Research v0.8 metrical correction and balanced fine-tuning

[bleugreen]

Summary

• instruments batch and single-file inference with optional per-clip/aggregate top-k outputs and metrical candidate masses while preserving legacy predictions
• adds canonical artist/duplicate-aware GiantSteps, GTZAN, and Ballroom partitions with checksums, provenance, global fingerprint disjointness, and immutable v0.7 selections
• implements correction baselines, onset-periodicity features, calibration/taxonomy reporting, and held-out-safe experiment tooling
• recovers and modernizes training with manifest-keyed HCQM caching, balanced sampling, time-stretch primitives, track-level validation aggregation, and a v0.8 research checkpoint
• commits the complete research report, machine-readable results, and final baseline/tuned prediction sidecars

Measured result

At 4% tolerance on the new canonical test folds:

Dataset	v0.7 Acc1	Fine-tuned Acc1
GiantSteps	70.99%	92.37%
GTZAN	69.01%	72.20%
Ballroom	59.70%	88.81%

The candidate fixes the canonical slow-octave bias, but it is not a general replacement for v0.7. On a separate 578-track strict three-estimator consensus check, v0.7 scored 95.50% and the candidate scored 93.60%; the non-Bass slice fell from 95.29% to 90.59%. v0.7 therefore remains the package default.

Grouped pseudo-label fine-tuning and v0.7/v0.8 selectors were rejected. The branch now includes a hierarchical DeepRhythm architecture that separates a 60–120 BPM base-rate head from a four-class metrical-level head, plus manifest-driven audio-domain stretching and posterior-preservation training support. The hierarchical experiment showed that 13/14 broad-validation failures had the correct base rate but wrong level. Level-only replay preservation reached 87.02% GiantSteps, 69.33% GTZAN, and 91.04% Ballroom, while scoring 91.67% versus v0.7’s 93.75% on the 48-track broad test. That is 64 additional canonical tracks at a cost of one broad-test track. Its weights are not shipped because the replay corpus is private and it misses the no-regression gate; the report names a licensed, manifestable replay corpus as the remaining requirement.

Verification

• 113 tests pass, including hierarchical tempo composition, canonical fold-to-cache integration, and joint domain/tempo sampling
• Ruff is clean
• canonical partition validation passes
• committed result tests verify test-manifest and weights hashes
• final prediction/details sidecars are archived under docs/research/results/sidecars