Improve beat acquisition and genre accuracy behind expanded corpus

[brxs]

Summary

Closes #77.

This improves the Rust beat estimator's acquisition speed and genre coverage while preserving its blank-when-unsure contract.

• locks a versioned 20-clip / 534-second corpus through Git LFS, with deterministic regeneration metadata, SHA-256 hashes, coverage validation, and a model-free verifier
• records the unchanged shipping baseline and owner-approved targets before changing production code
• ships the corpus-selected adaptive detector: complementary six-second band-flux and spectral-flux views, impulsiveness-based arbitration, and a two-second change probe with corroboration and recovery quarantine
• applies the same detector to live streams and decoded-track analysis without moving work onto the realtime audio callback or changing the published analysis surface
• documents the decision in ADR-0035, the rejected candidates, final per-clip measurements, clean-checkout reproduction, and completed hardware validation

Measured result

Metric	Unchanged estimator	This PR	Approved gate
Worst required rhythmic acquisition	20 s or never	9 s	<= 10 s
House -> dub recovery	16 s	8 s	<= 8 s
Dub -> house recovery	15 s	1 s	<= 8 s
Wrong steady / intro / beatless seconds	0	0	0
Wrong tempo-change seconds	1 / 3	1 / 0	<= 1 each

The final harness processes all 534 seconds in about 1.9 seconds (roughly 280x realtime on the validation machine). Every approved profile passes, including breakbeat/jungle, swung/2-step, sparse/minimal percussion, delayed rhythmic intros, opposed tempo changes, and the legacy beatless controls.

Implementation

• Main evidence:
  • six-second band-split log-energy flux
  • six-second spectral flux from a reused 2048-frame FFT buffer
  • spectral priority only when onset impulsiveness clears the measured threshold
  • band fallback only when spectral evidence is absent or metrically agrees
• Change handling:
  • a two-second spectral probe never displays directly
  • two mutually consistent contradictions are required while the main detector still confidently holds the old metrical clock
  • the first contradiction consumes the permitted one stale tick; the second blanks and quarantines the stale long-window result
  • reacquisition resumes only after the main detector and a confident probe agree
• Gate agreement now uses the corpus's existing binary/ternary metrical levels within 8%, while preserving three-stable acquisition and one-miss grace.
• rustfft 6.4.1 is pinned directly rather than used transitively.
• FFT spectra and scratch buffers are reused; all estimator work remains on the existing non-realtime analysis thread.
• AnchorGate, phase-anchor input, IPC/store state, engine callback, reset ordering, and downstream fallback contracts remain unchanged.

There are no prompt-, filename-, or genre-specific production branches.

Corpus and review checkpoints

The commit history deliberately preserves the measurement-first workflow:

1. expand and lock the corpus
2. measure the unchanged shipping estimator
3. approve numeric targets
4. accept ADR-0035
5. ship the selected estimator and hard regression gates
6. document and complete native/hardware validation

The committed corpus contains 20 stereo PCM16 / 48 kHz WAVs (97.8 MiB): ten immutable legacy fixtures, six new genre fixtures, two short-intro scenarios, and two opposed tempo-change scenarios. Missing fixtures, LFS pointer files, hash drift, invalid WAV contracts, incomplete coverage, unapproved targets, or metric regressions fail explicitly.

Reproduce it with:

backend/.venv/bin/python backend/scripts/spike_beat_corpus.py --verify
cd src-tauri
cargo test analysis::beat_corpus::the_expanded_spike_corpus_meets_approved_targets -- --nocapture

Validation

• just check
  • Ruff format/check
  • ESLint and TypeScript
  • Clippy across the workspace with warnings denied
  • 176 backend tests
  • 572 frontend tests
  • 210 Rust shell tests (one timing diagnostic ignored)
  • 109 Rust engine tests
• git lfs fsck: pass
• model-free corpus verifier: 20 clips / 97.8 MiB pass
• fresh local clone + git lfs pull + frontend build + verifier + approved corpus test: pass
• native mrt2_small session on MacBook Pro Speakers at 48 kHz:
  • real house stream acquired 130.13 BPM at 0.881 confidence
  • 35-second ambient/no-drums stream remained blank
  • stop cleared BPM, confidence, and live beat immediately
• completed owner listening checklist: acquisition, both tempo-change directions, model-switch reset, phase meter, synced echo, beat-quantised loops/freeze pads, two-deck audio, and CPU behaviour all pass

Documentation

• docs/spike-beat-detection-issue-77.md — unchanged baseline, target derivation, candidate matrix, and final per-clip results
• docs/adr/0035-dual-envelope-beat-detection-with-fast-change-invalidation.md — selected architecture and tradeoffs
• docs/issue-77-hardware-checklist.md — native and owner listening evidence
• .plans/issue-77-beat-estimator-acquisition-and-genre-accuracy.md — completed plan and retrospective