Fix VAE-aligned chunking in full-song inference

[zonnart17-collab]

Summary

Fix cumulative timing drift and padded output tails in both SonicMaster full-song inference paths.

• centralize the published checkpoint geometry in fullsong_chunking.py;
• use the loaded VAE's actual hop length with the trained 645-frame main sequence and 215-frame carry sequence;
• apply the same geometry to infer_single.py and inference_fullsong.py;
• keep duration conditioning fixed at the trained value of 30 seconds;
• validate VAE sample rate, model sequence length, main encode, decode, and carry geometry at runtime;
• trim the final padded tail to the exact input length without allowing trim to hide stitched underflow.

infer_single.py no longer exposes --fs, --chunk_duration, or --overlap_duration. Input audio is resampled to the published checkpoint sample rate of 44.1 kHz.

Background

The previous 30-second waveform chunk contained 1,323,000 samples at 44.1 kHz, but the Oobleck VAE encodes it to 645 frames and decodes those frames to 1,320,960 samples.

That 2,040-sample discrepancy accumulates at every stitched chunk boundary. The previous nominal 10-second overlap also used 441,000 samples instead of the VAE-native 215 × 2048 = 440,320 samples.

The corrected geometry is:

• main: 645 × VAE hop = 1,320,960 samples;
• carry: 215 × VAE hop = 440,320 samples;
• stride: 880,640 samples;
• trained duration conditioning: 30 seconds.

Validation

• py_compile: passed
• tests/test_fullsong_chunking.py: 11 tests passed
• full-song multi-chunk inference preserves the exact resampled input length
• main/carry geometry is 645/215 frames with zero cumulative boundary drift

No changes are made to model.py, model configuration, training, preencoding, latent batch inference, dataset generation, or evaluation code.