Fix: Pre-download models in Dockerfile to avoid runtime hang

Root cause analysis:
- preload_from_hub ignores HF_HOME (documented limitation)
- Previous approach caused permission conflicts → fallback to /tmp
- local_files_only=True + no models in /tmp = startup failure

Solution (best practice from research):
- Manually download models in Dockerfile to default ~/.cache location
- Download as 'user' to avoid permission issues
- Keep local_files_only=True (models already cached)
- Build logs show download progress (transparent to user)

Changes:
- README.md: Remove preload_from_hub (has limitations)
- Dockerfile: Add manual model download step (lines 100-119)
- All models download during build → instant runtime startup

Benefits:
✅ Build logs show download progress (no "hanging" confusion)
✅ Runtime startup is instant (models already in image)
✅ No permission conflicts (consistent user ownership)
✅ Works on free tier (no /data persistent storage needed)

Updated: Dockerfile

Dockerfile

@@ -97,6 +97,27 @@ RUN ls -la $HOME/app/code/ImportLDraw/__init__.py || \
 RUN pip install --no-cache-dir --upgrade pip && \
     pip install --no-cache-dir -r requirements.txt
 
+# 预下载模型到默认缓存目录（构建时一次性下载，避免运行时卡顿）
+RUN echo "📥 Pre-downloading models during build (this may take 2-3 minutes)..." && \
+    python3 -c "\
+from huggingface_hub import hf_hub_download; \
+from transformers import CLIPModel, CLIPProcessor; \
+print('Downloading shape_gpt.safetensors (7.17 GB)...'); \
+hf_hub_download(repo_id='0xZohar/object-assembler-models', filename='shape_gpt.safetensors'); \
+print('✓ shape_gpt downloaded'); \
+print('Downloading shape_tokenizer.safetensors (1.09 GB)...'); \
+hf_hub_download(repo_id='0xZohar/object-assembler-models', filename='shape_tokenizer.safetensors'); \
+print('✓ shape_tokenizer downloaded'); \
+print('Downloading fine-tuned adapter (1.68 GB)...'); \
+hf_hub_download(repo_id='0xZohar/object-assembler-models', filename='save_shape_cars_whole_p_rot_scratch_4mask_randp.safetensors'); \
+print('✓ adapter downloaded'); \
+print('Downloading CLIP model (~600 MB)...'); \
+CLIPModel.from_pretrained('openai/clip-vit-base-patch32'); \
+CLIPProcessor.from_pretrained('openai/clip-vit-base-patch32'); \
+print('✓ CLIP downloaded'); \
+print('✅ All models pre-downloaded to ~/.cache/huggingface')" && \
+    echo "✅ Model weights cached successfully"
+
 # 创建启动脚本（先启动 Xvfb，再运行应用）
 RUN echo '#!/bin/bash\n\
 echo "🖥️  启动虚拟显示服务器 Xvfb..."\n\