Re-eval with max_new_tokens=256 fix; n=5 per condition

SFT prime (1 epochs, lora r=16) on teacher trajectories

initial commit