Re-encode LibriSpeech supplement from FLAC to WAV (PCM16, 16 kHz mono) for format uniformity

Fix per-speaker severity labels to match canonical mapping (TORGO + UA-Speech)

Canonicalize metadata across all shards:

Add LibriSpeech train-clean-100 'Normal' supplement (train+test, speaker-disjoint) to break the Normal=TORGO-only corpus-severity confound.

Add train/test splits (test=1500 stratified by severity)

Add held-out evaluation set (unseen speakers, severity metadata)