feat(meter): loudness-accent stream for meter + downbeats on non-4/4 (PR-G4)

[slittycode]

What

PR-G4 of the genre-generalization program (stacked on #212): meter and downbeat detection learn to hear loudness accents, closing the measured odd-meter gaps. The onset-count accent stream that analyze_time_signature ran on is loudness-blind — a kick on every beat yields identical counts at every bar position no matter which beat is accented, which is exactly how the corpus odd-meter clips (and most real four-on-the-floor material in 3/4, 6/8, 7/8) are shaped. Measured before this PR: 3/4's best candidate edged 4/4 by 0.8% (dominance 1.067 vs 1.059, position means all ~1.0), nowhere near the 20% override margin, shipping 4/4 at confidence 0.03.

Design

1. Loudness-accent stream: the per-beat low-band (kick) loudness from beat_data — the same signal the downbeat phase heuristic already uses — folded at each candidate bar length, dominance maximized over bar phase, clamped to 10 (near-silent off-beat positions explode the ratio without carrying more information), neutral (1.0) when the fold has fewer than 4 bars or the caller has no beat_data.
2. Harmonic collapse: a 3-periodic accent folded at bar 6 scores high but describes the shorter bar twice. The discriminator is the echo position's prominence — how far the position at (peak + divisor) rises above the remaining positions — because the raw echo/peak ratio measures ~0.72 for BOTH true 3/4 and true 6/8 (baseline noise), while prominence measures 0.262 vs 0.000 on the corpus clips.
3. Decision: the shipped meter now picks the best count-dominance × loudness-dominance product, overriding 4/4 at a 15% margin when loudness contributed (calibrated: 6/8's true combined margin is ~17%) and the original 20% count-only margin otherwise — legacy callers and --fast (which pass no beat_data) are byte-identical.
4. Pipeline: beat_data extraction is hoisted above the meter call in analyze.py (one extraction, same object reused by rhythm detail/groove/sidechain as before).
5. Evidence surface: timeSignatureCandidates entries gain an additive loudnessDominance key, landed on both sides (schema doc, TS type, client parser — which would otherwise have silently dropped it, tripwire Preserve signal monitor playback during analysis and harden run flow #3 — and the full-payload fixture).

Measured burn-down (the PR's gate)

Clip	Before	After
grid_3_4_90	meter 4/4 ✗ (gap), downbeats F1 0.286 (gap)	meter 3/4 ✓, downbeats 0.875 ✓ — both now gating
grid_6_8_110	meter 4/4 ✗ (gap), downbeats 0.400 (gap)	meter 6/8 ✓, downbeats 0.875 ✓ — both now gating
grid_7_8_140	meter 4/4 ✗ (gap), downbeats 0.182 (gap)	meter 7/8 ✓, downbeats 0.875 ✓ — both now gating (tempo smear 142.6 stays a known gap)
all ten 4/4 rhythm clips (incl. 2-step, halftime, breakbeat broken kicks)	4/4	4/4 — no regressions, broken-kick clips at confidence 1.0

_KNOWN_GAPS_BY_ID shrinks from 5 entries / 11 checks to 4 entries / 9 checks; every remaining entry is tempo-family (octave halving at 174/190, halftime-174's 2:3 error, 7/8's smear) owned by the PR-G3 promotion gate, plus the PR-G5 shuffle gap.

Behavior change disclosure

This PR changes the shipped timeSignature (and therefore downbeats) on material with loudness-accented non-4/4 bars — that is its purpose. The pre-registered gate is the synthetic corpus: all 38 clips green, 113 active checks passed / 0 failed / 10 informational. What it does NOT change: any caller without beat_data (including --fast), the golden (passes without re-baseline — the pipeline reorder shifted no curated value), or 4/4 material with flat or broken-kick accents (measured: all ten 4/4 rhythm clips unchanged). Honest caveat: the 15% accent margin is calibrated on synthetic accents; real-audio meter accuracy has no staged corpus yet (GTZAN-Rhythm is beat/downbeat-annotated and can serve the downbeat half via the existing beat-eval harness when the beat_this gate work resumes).

Validation

1. Backend: full unittest discover — 1357 tests, OK (3 skipped) (exit code captured directly, not piped).
2. Synthetic eval: 38/38 clips, 113 active passed / 0 failed / 10 informational.
3. CI-mirror fundamentals gate (default manifest): 15/15, allPassed.
4. UI npm run verify: 51 smoke passed (one deep-equal unit test updated for the new candidate key).
5. New unit tests: 8 in test_time_signature_accent.py (dominance math, neutrality guards, cap, harmonic collapse both directions, no-beat_data fallback).

🤖 Generated with Claude Code