3-tier resilience: GPU retry -> HF Inference -> deterministic; add torchvision for VLM

Add torchvision (Qwen2-VL video sub-processor dep); switch default reasoning model to E4B (works on ZeroGPU CUDA emulation)

Pin kernels<0.15 to fix transformers hub_kernels LayerRepository error

Initial deploy: Document Integrity Verifier (LEGX)