Prevent hard clipping across audio output backends

[brxs]

Summary

Fixes #78.

Replace destructive output hard clipping with a shared, no-boost peak-protection policy across the PyTorch, MLX, TensorRT, and TFLite paths.

• Preserve in-range audio unchanged.
• When decoded audio exceeds the selected PCM ceiling, attenuate the entire waveform by one gain factor so sample ratios are preserved.
• Reject non-finite waveforms instead of silently converting corrupted samples.
• Trim decoder padding before measuring the peak.
• Expose --peak-ceiling-dbfs across the CLIs, defaulting to 0 dBFS.
• Allow API callers to select output_peak_policy="raw" when they need unmodified floating-point output for downstream mastering.
• Route optimized CLI and Gradio WAV exports through the same conversion behavior.

This avoids boosting quiet generations while preventing out-of-range samples from being flattened at full scale.

Implementation

The common PyTorch path now centralizes peak detection, gain calculation, PCM16 conversion, and WAV serialization in stable_audio_3.audio_output. The optimized backends use small, self-contained equivalents so their standalone bootstrap bundles do not depend on the root package.

TensorRT

The decoder ONNX rewrite now removes the baked [-1, 1] clip, PCM scale, and integer cast, exposing sample-major FP32 audio_unbounded. Both eager and CUDA-graph inference can therefore validate and peak-protect the decoded waveform before PCM scaling or narrowing hides non-finite and extreme values.

The rewrite also publishes uniquely named external-data sidecars before atomically replacing the ONNX model. This avoids sidecar growth on repeated rewrites and prevents readers from observing a model that references incomplete weights.

Existing TensorRT engines with the legacy clipped pcm output remain usable and produce a compatibility warning. Engines must be rebuilt and republished to gain the new unbounded-output behavior.

Configuration and compatibility

• Default ceiling: 0 dBFS, matching the full PCM16 range without boosting.
• Lower ceilings can reserve mastering headroom, for example --peak-ceiling-dbfs -1.
• Sample rate remains derived from each model/backend configuration; peak protection does not hardcode 44.1 kHz.
• Standalone MLX, TFLite, and TensorRT bundle/import behavior is preserved.

Validation

• 70 focused tests passed across shared peak protection, CLI plumbing, and optimized backends.
• 9 TensorRT ONNX/ONNX Runtime rewrite tests passed, including non-finite/extreme-value preservation, repeated rewrites, and injected model-swap failure.
• 11 TFLite cross-platform compatibility tests passed.
• MLX and TFLite Gradio import/help smoke tests passed.
• Standalone MLX, TFLite, and TensorRT CLI import/help smoke tests passed.
• Targeted Ruff, formatting, byte-compilation, bootstrap syntax, and diff checks passed.

CUDA/TensorRT engine replay was not available on the development Mac, so GPU execution remains for upstream CI or maintainer validation.