Fix: skip progress callback under ZeroGPU (step_callback closure not picklable across worker process)

Fix: lazy SAM2 construction inside @spaces.GPU (avoid poisoning parent process CUDA state)

Fallback: use native demo.launch() under ZeroGPU (custom uvicorn.run() causes silent shutdown)

Fix: disable Gradio SSR companion node server (port conflict)

Fix: temporarily unpatch ZeroGPU torch emulation during SAM2 Hydra construction

Fix: build SAM2 on CPU then .to(device) - Hydra construction fails under ZeroGPU CUDA emulation

Fix: move HF_HOME writability check before gradio import

Fix: HF_HOME fallback for writable modules cache (Florence-2 trust_remote_code)

Deploy FastAPI+WebSocket+Gradio microservice (Etapes 1-6)