Tracked-beat corroboration: rigid grids adopt on syncopated material (Stage 10)

[robmorgan]

Fixes the rigid-grid non-adopter class (ROADMAP Stage 10) — the misaligned beat markers first observed on Hot Stuff in the desktop app.

Problem

MSBWY, Hot Stuff, and Somebody To Love fit exact BPMs but sat at phase_lock 0.109–0.181 against the 0.3 adoption gate, shipping wandering DP grids (intervals 475–569 ms on constant-tempo tracks). Rival-phase landscape mapping showed why: swung bassline energy under 150 Hz scores competing phases at scattered subdivision offsets (Hot Stuff peaks at 6–10/32 period, MSBWY at 25/32).

Both roadmap candidates prototyped first — and failed

• Slot exclusion (anti-phase ± tolerance): rivals aren't at one excludable slot; stacking zones until specific tracks pass is threshold-lowering in disguise. Only helped cold_heart (already an adopter).
• Onset-sharpness weighting (squared envelope): helped two tracks, made Hot Stuff worse (0.109 → 0.072) — disco bass is as punchy as the kick inside the kick band.

Both negative results are recorded in LEARNINGS.md.

What landed: corroboration by an independent estimator

A below-threshold fit adopts when ≥ 60% of the DP tracker's beats — full-band novelty + dynamic programming, independent of the kick-band phase circle — land on the rigid grid within 25 ms. Measured populations:

Case	phase_lock	agreement @25 ms	outcome
msbwy	0.112	0.98	adopts
hot_stuff	0.109	0.90	adopts
somebody	0.181	0.28	stays out (genuine estimator disagreement)
tempo-ramp control	0.77 (!)	0.25	correctly rejected
jittered-kicks control	0.99	0.61	adopts (via lock, as before)

The clusters sit at ≤ 0.28 and ≥ 0.90; the 0.6 threshold splits the gap with margin on both sides. The 0.3 lock gate and the smeared sanity floor are untouched.

Corpus results (qa/bpm_accuracy, vs saved baseline)

Track	beat F	mean offset	downbeat F
msbwy	0.952 → 1.000	−5.3 → −0.6 ms	0.0 → 1.0
hot_stuff	0.934 → 1.000	−9.8 → +0.1 ms	0.94 → 1.0
33rd Rate Revs X	1.0 → 1.0	−4.1 → −0.1 ms	0.0 → 1.0

13/16 rows byte-identical; zero regressions.

Tests

New unit fixtures: a syncopated train (kick + beater click + swung low-band stabs) that deflates the lock, adopts via corroboration, and lands the phase on the kicks; and a drifting-tracker control sweeping a full period that must not adopt. Existing ramp/jitter/degenerate tests unchanged and green.

PREANALYSIS_VERSION/MIN_COMPATIBLE_VERSION → 9/9 per the CLAUDE.md analysis version policy (beat positions change on the corroborated class; cached sidecars regenerate once).

🤖 Generated with Claude Code