audit: full bug + architecture review (2026-05-30)

[slittycode]

Summary

• Land the 2026-05-30 full bug + architecture review report (audits/full-review-2026-05-30.md) under the existing dated-audit convention.
• Generated by a 30-agent workflow: 7 bug-hunt + 7 architecture dimensions in parallel, each candidate bug independently verified by an adversarial skeptic agent against the real source.
• Totals: 14 confirmed bugs (5×P1, 6×P2, 3×P3), 2 refuted (with reasons), 39 architecture findings.

What it flags (themes worth fixing as units)

1. Staged-run lifecycle robustness — cancel/cleanup/terminalizers were hardened for the measurement stage but never extended to its peers, so pitch-note + MT3 orphan multi-GB subprocesses on cancel and interpretation can strand in running forever (feat: value-first backport — genre profiles, MixDoctor, broken-pipe fixes #1, Preserve signal monitor playback during analysis and harden run flow #3, [codex] Prepare hosted runtime foundation #6, [codex] Stabilize backend CI subprocess assertions #7).
2. Chain-of-custody integrity — parser silently drops 4 prompt-blessed reverb/vocal subfields (Add phase1 visual story v2 deck assets and build scripts #2), normalizeKey self-corrupts major→majoror and fires a false contradiction ERROR (Codex/backend upload hardening #11), extractBPMFromText only checks the first match (#3q #14), and the citation check validates existence-not-relevance.
3. Frontend resilience — no React error boundary, monitorAnalysisRun has no wall-clock timeout (fix: chunk Demucs inference and stop boot-time memory spikes #4), retry handlers swallow failures into unhandled rejections (Codex/mix chain order guidance #10).
4. DSP honesty — truePeak rounded to 1 decimal silently hides inter-sample overs (Harden backend uploads, add integration tests, remove dead fields #8); PLR subtracts dB LUFS from linear true-peak amplitude (Add harmony and structure lanes with backend upload hardening #9).
5. No executable contract guard — Phase 1 contract lives in three hand-maintained representations with no cross-app test; the rename tripwire bites precisely because nothing executable defends it.

Where it fits

The audits/ directory is the canonical home for dated review/audit reports (nightly-2026-05-19.md, phase2-recommendation-surface-2026-05-24.md, etc.). This is past-tense; not consumed by either app, not on the request path.

Test plan

• No code changes — markdown-only documentation.
• Verify the file renders cleanly on GitHub.

🤖 Generated with Claude Code

[slittycode]

Verdict: APPROVE (posted as COMMENT — can't self-approve)

Summary

Adds audits/full-review-2026-05-30.md — a 485-line dated audit report landing in the canonical audits/ directory documented in CLAUDE.md. No code changes. The technical claims are consistently sourced with file:line references, the methodology is transparent (adversarial skeptic pass, two refuted findings documented), and the severity taxonomy is applied consistently. Spot-checked the five most falsifiable claims against the codebase description and CLAUDE.md contracts; all held.

Findings

Worth considering:

Bugs #3 and #6 describe the same defect at the same location (server.py:2008, missing try/except around _execute_interpretation_attempt setup) and recommend the same fix (wrap in try/except → fail_interpretation_attempt). The distinction drawn is trigger-reachability (artifact-resolution failure for #3 vs. any pre-Gemini exception for #6), but the gap being closed is identical. This inflates the confirmed count by one (14→13 if collapsed) and means the "5×P1 + 6×P2" breakdown in the PR description is slightly off. Harmless for a research document, but worth collapsing in any follow-on tracker so the fix list doesn't spawn two tickets that close the same line of code.

Everything else checks out — the severity downgrades (P1→P2 for #8, P2→P3 for #14), the "not-real" calls in §4 (both refutations are substantive, not hand-waving), and the cross-cutting theme grouping are all well-reasoned and internally consistent.

Test results

N/A — markdown only.

Phase boundary check

Clean. The document describes potential phase boundary violations in the existing codebase (notably Bug #2, dropped reverb/vocal subfields breaking the citation allowlist) but does not itself introduce any code that crosses a phase boundary.

---

Generated by Claude Code