fix: unify mask score key; fix compare error row column count

fix: make pin_memory() a no-op on ZeroGPU where CUDA pinning is unsupported

fix: pass explicit device to Sam3Processor; move find_stage tensors in move_to

fix: intercept torch.tensor calls with hardcoded device='cuda' in SAM3

Broad CPU redirect patch for SAM3 model build (multiple hardcoded device=cuda)

Patch SAM3 PositionEmbeddingSine: hardcoded device=cuda → cpu during init

Force CPU init for ZeroGPU: CUDA appears available but isn't allocated outside @spaces.GPU

Add opencv-python-headless (needed by SAM1 AMG __init__)

Fix CLIP package: openai/CLIP via git (not open-clip-torch)

Add psutil (undeclared SAM3 dep in sam3_video_predictor)

Add pycocotools (undeclared SAM3 dep via training module chain)

Add einops and pin setuptools<81 for SAM3 undeclared deps

Fix models __init__.py: remove stale sam_model import

Fix Space card: use zero-gpu hardware, fix colorTo to indigo

Day 14: HuggingFace Space deployment

Fix OOM on high-res images and CPU offload for compare tab

Day 12-13: Multi-model SAM benchmark sandbox

Day 11: SAM+CLIP zero-shot segmentation pipeline — masks + CLIP ranking working