap", RelativePath: "artifacts/session-recap.md"}},
-			AnalyzeState: &stage.AnalyzeStateProjection{
-				Session:    map[string]manifest.AnalyzeArtifactRecord{"session_recap": record},
-				Invocation: map[string]manifest.AnalyzeArtifactRecord{"session_recap": record},
-			},
-		}, errors.New("analysis failed")
-	}}
-	_, err := executeStages(context.Background(), cfg, []stage.Stage{stageToRun}, RunOptions{})
-	if err == nil || !strings.Contains(err.Error(), "may contain only analyze-owned state projection") {
-		t.Fatalf("executeStages() error = %v", err)
-	}
-}
-
-func TestExecuteStagesExposesSelectedForceDecisionToStage(t *testing.T) {
-	for _, force := range []bool{false, true} {
-		t.Run(strings.ToLower(strings.TrimSpace(map[bool]string{false: "ordinary", true: "forced"}[force])), func(t *testing.T) {
-			cfg := testConfig(t)
-			captured := !force
-			stageToRun := projectionStage{name: "prepare", run: func(env *stage.Env, _ *manifest.Manifest) (*stage.StageResult, error) {
-				captured = env.Force
-				return &stage.StageResult{}, nil
-			}}
-			if _, err := executeStages(context.Background(), cfg, []stage.Stage{stageToRun}, RunOptions{Force: force}); err != nil {
-				t.Fatalf("executeStages() error = %v", err)
-			}
-			if captured != force {
-				t.Fatalf("stage env force = %v, want %v", captured, force)
-			}
-		})
-	}
-}
-
-func appAnalyzeRecord(key string, status manifest.AnalyzeArtifactStatus, producerRunID string, at time.Time) manifest.AnalyzeArtifactRecord {
-	record := manifest.AnalyzeArtifactRecord{
-		Key:           key,
-		Status:        status,
-		ProducerRunID: producerRunID,
-		UpdatedAt:     at,
-	}
-	if status == manifest.AnalyzeArtifactFailed {
-		record.Error = "scriptorium failed"
-		return record
-	}
-	if status != manifest.AnalyzeArtifactCurrent {
-		record.FingerprintVersion = manifest.AnalyzeFingerprintContractVersion
-		record.Fingerprint = strings.Repeat("b", 64)
-		return record
-	}
-	record.FingerprintVersion = manifest.AnalyzeFingerprintContractVersion
-	record.Fingerprint = strings.Repeat("a", 64)
-	record.OutputSize = 42
-	record.Output = &manifest.ArtifactRecord{
-		Kind:          key,
-		SourceID:      artifacts.ConfiguredArtifactSourceID(key),
-		LocalPath:     "artifacts/" + strings.ReplaceAll(key, "_", "-") + ".md",
-		ProducerRunID: producerRunID,
-		Checksum:      strings.Repeat("c", 64),
-		Contract: &artifactmodel.ContractMetadata{
-			MediaType: "text/markdown", SchemaID: "narratio." + key, SchemaVersion: "1",
-		},
-	}
-	return record
-}
-
-func analyzeArtifactOutputKeys(outputs []manifest.ArtifactRecord) []string {
-	keys := make([]string, 0, len(outputs))
-	for _, output := range outputs {
-		keys = append(keys, strings.TrimPrefix(output.SourceID, "narratio.artifact."))
-	}
-	return keys
-}
-
-type analyzeProjectionFailingStore struct {
-	delegate   *manifest.LocalStore
-	shouldFail func(*manifest.Manifest) bool
-	failed     bool
-}
-
-func (s *analyzeProjectionFailingStore) Create(ctx context.Context, sessionID string) (*manifest.Manifest, error) {
-	return s.delegate.Create(ctx, sessionID)
-}
-
-func (s *analyzeProjectionFailingStore) Load(ctx context.Context, path string) (*manifest.Manifest, error) {
-	return s.delegate.Load(ctx, path)
-}
-
-func (s *analyzeProjectionFailingStore) Save(ctx context.Context, path string, m *manifest.Manifest) error {
-	if !s.failed && s.shouldFail != nil && s.shouldFail(m) {
-		s.failed = true
-		return errors.New("injected analyze projection save failure")
-	}
-	return s.delegate.Save(ctx, path, m)
-}