Fix adapter binding: load via AutoModelForImageTextToText (Qwen3.5 is multimodal); add lora_B bind guard

Add /diagnose: adapter-on vs adapter-off control to verify the fine-tune is active

Add thinking toggle + stable /generate API endpoint for evaluation

Fix validator crash on bare-string responses; harden run() so output always shows

Fix ZeroGPU: load adapter on CPU, defer CUDA placement; dtype rename

Add spaces dependency for ZeroGPU; make missing-ZeroGPU fallback loud

Use explicit CUDA placement instead of device_map=auto for ZeroGPU compatibility

Add commerce-ops agent demo: two trained tasks + guardrail validation