Apples-to-apples dataset replication (workshop sprint)

Replicate the finished MedQA (text) + NIH ChestX-ray14 (imaging) standard battery on every remaining dataset, so the shortcut-cascade result is shown consistent across datasets (apples-to-apples). Priority 1 for the MICCAI workshop submission (deadline 31 Jul). Priority 2 = extra experiments; priority 3 = infra/cosmetic. Golden rule: real data + real API only, results under experiments//results/.