Restore max_new_tokens to 512 (4-bit gen is fast: ~25 tok/s on GPU)

Load Qwen2.5-Coder-7B in 4-bit (nf4) inside the GPU worker

ZeroGPU: load model on GPU inside @spaces.GPU (canonical), not at import

ZeroGPU: raise GPU budget 120->180s, cap max_new_tokens 512->256

ZeroGPU: force model.to(cuda) in fn (ignore stale is_available); no cuda at import

diag: log cuda availability + model device + gen timing; force model.to(cuda) in fn

ZeroGPU: keep model GPU-resident (canonical pattern)

Load the LLM once at startup instead of per ZeroGPU call

GDScript RAG assistant: app + corpus (index added later via Colab)