0.3.0

[0.3.0] - 2026-03-11

Changed

• Simplified public API: internal modules (inference, mel, postprocessing, output) are now private
• Renamed core types: InferenceRuntime → Runtime, InferenceSession → Model, BeatInference → BeatPredictor, MelProcessor → MelExtractor, PostProcessor → PeakPicker
• Renamed methods: process → predict (BeatPredictor), process → extract (MelExtractor), process → decode (PeakPicker)
• Re-exported beat_counts and calculate_bpm from the crate root

Added

• BeatThis::from_models constructor for building from pre-loaded models
• BeatThis::beat_model_mut accessor for runtime-specific operations (e.g. ORT profiling)
• analyze_audio_timed method with per-stage timing via TimedAnalysis and AnalysisTiming

Fixed

• Audio chunk padding: correctly handle last chunk border trimming instead of always using full CHUNK_SIZE, fixing potential out-of-bounds access on short audio

Beat model weights (large / FP32)

Pre-converted FP32 beat model (beat_this.onnx, ~79 MB) for scripts/download-models.sh. SHA-256 in the .sha256 sidecar. The small model used by the test suite is committed to the repo; this asset is only needed for full-accuracy inference.