Code snapshot with correctness fixes (fp16 fallback on T4, agent reset validation, collect/evaluate fixes); four-box wording and Stage-4 validation note in the card
Release the on-policy finetuned SMWM-c3 (step 60k) as jepa/smwm-c3/model.pt, keep the 145k model as pretrain.pt, add the 300-scene benchmark and a new code snapshot