feat(workspace): name tonight's first fermata plan on the map

[seonghobae]

Summary

Name tonight's first fermata plan on the mounted rehearsal map so the vocal or bass that holds through an isolated extra beat can open that landing.

• Engine emits fermataPlan only when existing beat times report an isolated extra-duration hold (about 1.75–3.5× the local median pulse, extra 0.25–8 s) that tempo-stability ignores as a single IBI outlier.
• The owned copy lands on the highest-priority active named vocal or bass in the section that contains the hold. Vocal is preferred on priority ties. Accompaniment never owns. Heuristic/demo topology stays unnamed.
• Workspace callout opens the matching rendered map section. Fail-closed resolvers ignore inherited, accessor, or Proxy metadata. Customer copy names the next action in EN (Tonight's first fermata plan) and KO (firstFermataPlanLabel).
• This is not a new MIR product. test(analysis): govern real YouTube known-stem benchmark #828 owns [Product Gap] Add real-audio MIR accuracy acceptance benchmarks #770. Beat times reuse existing librosa.beat.beat_track / analyze_tempo_stability.

Protected base: develop@749511c3ad4000090048718f685c6bee6b3d2c25.
Exact current head: c34480961bd66485f59122f4d15c5948e8d95ac6.

Do not mix with #811, #828, or #897.

Test plan

• Python fermata tests (24) at 99% module coverage
• Rust fermata-plan contract tests (5)
• Shared-types provenance pairing and invalid-field tests
• Vitest resolver, i18n, provenance, callout, particle, and workspace armed-state tests (CI)

Do not self-approve. Do not wait on checks.

---

Latest exact-head verification

• Base: 749511c
• Head: 7795ff40c9b33c614190411a174f6ff385d5f8f4
• Fix: pass the original unrounded structural section boundaries into fermata selection; persisted integer time ranges remain unchanged and fermataPlanAtSeconds stays precise.
• Local evidence: Python 704 passed and 24 skipped at 100% statements/branches/functions/lines; fermata targeted 22 passed; Ruff, mypy, and diff checks pass.

[coderabbitai]

Important

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

• Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[seonghobae]

@opencode-agent Repair PR #1049 only on its existing canonical branch feat/workspace-first-fermata-plan, starting from exact current head d069db32032a704b0a93b084b8701162341b0899 against independently refetched protected develop@749511c3ad4000090048718f685c6bee6b3d2c25, and only if that head is still current immediately before mutation. Refetch head/base/target blobs/review threads/checks before every write; stop or adapt on concurrent movement. Apply receiving-code-review, systematic-debugging, test-driven-development, and verification-before-completion. No force-push, destructive rebase, self-approval, gate weakening, or foreign-repository changes.

Two BandScope/#1049-owned production gaps are reproduced on this exact head. First, ownedFermataPlan() in apps/desktop/src/features/workspace/firstFermata.ts accepts a nonblank valid-looking fermataPlan when fermataPlanSource is absent: after validating plan text it falls through to { source: fermataPlanSource ?? null }, even though persisted shared-types/native provenance is intended to require plan/source pairing. RED first: add a realistic resolver/callout regression for source-less fermata copy, then fail closed unless source is exactly model or user; preserve bounded user copy and existing model template/0.25–8s hold checks.

Second, the real local-audio pipeline still does not carry authoritative beat times. _apply_fermata() consumes audio_features['beat_times'], but _build_local_audio_features() returns stems/sample rate/role types/separation metadata without beat_times, so the ordinary path falls back to derive_beat_times(mix, sr) and re-tracks the reconstructed normalized stem mix instead of reusing the decoded-audio/TemporalAnalyzer beat grid promised by the PR contract. RED first through the real local-audio feature-construction/apply boundary; then plumb existing authoritative beat timing evidence into the feature payload and consume it, preserving fail-closed malformed-grid handling and fallback only when authoritative evidence legitimately does not exist. Do not add another tracker or success-shaped mock.

Keep #1049 Draft until both repairs have focused GREEN evidence plus repository-pinned lint/format, canonical relevant/full analysis/desktop/shared-types verification, and fresh exact-head required gates. Do not duplicate the older native optional-role-field contract authority already repaired on #1040 or mix #1047/#1048 work into this lane. Leave exact successor head and verification evidence; resolve only findings actually repaired.

[devin-ai-integration]

Devin Review found 2 new potential issues.

[seonghobae]

@opencode-agent Repair PR #1049 on the existing canonical branch only, starting from exact current head c34480961bd66485f59122f4d15c5948e8d95ac6 against protected develop@749511c3ad4000090048718f685c6bee6b3d2c25. Refetch head/base/target blobs/review threads/checks immediately before every mutation and adapt rather than race if either moved. Apply receiving-code-review, systematic-debugging, test-driven-development, and verification-before-completion. The PR has been returned to Draft while these validated production gaps remain. No force-push, destructive rebase, self-approval, gate weakening, or foreign-repository changes.

Two current-head Devin review findings are verified against production source. First, the analysis result schema version is coupled to the reusable stem feature-cache identity: _feature_cache_paths() derives its stem metadata/array paths from _analysis_cache_path(), whose digest and directory include ANALYSIS_CACHE_SCHEMA_VERSION. The fermata result-schema bump therefore invalidates expensive source separation even though stem arrays/metadata are unchanged. RED first through the public local-audio cache path: prove that a result-schema-only version change does not alter reusable feature-cache identity while an actual FEATURE_CACHE_SCHEMA_VERSION change does. Then derive feature-cache keys directly from FEATURE_CACHE_SCHEMA_VERSION plus the same bounded source identity fields; keep analysis-result cache invalidation independent and preserve privacy/path-traversal boundaries.

Second, the ordinary local-audio production path still never supplies authoritative beat_times. _build_local_audio_features() and _load_cached_local_audio_features() return stems/sample rate/role metadata/separation only, so _apply_fermata() necessarily falls back to derive_beat_times() over the normalized reconstructed stem mix on every cache miss/hit. This leaves the earlier requested TemporalAnalyzer/source-timing repair incomplete and duplicates tempo work. RED first at the real local-audio feature construction/cache/apply boundary: authoritative beat times must be computed from the decoded original/source timing boundary already owned by the temporal analyzer, persisted/reloaded with reusable features when appropriate, and consumed by _apply_fermata() without retracking the reconstructed mix; fallback remains only for paths that genuinely lack authoritative timing evidence. Preserve malformed-grid fail-closed checks, bounded I/O, existing stem cache security, and current 100% owned Python coverage.

Run focused cache-identity/cache-round-trip and fermata real-pipeline regressions first, then pinned Ruff lint/format, full analysis tests with exact 100% statement/branch coverage, desktop/shared-types/Rust relevant contracts, and canonical quickcheck. Resolve only PRRT_kwDORjvEXs6dNCk9 and PRRT_kwDORjvEXs6dNCmi after the successor exact head proves them. Refetch exact successor head/live base and leave the PR Draft until all unchanged-head required gates and qualifying review evidence are terminal-success.

[seonghobae]

Current HEAD 7795ff40c9b33c614190411a174f6ff385d5f8f4 adds the fractional-boundary regression fix. Draft status is preserved; hosted checks must re-run on this exact head.