PR-A5: accuracy benchmark loop — synthetic corpus + CI threshold gate#22

Merged

MacFall7 merged 11 commits into

mainfrom

claude/a5-benchmark-loop

Jul 10, 2026

Commits on Jul 10, 2026

test: add failing regression for silent 120-BPM tempo fallback
claude
committed
feat: tempo/beat evidence — honest unknowns + shared engine fallback chain
claude
committed
test: cover BPM confidence/candidates, meter, chains, and stretch/drum wiring
claude
committed
test: add failing regression for cross-part drum bleed duplicates
claude
committed
feat: calibrate drum MIDI — per-part velocity, bleed de-dup, tom split, config
claude
committed
test: cover per-part velocity, ghost survival, tom split, and config validation
claude
committed
feat: accuracy benchmark loop with a synthetic corpus and CI threshold gate
claude
committed
test: cover the benchmark harness, threshold enforcement, and provenance
claude
committed
ci: install test + web extras so the benchmark workflow can run pytest
claude
committed
ci: put repo root on sys.path (pytest pythonpath) so `benchmarks` imports
claude
committed
test: strip ANSI in the CLI help assertion (CI forces color)
claude
committed