Guard resample overshoot before 16-bit save (stop silent clipping)

[wilke0818]

Summary

The audio-copy sanitize path (downmix → 16 kHz resample → 16-bit save) let torchaudio silently hard-clamp any sample the resampler pushed past full-scale. That's destructive (flattens peaks) and raises no error. On a source that wasn't already clipped it manufactures clipping — e.g. diadochokinesis-(v2)-tuh: source peak 0.998 → resample 1.046 → saved 1.000 with clipping introduced.

Measured on the adult 07_01 build: resample overshoot fired on ~18% of recordings (all previously silently clamped); 14 of those had clean (un-clipped) sources, so those clamps are unambiguously wrong.

Fix

_guard_resample_overshoot(resampled_audio, in_peak) in dataset.py: when the resampled peak exceeds 1.0, rescale the waveform down to the input peak so no sample exceeds [-1, 1] — preserving gain and waveform shape, introducing no clipping — and log a warning. No-op when the resampled peak is already in range, so non-overshooting audio is byte-identical to before. Wired into _copy_audio_files_parallel right before save_to_file.

Rescale-to-input-peak (rather than clamp, or rescale-to-1.0) keeps every file at its own source level whether or not it overshot; overshoot only occurs on near-full-scale material, so the rescale factor is always small.

Verification

• New tests/test_resample_clip_guard.py (2 tests): overshoot → rescaled to input peak with a real 16-bit round-trip showing 0% clipping; in-range → no-op.
• Real case confirmed end-to-end: diadochokinesis-(v2)-tuh old-path saved peak 1.000 with clipping, new-path saves 0.998 with 0% clipping.

Independent of the language-aware work (#326); touches only the audio-copy path.

🤖 Generated with Claude Code

[wilke0818]

Addressed the code review (pushed in 3e2ca89):

• Carry Audio.metadata through the rescale — it was dropped only on the overshoot branch (asymmetric with the no-op branch).
• No-op on zero-sample waveforms — don't call .max() on an empty tensor.

Kept by design:

• Global peak-renormalize (rescale the whole file to the input peak) rather than a local limiter/clamp — clamping is the destructive behavior this PR removes; renormalize preserves waveform shape. Only fires on overshoot (~18% of recordings), and overshoot only occurs on near-full-scale material so the gain change is small.
• The ≤1-ULP floating-point residue after rescaling an already-full-scale source is below the 16-bit quantization step, so it saves as the intended full-scale sample — no audible/analytic clamp.

Verified against the adult 07_01 build: overshoot fired on ~18% of recordings (all previously silently clamped); 14 had clean sources (the unambiguously-wrong clamps).