Trim prefetch to 1 pack, log model-pack failures, JSON-constrained cards, rotate fallback names

Add MODE switch: LOCAL (on-device, default) vs MODAL (GPU endpoints)

ZeroGPU: move models to CUDA inside @gpu calls, not in the main process

ZeroGPU: read models from module globals, not pickled args

Integrate ZeroGPU Space path with card-cache, sampling, and diversity fixes

finished base

Build Tabras card duel prototype