…ew (#202)

Review of #201 found the committed gates had never been run: 'asa verify'
was red on main (3 failing default-manifest checks) and the synthetic
manifest failed 52 of 140 checks. This commit makes both gates green with
honest measurements and records the first accuracy baseline.

Generator (build_synthetic_corpus.py):
- Engineered band-disjoint one-shots replace product drum samples in
  fixtures: the kick's pitch-sweep onset/click landed in the snare band
  (16 kicks counted as snares) and product hats never cleared the hat
  detector's floor. Kick = 50 Hz sine, 20 ms fade-in (sharper attacks
  splatter into the snare band), 0.20 s length (its low-band tail
  otherwise turns -64 dB snare residue into countable local maxima);
  snare/hat = band-passed bursts with an extra HP pass.
- Grid clips are a bare accented-kick pulse: 'and' hats bin ambiguously
  in analyze_time_signature's ±half-beat onset counting and flipped 4/4
  reads to 3/4 or 6/8. Swing clips keep swung hats but check BPM only.
- Chord labels computed from pitch classes, flat-spelled to match the
  Viterbi vocab (hardcoded sharp table scored 0.25 on F# major).
- Bass fixture gets --transcribe (was missing → guaranteed noteF1 0).
- expected blocks carry active checks only; ground-truth-for-later
  (swing %, hit times) moves to a separate 'truth' key.
- Dropped the [:28] truncation that silently discarded multi_F_major;
  one-shots carry their own seeds (order-independent determinism).

Harness (fundamentals_evaluation.py):
- Enharmonic pitch-class folding for key + chord comparisons
  ('C# Minor' == 'Db minor', 'Gb' == 'F#'); relative-key table replaced
  with pitch-class math.
- knownGaps: per-track check names that still run and report scores but
  don't gate — measured baseline weaknesses stay visible as improvement
  targets (8 at baseline: odd-meter detection + downstream downbeats,
  tempo octave at 174, 7/8 tempo wobble).

Trust layer (fundamentals_quality.py):
- bpmAgreement=True with confidence >= 0.5 is authoritative: a BPM
  measured exactly right and cross-confirmed by Percival was being
  hedged as 'cross-check not strong enough'.

CI: the default-manifest fundamentals gate (render + --fail-on-skip) now
runs in the backend job — #201 merged green because CI never executed
the gate it shipped.

Baseline: audits/accuracy-baseline-2026-07-03.md — 29 clips, 54/54
active checks green, 8 knownGaps quantified. Backend suite: 1283 tests OK.


Claude-Session: https://claude.ai/code/session_01T4wfz87k6kzJqkKLKZE7YL

Co-authored-by: Claude <noreply@anthropic.com>