S2PD: Physically grounded augmentations and stable parameter updates for cross-domain few-shot learning