fix: harden _run_classification to handle Gradio 4+ FileData Pydantic model

fix(input): set gr.Image type='filepath' for Gradio 4+ API file handling

fix(adapter): return dict mapping {'path': cam_img_path} for Gradio 4+ FileData serialization

fix(adapter): save Grad-CAM overlay to temp filepath for Gradio FileData compatibility

fix(app): add comprehensive image_input type handling for dictionary/path/PIL objects

fix(app): import spaces first and wrap demo.launch() inside main guard for ZeroGPU supervisor

fix(app): call demo.launch() without hardcoded server_port for ZeroGPU container socket compatibility

fix(app): place demo.queue() and demo.launch() inside if HAS_GRADIO block

fix(app): add demo.queue() and standard main guard launcher for Hugging Face Gradio supervisor

fix(app): launch demo unconditionally at root level to prevent container exit 0

fix: check HAS_ZEROGPU before @spaces.GPU to prevent connection reset on CPU hardware

Upload app.py with huggingface_hub

Upload app.py with huggingface_hub

Fix L1 gatekeeper ResNet-50 compatibility, integrate L1-L2-L3 pipeline and return Base64 GradCAMs

Hotfix: extract model_state_dict from checkpoint

Add Bucket upload and huggingface_hub

Track model weights with LFS