fix(distill): train from in-memory labels; never re-read the just-written file

[ronaldtse]

Summary

• Arabic tiny labeled all 11,790 srcs, then the post-labeling re-read of teacher_labels.jsonl inside the same container served a stale volume replica (2 visible pairs of 11,790 written; the persisted file is healthy at 27,324 valid lines across runs)
• the spec lacked labels_complete so no guard fired: the run trained on 2 pairs and saved garbage weights
• fix: keep the freshly generated labels in memory and train from them directly; the file is read only on resume runs; the torn-view guard (<50% pairs) is now unconditional; final volume commit after the label loop (the 3200-modulo could skip the tail)

Test plan

• CI green
• ara-diac-tiny relaunch logs 'trainable label pairs: ~11790' and trains