From 7619cf6c17b7b6cb8cf6dba034c2616f29ccde4c Mon Sep 17 00:00:00 2001 From: Lucas Winther Date: Fri, 30 Oct 2026 05:55:19 +0100 Subject: [PATCH] chore: untrack SDD workspace report Co-Authored-By: Claude Opus 5.5 --- .../final-fix-report.md | 13 ------------- 1 file changed, 13 deletions(-) delete mode 100644 .superpowers/sdd/2026-10-30-synthetic-benchmark/final-fix-report.md diff --git a/.superpowers/sdd/2026-10-30-synthetic-benchmark/final-fix-report.md b/.superpowers/sdd/2026-10-30-synthetic-benchmark/final-fix-report.md deleted file mode 100644 index 09ed28b..0000000 --- a/.superpowers/sdd/2026-10-30-synthetic-benchmark/final-fix-report.md +++ /dev/null @@ -1,13 +0,0 @@ -# Final-review fixes report - -Verification: `bun test --timeout 300000` -> 89 pass, 1 skip, 0 fail (90 tests, 19 files). `bun run typecheck` -> clean. Each fix was test-first (quantize and dataset tests seen red before the fix; report/score tests written before the implementation). - -1. Per-mode tick (src/bench/report.ts): uses median merged-guitar F1 >= 0.85 and appends "k/n pass". Test "per-mode tick uses the median rule with a k/n pass count" (4/5 pass, median passes -> "✓ 4/5"; 1/5 -> "✗ 1/5"). -2. R12 (src/pipeline/quantize.ts): chord duration = median of members' snapped durations (even count: rounded mean of the middle two), min 1 step. Tests: updated chord test (8,4,8 steps -> 8) plus a 70 ms ghost-note case that keeps a 16-step chord. -3. R11: SongScore gains `gpPitch` (MatchResult from truthFromGp(outGp), onsets/offsets + grid.firstDownbeat) and `barAgreement` {refBars,outBars,timeSigMatch}; scoreSong takes optional `outGp`; runner passes res.gp. truthFromGp now also returns `bars` (masterBars count). Report columns: "GP-file guitar F1", "Bar count match", "Time sig match", "Techniques" (= "n/a (tier 1 has no technique marks)"). Tests in tests/eval/score.test.ts (downbeat shift, null without GP, bar-count mismatch) and tests/bench/report.test.ts. -4. results/.json: new `resultsJson()` in report.ts drops `matches`; runner uses it (score.json untouched). Test "results JSON omits per-note matches". -5. `resample` moved to src/dsp/resample.ts, re-exported from src/audio/wav.ts; transcribe.ts imports src/dsp/mono.ts and src/dsp/resample.ts. Grep for "audio/wav" under src/pipeline, dsp, score, gen, eval finds only comments. -6. src/bench/dataset.ts: `RENDER_VERSION = 1` (bump comment), written to `/render-version` before song.json; buildSong rebuilds if missing/different. Test: old and missing marker both trigger a rebuild (mp3 mtime advances). -7. README caveat on oracle-stems; models/demucs/README.md with provenance. - -Note: the report column count grew (3 extra metric columns + Techniques); committed results/ baseline files were not regenerated (controller re-runs the benchmark).