fix(audio): establish canonical local-audio resource policy

[seonghobae]

Buyer/security gap

Advances #781 by making local-audio resource admission one versioned policy before expensive decode, resampling, downmixing, transcription, separation, or downstream analysis. The 100 MiB encoded-byte ceiling remains shared across native local-file/YouTube bootstrap, the desktop bridge, Python request/decode admission, and YouTube download admission.

Exact current identity and ownership

• Protected integration base: develop@acdbea6344fe1231c39535b575f4de35e4c607c9.
• Exact current head: c2cc5bbeda6628fa9999401d6b0d228cb9b6bb9c.
• Branch: fix/audio-resource-policy-781.
• Draft retained; this lane is not Ready or merged.
• Dependency predecessor: fix(score): bound native PDF reads before allocation #865, current PR head faf565d92d464de8edb968ee123e6e953eb1b5d6, which must integrate into protected develop first.
• feat(analysis): enforce one canonical audio resource policy (#781) #985 is reference-only for this work; its unique source-container preflight requirement is integrated here, so no competing canonical owner is created.

Implemented policy-v1 boundary

• AudioResourcePolicy owns the 100 MiB encoded ceiling, target rate, decoded duration/sample/memory ceilings, numeric validation, and inclusive source-container bounds of 8 kHz–192 kHz and mono/stereo.
• preflight_audio_metadata() reads soundfile.info() from the already-open caller-owned handle, rejects malformed/overlong/unsupported source metadata before transformation, and rewinds the handle for the decoder. It does not open paths, follow URLs, or decode PCM.
• TemporalAnalyzer, AudioStemSeparator, and bass transcription run that preflight before librosa.load(..., sr=..., mono=True, duration=...); decoded output is then revalidated before beat tracking, pYIN, or Demucs/model work.
• Decoder calls request one sample beyond the accepted duration so longer sources cannot be silently truncated. Empty, non-finite, wrong-shaped/rate/dtype, over-budget, and float32-overflowed model outputs fail closed.
• Native desktop-core and analysis.ts retain filesystem-observed/request metadata byte admission for local-file and YouTube bootstrap. YouTube additionally rejects announced oversize before download, uses max_filesize and an in-flight progress abort, removes only owned in-directory partials, and canonicalizes completed paths before cleanup or success.
• Non-CPU Demucs output crosses to host CPU before NumPy conversion; this is a device-to-host safety boundary, not GPU budget or CPU/GPU parity evidence.

Security Notes

• Source bytes, container headers, decoder output, model output, local paths, and YouTube metadata are untrusted.
• Admission is allowlisted by finite byte, duration, source-rate/channel, decoded-shape/dtype/finiteness, and memory/sample contracts. Parser/rewind/decoder failures return bounded payload-free errors and do not disclose paths, URLs, cookies, or audio content.
• The source probe uses an already-open handle and rewinds it; no generic filesystem, URL, subprocess, IPC, or network capability is added. The completed YouTube path remains subject to canonical containment and point-in-time race limitations documented in doctoring.
• Focused tests cover parser/rewind failures, source duration/rate/channel rejection before librosa.load in all three consumers, decoded/model-output failures, empty-array shape invariance, and owned YouTube cleanup.

Exact-current-head verification

Local verification ran on exact head c2cc5bbeda6628fa9999401d6b0d228cb9b6bb9c:

• ./scripts/harness/quickcheck.sh — passed documentation/security/supply-chain/bootstrap checks, ESLint, ruff/format, Bandit, mypy, desktop 198/198 with 100% coverage, shared-types 20/20 with 100% coverage, Python 761 passed/24 skipped with 100.00% coverage, and Vite build.
• Focused source-policy suite — 108 passed, 1 skipped; targeted privacy/YouTube/policy regressions — 73 passed.
• git diff --check and all local harness checks passed. The macOS Demucs subprocess emitted an existing native worker segmentation-fault log during fallback coverage, but the harness completed successfully with exit 0; this is recorded as platform evidence, not claimed as a clean accelerator run.

Fresh hosted checks for this exact head are queued or still running, including CodeQL, OpenCode Review, Required Noema Review, PR Review Merge Scheduler, and close-empty workflow jobs. Queued, skipped-required, failed, stale-head, predecessor, draft-skipped, model-only, self/author, or administrative evidence is not passing evidence. No qualifying independent non-author last-push approval is present; the old Cursor CHANGES_REQUESTED review state is not transferred as a current-source finding, and the current review thread inventory has zero unresolved threads.

Remaining #781 / merge boundary

Keep Draft and unmerged until #865 integrates, this stack is reconstructed on the resulting protected develop, and one unchanged exact head has terminal-success required CI/security/SAST/SBOM/supply-chain/package/release/central-review gates, native/Rust and applicable CPU/GPU evidence, exact owned statement/branch coverage and docstrings/rustdoc, zero valid unresolved findings, and a qualifying independent approval. Remaining work includes explicit per-job CPU/GPU/VRAM admission budgets, cancellation/resource measurements, and realistic whole-product CPU/GPU numerical/performance parity evidence. Do not infer completion from local proof, MERGEABLE, Ready state, or auto-merge.

Repository ownership boundaries

• fix(security): establish canonical npm, PDF.js, Nanoid, and Undici baseline #783 owns the protected-base nanoid/pdfjs-dist/undici dependency-security baseline.
• fix(security): drop persisted credentials before dependency lifecycle code #894 owns dependency-lifecycle checkout credential hardening.
• test(quality): enforce Python branch coverage #861 owns repository-wide Python statement/branch coverage baseline integration where applicable.
• Central .github remains the authority for hosted review/coverage tooling; no local workaround or repeated unchanged dispatch replaces it.

Latest exact-head maintenance

Current head: 505a595. Protected base: develop@749511c3ad4000090048718f685c6bee6b3d2c25.

Merged current origin/develop through ordinary merge commit b4a3513; the only conflict was CHANGELOG.md and both canonical audio-policy and current develop entries were retained. Exact-code review confirmed source-container metadata preflight is already wired before temporal, stem-separation, and bass-transcription decode; no duplicate policy authority was added.

Fixed the remaining valid canonical-lane defect: ChordRecognizer now uses zero-element shape-invariant y.size, with regressions for (0, 2) and (2, 0) layouts.

Local exact-head verification: 783 Python tests passed and 24 numeric-parity tests were skipped by platform policy; production statements/branches/functions/lines are each 100%. Ruff check, Ruff format check, mypy, and Bandit passed. No dependency, lockfile, workflow, model, network, filesystem, IPC, or reviewer-credential authority changed beyond the ordinary develop merge and this focused chord guard.

Keep Draft and do not transfer predecessor-head evidence.

[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 Take the canonical #781 owner lane on the existing fix/audio-resource-policy-781 branch only. Fresh exact head 223dd78126deeb3f12a68dc140f6a83fbe422225, protected base develop@acdbea6344fe1231c39535b575f4de35e4c607c9, logical predecessor #865 f86e266b2ab2dc5a95e6b4a484e777b29f0feeaf. Do not create a competing PR, rebase/force-push, touch foreign repos, or suppress #783 dependency findings.

First repair the exact current-head CI blocker with repository-pinned tooling, not guessed formatting: CI run 32336903836, job 96328111793, actual checkout 92c5d3db777cbea11bd73f5f3e7dccf5482cb4cf passed docs/security/supply-chain/desktop lint and then failed first at services/analysis-engine/src/bandscope_analysis/__init__.py:3:1 Ruff I001. Run the pinned Ruff fixer/checker for that file, preserve the privacy-filter import-before-API semantics, and commit the smallest formatter-equivalent repair.

Then, on the resulting exact head, preserve the unique non-duplicative #781 evidence currently stranded in competing PR #985 (feat/canonical-audio-resource-policy-781@d2cf2047af790cddf02b3957856d246638a754b5) by integrating it into this canonical lane with TDD rather than merging/cherry-picking #985 wholesale:

• fix(audio): establish canonical local-audio resource policy #866 already owns cross-boundary native desktop-core + desktop bridge + service admission + temporal/separation + YouTube resource policy. Keep that authority and its one-sample-over decode probe, payload-safe errors, device boundary, and existing tests.
• Add source-container metadata admission before any librosa.load(... sr=..., mono=True, duration=...) transform can hide the original source duration/sample-rate/channel count. Use one canonical policy-owned source-rate/channel contract, not helper-local constants; the feat(analysis): enforce one canonical audio resource policy (#781) #985 candidate evidence is 8 kHz–192 kHz and mono/stereo. Validate exact product compatibility before adopting those bounds. The preflight must inspect the already-open handle without PCM decode, rewind it for the decoder, fail closed/payload-free on malformed metadata, and keep the existing post-decode validation.
• Add realistic REDs proving >15-minute source metadata, unsupported original source rate, and >2 original channels are rejected before librosa.load in every local decoder path that transforms these properties (at minimum TemporalAnalyzer and AudioStemSeparator; include bass transcription if its bytes decoder has the same bypass). Preserve encoded-byte checks before metadata parsing.
• Preserve feat(analysis): enforce one canonical audio resource policy (#781) #985's shape-invariant chord empty-buffer regression: zero-element arrays such as (2, 0) must return no chords rather than bypass the intended empty-audio behavior merely because len(y) != 0; non-empty malformed layouts must still fail closed under the canonical policy.
• Reconcile stale reason-code tests against this branch's published cross-boundary error contract; do not replace fix(audio): establish canonical local-audio resource policy #866's established policy/error architecture merely to match feat(analysis): enforce one canonical audio resource policy (#781) #985.

Run focused RED→GREEN tests, repository-pinned Ruff check/format, Bandit, mypy where applicable, then canonical quickcheck. If a finding belongs to #865 or #783 rather than this exact branch, prove the first causal boundary and leave it with its owner rather than adding a leaf workaround. Commit only to this branch and report resulting exact head and evidence.

[seonghobae]

@OpenCode repair exact head 223dd78126deeb3f12a68dc140f6a83fbe422225 only if fix/audio-resource-policy-781 is still unchanged. CI run 32336903836, job 96328111793, actual merge checkout 92c5d3db777cbea11bd73f5f3e7dccf5482cb4cf fails first at services/analysis-engine/src/bandscope_analysis/__init__.py:3 with Ruff I001 (uv run ruff check src tests). Run the repository-pinned Ruff/isort fixer on that file rather than guessing import order, inspect the resulting diff, and commit only the formatter-equivalent import-block change; do not alter the diagnostic privacy behavior, dependency/security policy, gates, lockfiles, stack ancestry, or other files. Then run focused uv run ruff check src/bandscope_analysis/__init__.py and uv run ruff format --check src/bandscope_analysis/__init__.py, followed by the normal exact-head CI. If the head moved, inspect the intervening delta and do not race the writer.

[seonghobae]

@opencode-agent Please review the current exact PR head c2cc5bbeda6628fa9999401d6b0d228cb9b6bb9c. Publish only a current-head APPROVED or CHANGES_REQUESTED verdict; do not rely on predecessor evidence or a provider-unavailable fallback. The branch remains subject to required Checks, unresolved-thread, independent-approval, and protected-merge rules.

[seonghobae]

Exact-head maintenance update for 505a595:

• Ordinary-merged current origin/develop as b4a3513; CHANGELOG.md was the only conflict.
• Fixed y.size zero-element chord handling for empty layouts (0, 2) and (2, 0).
• Full local verification: 783 passed, 24 numeric-parity tests skipped by platform policy, 100% statements/branches/functions/lines; Ruff, mypy, and Bandit passed.
• Existing source metadata preflight remains wired before all three local decoder paths; no duplicate policy authority added.

Keep Draft; predecessor evidence does not transfer.

[seonghobae]

@opencode-agent review\n\nReview only current PR head 505a595 against protected develop base 749511c. Revalidate the canonical local-audio resource policy, source metadata preflight before decode, post-decode limits, empty-layout chord handling, payload-safe diagnostics, exact tests, and current security checks. Do not reuse predecessor-head evidence or provider-unavailable results.

[seonghobae]

Queued @opencode-agent for PR #866 at head 505a595d481f8ba03abd8d13e7c17202918c833f. Central exact-name Actions artifacts are the durable dispatch ledger; existing review workflows remain authoritative for the final verdict and failure evidence.

[seonghobae]

Canonicalization note (2026-08-30): this draft covers the same Issue #781 audio-resource-policy Gap as #985. The maintained successor is #985, now carrying the shared encoded/decoded admission policy and the current fail-closed numeric-metadata fix at head 0e2486a33bb83a644f71fc0e7e57f090c7fcdb02. This PR remains unmerged; no protected-gate or approval evidence supports merging it independently.