fix: pad mel with spec_min instead of 0.0 in acoustic collater

[KakaruHayate]

Problem

The acoustic collater pads mel with raw log-mel 0.0, which is not a neutral value: norm_spec maps it to +1.0 — the top of the normalized range, i.e. maximum loudness. Every batch fills the padded tail of shorter samples with full-loudness garbage.

Two consequences:

1. Backbone receptive-field leakage. The diffusion backbones (WaveNet / LYNXNet / LYNXNet2) receive no padding mask. LYNXNet2's receptive field is ~181 frames (kernel_size=31 × 6 layers), so the fake full-loudness signal leaks into the trailing valid frames of every non-longest sample. The loss mask (mel2ph > 0) only hides padding frames themselves — the contaminated valid frames near the boundary are fully counted. This is a systematic bias on utterance tails, exactly where breathy endings and vibrato decay live.

2. Aux decoder trained on garbage. The aux decoder (shallow diffusion) loss is not masked at all: with zero-padding it was actively trained to predict maximum loudness on padding frames from near-zero condition.

Measured on an untrained LYNXNet2 (1024ch × 6L): valid-frame output shift up to ~1.3% within the receptive field of the padding boundary, zero beyond it.

Fix

Pad with spec_min (-12 by default) instead. It maps to -1.0 (silence) after normalization and sits next to the mel extractor's true silence floor log(1e-5) = -11.51, so padded regions now look like ordinary trailing silence — consistent with what the model sees at inference.

One-line behavioral change + comment; falls back to -12.0 if spec_min is absent from config.

Compatibility

• Checkpoint-compatible: no parameter/shape changes.
• Training data distribution changes slightly (padded tails now silence instead of max loudness). Models trained before/after differ only in padded-region behavior; fine-tuning across this change is safe.
• Inference is unaffected (no padding at inference).