Harden prepare and transcribe transitions

docs/internal/stage-prepare.md

@@ -30,10 +30,12 @@ Materialize canonical current-session inputs before processing stages.
 
 - validates required config/store state.
 - enforces local audio vs S3 audio mutual exclusivity.
+- rejects duplicate explicit local audio sources after resolution.
+- gives distinct local source paths with the same basename deterministic unique
+  prepared filenames so neither source is overwritten.
 - materializes S3 audio through spool/cache-aware logic.
 - scans enabled configured artifact inputs for `narratio.previous_session.artifact.*` requirements.
-- when previous requirements exist:
-  - clears managed `previous/` state;
+- clears managed `previous/` state on every invocation, then, when requirements exist:
   - resolves the pointer-selected previous source through the shared resolver;
   - downloads previous manifest/artifacts;
   - records previous inputs in `manifest.inputs`.
@@ -47,6 +49,7 @@ mapping, while the isolated legacy reader rejects ambiguous fallback matches.
 - only `prepare` hydrates canonical `previous/` cache state.
 - managed previous artifacts are stored under `previous/artifacts/**` without
   duplicate `artifacts/artifacts/` nesting.
+- managed `previous/` state represents only the current requirement set.
 - `manifest.inputs` ordering is deterministic (`kind`, `path`).
 
 ## Related Contracts And Tests

docs/internal/stage-transcribe.md

@@ -15,16 +15,19 @@ Generate raw per-speaker transcripts from prepared audio using WhisperX.
 ## Key Behavior
 
 - discovers prepared audio from manifest inputs or canonical audio directory.
-- derives speaker ID from `.flac` basename.
+- derives the transcript identity from the prepared `.flac` filename.
 - dispatches WhisperX requests through a bounded worker pool.
 - validates each output as JSON.
-- writes run-local outputs then materializes canonical transcript outputs.
+- writes run-local outputs then materializes canonical transcript outputs only
+  after every planned request succeeds.
 
 ## Invariants
 
-- speaker basenames must be unique.
+- prepared audio identities must be unique; prepare disambiguates distinct
+  source paths that share a basename.
 - output path returned by adapter must match requested output path.
-- each successful output is validated before stage success.
+- each successful output is validated before stage success, and cancellation or
+  incomplete dispatch cannot be reported as a successful result.
 
 ## Related Contracts And Tests
 

docs/roadmap/implementation.md

@@ -840,6 +840,8 @@ cancellation at each phase, partial completion, duplicate source, same-basename
 distinct source, retry, and no-false-manifest tests. This is the primary TST-008
 stage; Stage 25 adds output/resolver boundary cases.
 
+**Status:** Completed.
+
 ## Stage 25 — Enforce output-path authority and shared singleton resolution
 
 **Read first:** `audit-findings.md` lines 3306–3330 (ARC-006), 3515–3537