V5 retrain: 840 examples with 20 compliance fixes (Q24/Q37 Arabic greeting fix)

Swap to 40-question evaluation mode (loads adapter + runs all 40 Qs)

Fix OOM: batch=1 accum=16, seq_len=1024 (70B needs conservative memory)

Hayat Meditron3-70B V5: Single A100-80GB QLoRA training (battle-tested config)

initial commit