revert: Let Unsloth handle base model loading directly

fix: Explicit snapshot_download base model to /data/models/

fix: Use arka2696/gemma-4-31B-base for base model (uploaded from HPC)

fix: Let Unsloth handle base model download via HF_HOME cache

fix: Correct base model ID to google/gemma-4-31B

fix: Hardcode base model ID + clean incomplete downloads

fix: Check for .safetensors files + pass HF_TOKEN for gated model

fix: Handle already-patched adapter_config with local path

fix: Cache base model in /data/models/ bucket — stop re-downloading

fix: Thread-safe model loading — prevent duplicate parallel loads

fix: Fast startup for HF Space — background model load + persistent cache

feat: Add instant demo gallery + Dockerfile for HF Space

docs: Complete DEPLOY.md with lessons learned from deployment

docs: Complete HF Space README with full project details

fix: Monkey-patch gradio_client schema introspection bug

fix: Use gr.Interface with show_api=False, add h5ad upload

fix: Replace Chatbot with Blocks layout to avoid Gradio schema bug

fix: Use tuple chat format for Gradio ZeroGPU compatibility

fix: Lazy-load models inside @spaces.GPU for ZeroGPU compat

fix: Remove gradio/spaces from requirements, pin huggingface_hub<1.0

fix: Pin gradio>=5.6.0 for huggingface_hub compatibility

fix: Vendor Phoenix flow_simple.py, remove phoenix package dependency

feat: Add HuggingFace ZeroGPU Space deployment