v1.0.0

[danigb]

[1.0.0] - 2026-05-30

Parity-with-the-reference release. Parity with the Python
beat_this reference is now verified by a
committed golden test (tests/python_parity.rs), not just argued by construction:
F-measure == 1.0 (standard FP32 model) and ≥ 0.99 (small model) at the ±70 ms MIR
window for both beats and downbeats.

Added

• Golden parity test against the Python reference (tests/python_parity.rs +
  scripts/gen_golden.py), runnable on a fresh clone with the committed small model
• "Parity with the Python reference" section in the README documenting the verification
  and the remaining known divergence

Changed

• deduplicate_peaks keeps fractional merged-peak frame positions instead of
  rounding to an integer frame, matching the Python reference (removes a ≤10 ms
  divergence on merged adjacent peaks)
• beat_counts now ports the reference's infer_beat_numbers: pickup-measure
  (anacrusis) beats are numbered so they lead into the first downbeat, matching the
  Python .beats/JSON count column (beat/downbeat times are unchanged)
• The ort (ONNX Runtime) backend is now behind an off-by-default ort Cargo
  feature. The default build is pure-Rust rten only — no libonnxruntime needed.
  Build/test the ort backend (and its --runtime ort, cross-runtime parity tests, and
  op-level --profile) with --features ort.

Notes

• Known remaining divergence: the resampler (rubato sinc vs Python soxr) differs
  sub-perceptually for inputs not already at 22050 Hz; inputs at 22050 Hz resample
  exactly. Decode precision (f32 + symphonia vs float64 + torchaudio) differs negligibly.
• Post-processing is "minimal" only — the optional --dbn path (madmom DBN) is
  intentionally not implemented; the model is designed to be accurate without it, and
  Python's default is also "minimal", so default-vs-default output matches.
• The "identical timestamps" claim is scoped to the two Rust backends (rten vs ort),
  verified by tests/cross_runtime.rs.