Fix: import sys as _sys2 in inserted block (sys unbound in finetune main)

Fix: _ilu2.module_from_spec (importlib name unbound in patch context)

v2: warm-start LM via peft's own loader (_load_language_model — fixes 730-missing-keys no-op); data fix retained

R3 emotion-LoRA warm-start wrapper: prep mixed data, fetch fullv3 as init, warm-start patch, 800-step train, push emo3/