Replace corrupt jazz.00054.wav (6 refs) with jazz.00055_10s.wav (matches the rest of the dataset, all 10s segments)

Add 10 cross-modal JSONL splits