feat(eval): genre-generalization synthetic corpus + honesty checks (PR-G2)

[slittycode]

What

PR-G2 of the genre-generalization program (plans/genre-generalization-program.md, stacked on #210): expands the synthetic fundamentals corpus from 29 to 38 clips so the accuracy gates span broken-kick, halftime, shuffle, breakbeat, beatless, and 85–190 BPM material instead of only steady four-on-the-floor 4/4.

New clips (all deterministic, --check-verified)

1. twostep_132 — classic 2-step broken kick (kick 1 + "and" of 2, backbeat snares), straight 8th hats.
2. halftime_140 / halftime_174 — kick 1 / snare 3, 8th hats carrying the notated tempo (the tempo-octave trap).
3. breakbeat_136 — syncopated kick pattern (anticipated second kick + "and" of 3).
4. shuffle16_130_62 — four-on-floor kicks + 16th-grid shuffled hats at 62% (UKG shuffle).
5. ambient_beatless_70 — two slow A-minor triads, no percussion: the abstention clip.
6. grid_4_4_85, grid_4_4_150, grid_4_4_190 — tempo extremes.

New harness capability

fundamentals_evaluation.py gains an additive honesty check family (expected.honesty) asserting what the pipeline must NOT claim on evidence-free material: maxBpmConfidence, swingDetailAbsent, meterSources. Unit-tested in tests/test_fundamentals_evaluation.py.

Measured baseline (calibrated on this machine, 2026-07-13)

Passing at baseline — now permanently gating: 2-step (beatGrid F1 0.954, downbeats 0.875), breakbeat (0.954/0.875), halftime at 140 (139.9 BPM measured — the hat grid carries it), 85 and 150 BPM grids, and the beatless-ambient abstention checks (bpmConfidence 0.114 ≤ 0.4, swingDetail absent, meter on the assumed-4/4 fallback, key still exact).

Failing at baseline — registered as knownGaps (report-but-don't-gate), each with a named owner:

Clip	Check	Measured	Owner
grid_4_4_190	tempo / beatGrid / downbeats	94.9 BPM (octave-halved); grid degrades to 0.625/0.667 — worse than at 174, where the grid survived	PR-G3
halftime_174	tempo / beatGrid / downbeats	117.0 BPM — a 2:3 ratio error, not clean halving; grid follows (0.370/0.308)	PR-G3
shuffle16_130_62	swing:swingPercent	None — 16th-grid shuffle is invisible to the 8th-IOI-windowed swing measurement	PR-G5

Full-manifest run: 38/38 evaluated, 94 active checks passed, 0 failed, 15 informational.

What this PR does NOT do

1. No product-path code changes — analyze.py and friends untouched; no schema/type/golden changes.
2. CI's default 4-clip manifest is unchanged; the 38-clip synthetic manifest remains the local accuracy gate (same split as the 2026-07-03 baseline).
3. The knownGaps failures are documented, not fixed — PR-G3 (tempo octave) and PR-G5 (16th-grid swing) own them.

Validation

1. Backend: ./venv/bin/python -m unittest discover -s tests — 1340 tests, OK (3 skipped).
2. Synthetic corpus determinism: build_synthetic_corpus.py --check — every clip double-rendered byte-identical.
3. Full synthetic eval: 38/38, 94 passed / 0 failed / 15 informational.
4. CI-mirror fundamentals gate (default manifest): 15/15, allPassed.
5. UI: npm run verify — 51 smoke passed (re-run solo after an initial load-flaked run; zero UI code touched).

Everything validated here is synthetic-corpus truth; real-audio confirmation of the tempo-octave behavior lands with the PR-G3 GiantSteps tempo re-run (corpus already staged locally).

🤖 Generated with Claude Code

[slittycode]

Note on CI: the main CI workflow only triggers on PRs based on main, so this stacked PR shows only claude-review. Full validation ran locally (backend 1340 tests OK, synthetic eval 38/38 with 94/0 active checks, CI-mirror fundamentals gate 15/15, UI verify 51 smoke). After #210 merges, rebasing this branch onto main and pushing will arm the full CI run before merge.