desktop: persisted waveform-peaks cache + async load path

[robmorgan]

Summary

Track loads no longer freeze the UI, and analyzed tracks load near-instantly on relaunch.

• .tspeaks binary sidecar persists the 3-band peaks pyramid's base level (u8-quantized, versioned 60-byte header). Keyed by the same mono-signal FNV hash as the .tsanalysis.json sidecar, so renamed and retagged files keep their cache. Corrupt/truncated/mismatched files silently recompute and rewrite; writes are temp+rename atomic. Upper pyramid levels rebuild from the cached base via the existing halving loop.
• Async load path: decode, peaks, and pre-analysis all run on a background worker; results install through an mpsc channel polled each frame, guarded by the existing analysis_generation counter (rapid successive loads discard stale results). The file panel shows a spinner while decoding.
• Grid unification: the UI beat grid now derives from the pre-analysis artifact (grid_from_artifact), deleting the duplicate synchronous detect_beat_grid_buffer pass — beat detection runs once per load, and BPM appears instantly on a sidecar hit.

Zero new dependencies; no library changes.

Testing

• 21 new unit tests: cache corruption matrix (magic/version/rate/hash/length/truncation/trailing garbage), quantization roundtrip, atomic-write temp cleanup, grid adapter, and headless load-worker tests against a hand-written WAV fixture (message ordering, cache hits, decode errors, stale generations).
• cargo test --all-targets (68 passed), cargo clippy --all-targets -- -D warnings, cargo fmt --check — all clean in desktop/.
• Live app verification on macOS: cold load computes and writes the sidecar (10,860 bytes for the 12 s test track — exactly header + 6×1800 buckets); relaunch logs Peaks: using cached sidecar; playback runs with zero underruns; a truncated sidecar triggers silent recompute and rewrite.

🤖 Generated with Claude Code