card: mark as the production model, add gen4 row, unified-naming note

Add PyTorch best.pt + data.yaml + train-args.yaml (consolidate from cards_day_2)

Publish gen3 (current) weights + standardized card