epidebug-mm-v2 (CPU reasoning seed)

Frozen MiniLM + tabular MLP + tiny CNN with reasoning heads: failure category, shuffled hypothesis ranking, reasoning-pair, weak rationale, follow-ups.

Trained on hardened distractors + expanded physics vignettes + AI4I harvest. Not a frontier reasoner; GPU LoRA on a larger LM is the scale-up path.

See local /workspace/epidebug-train/TRAINING_DONE.md for metrics and resume commands.