audio-verify

name	audio-verify
description	Verify warp/audio-engine behavior by offline-rendering the warp demo scenarios to WAVs and asserting beat alignment numerically with the audio-analyzer MCP. Use after changes to src/lib/beats/, src/demos/warp/, stretch-engine or tempo-track behavior, or when asked to verify audio output without listening.

audio-verify

Renders the five warp scenarios full-song through the OpenDAW offline engine and asserts beat alignment against expected times computed from the beat map. Replaces "needs human ears" with numbers. Requires: the dev server (HTTPS certs present), Playwright MCP, audio-analyzer MCP.

All thresholds below are calibrated from a full end-to-end run on 2026-06-10 (M-series Mac, Otherside.mp3 + 511-marker beat map).

Workflow

1. Start the dev server from the branch/worktree under test: npm run dev -- --port 5181 --host 127.0.0.1
2. Render each scenario (sequentially; measured ~15–30 s per full-song render on an M-series Mac — budget minutes only for slower machines): navigate Playwright to https://localhost:5181/audio-verify-debug.html?scenario=<s> for raw, varispeed, timestretch, grid-conform, grid-rigid. Poll the #verify-state element's data-verify-state attribute: setup → rendering → uploading → done. On error:<msg>: stop, report the message. WAVs land at .verify-output/verify-<scenario>.wav.
3. Expected times: node scripts/expected-beats.ts > /tmp/expected.json — JSON with projectBpm, gridTimes, fileTimes, fileTimesRigid, rigidClickTimes (render-relative seconds).
4. Analyze each WAV with audio-analyzer rhythm_analysis on two 20 s windows: [60, 80] s and [120, 140] s. Collect the detected beat lists. Window choice matters: both windows must be musically dense (the tracker needs stability ≥ ~0.8; Otherside's first ~30 s is sparse guitar and unusable) and away from divergence zero-crossings (file-vs-grid drift is NOT monotonic — it re-converges where the song crosses its own average tempo, near 200 s for this file; a window there cannot discriminate raw from locked).
5. Compare per window with the committed helper: python3 scripts/compare-beats.py /tmp/expected.json <list> "<detected beats>" → median/p90/max nearest-expected distance in ms.

How the metric behaves (read before judging numbers)

Nearest-expected distance saturates at half the inter-beat interval (~244 ms at 123 BPM). A misaligned render therefore reads as median ~120–244 ms — never seconds. Measured reference points:

• Locked scenarios measure median 30–46 ms (the onset-detection jitter floor on rendered audio is ~30–40 ms; sub-30 ms assertions are not achievable).
• Unaligned (random phase) measures median ~120–180 ms.
• The beat tracker follows the MUSIC's pulse, not metronome clicks mixed into the render — click-based assertions must compare against the click list explicitly and expect the tracker to sit on the music.

Assertions (per window, median nearest-distance)

Scenario	Compare against	Pass criteria
raw — sanity	fileTimes	≤ 60 ms (the render plays the file; if this fails the harness/render is broken: STOP)
raw — negative control	gridTimes	≥ 100 ms (the file does not sit on the grid; if raw "passes" the locked test the discriminator is broken: STOP)
varispeed	gridTimes	≤ 60 ms
timestretch	gridTimes	≤ 75 ms (WASM offline worker — see 2026-07-16 re-measurement)
grid-conform	fileTimes	≤ 60 ms (conformed grid + clicks + music all coincide on file times)
grid-rigid — placement sanity	fileTimesRigid	≤ 60 ms (music plays where the region was placed)
grid-rigid — negative control	rigidClickTimes	≥ 90 ms ([60,80] measured 92 ms on the WASM worker; [120,140] ≥ 100 ms)

Measured 2026-06-10 ([60,80] / [120,140] medians, ms): raw-vs-file 30/40, raw-vs-grid 180/122, varispeed 33/32, timestretch —/46, conform —/35, rigid-vs-fileRigid —/33, rigid-vs-clicks —/153. ("—" = [60,80] not yet measured for that scenario; the [120,140] margins suggest similar values, but the first run that fills those cells should not be surprised by small drift.)

Re-measured 2026-07-15 at SDK 0.0.159 (same windows): raw-vs-file 30/40, raw-vs-grid 174/118, varispeed 33/32, timestretch 43/68, conform 30/35, rigid-vs-fileRigid —/33, rigid-vs-clicks —/153; pitch ordering 0.983 > 0.953. timestretch [120,140] measures ~68 ms at 0.0.159 — over the nominal 60 ms line. This is NOT a harness/render-path artifact: renders from the legacy OfflineAudioContext path and the OfflineEngineRenderer path are byte-identical (same SHA-256), so the drift comes from the SDK's 0.0.159 Tape/PitchVoice changes shifting onset content at that window. Treat timestretch [120,140] medians up to ~70 ms as the current expected value; the [60,80] window remains the discriminating ≤60 ms assertion.

Re-measured 2026-07-16 on the wasm-engine-only branch (every scenario now renders on the WASM (Rust) offline worker — raw/varispeed/timestretch previously rendered on the TS offline worker). Same two windows ([60,80] / [120,140] medians, ms): raw-vs-file 30/40, raw-vs-grid 174/118, varispeed 33/32, timestretch 71/68, conform 30/35, rigid-vs-fileRigid 33/33, rigid-vs-clicks 92/153; pitch ordering 0.985 > 0.953. This is the WASM-worker recalibration, not a regression. All discriminations stay fully intact (every locked median sits far below the ~120 ms unaligned floor; varispeed at 33 ms proves the harness still measures tight alignment). Two cells shifted and warrant a note: (1) timestretch [60,80] rose 43 → 71 ms — the WASM Tape/ PitchVoice path smears onsets slightly more than the TS path did, uniformly across both windows (they now read ~68–71 ms together, where the TS worker read 43/68). Treat timestretch medians up to ~75 ms as the current WASM expected value on both windows. (2) rigid-vs-clicks [60,80] first-measured at 91.5 ms — nominally under the ≥100 ms negative-control line, but it still discriminates cleanly (the music sits 33 ms from fileTimesRigid vs 92 ms from the rigid click grid, a ~2.7× separation), so the control is not broken; the [120,140] window remains 153 ms.

Pitch (relative check): harmonic_analysis pitch-class distributions on [120, 140] s; Pearson-correlate each against raw's. Require corr(raw, timestretch) > corr(raw, varispeed) — timestretch preserves pitch, varispeed smears it. Measured: 0.987 vs 0.956 (2026-07-16 WASM re-measurement: 0.985 > 0.953). The margin is small because this window's detune is only ±50–85 cents; treat absolute values as informational and assert only the ordering.

Report a pass/fail table with the medians. Stop at the first failed scenario with the numbers collected so far.

Troubleshooting

• Page won't load / cert errors: dev server must be HTTPS (COOP/COEP); localhost-key.pem/localhost.pem must exist in the directory the server runs from. Check the port matches the URL.
• error:verify sink rejected upload: the middleware only exists in dev mode (apply: "serve"); a preview/production server has no /__verify. HTTP 413 = render exceeded the 150 MB cap.
• error:Transient detection returned fewer than two positions: timestretch needs at least two transients (the engine's minimum); the audio file is silent/featureless — wrong file or broken load.
• State stuck at rendering: check the browser console via Playwright; if

  5 min, capture console messages and report.

• Onset medians ~120 ms for ALL scenarios including raw-vs-file: the windows are probably in sparse/unstable material (check rhythm_analysis Stability, want ≥ 0.8) or the expected lists are stale — re-run node scripts/expected-beats.ts.