Fix: Remove use_safetensors=True for CLIP model loading

Root cause (Research findings):
- openai/clip-vit-base-patch32 main branch only has pytorch_model.bin
- model.safetensors exists in PR #21 (revision d15b5f2) but NOT merged to main
- Code forced use_safetensors=True → file not found → runtime failure
- Working version (commit b9fd613) did NOT have this flag

Solution:
- Remove use_safetensors=True from CLIPModel.from_pretrained (line 144)
- Remove use_safetensors=True from CLIPProcessor.from_pretrained (line 155)
- Let transformers auto-detect available weight files (pytorch_model.bin)
- Add documentation explaining why we use pytorch_model.bin

Security considerations (CVE-2025-32434):
- Using official OpenAI model (high trust, low risk)
- local_files_only=True prevents malicious replacements
- Model preloaded during build in controlled environment
- Same approach as working version b9fd613

Official documentation:
- CLIP model files: https://huggingface.co/openai/clip-vit-base-patch32/tree/main
- Safetensors PR: https://huggingface.co/openai/clip-vit-base-patch32/discussions/21
- HF Spaces preload: https://huggingface.co/docs/hub/spaces-config-reference

Changes:
- code/clip_retrieval.py lines 141-157: Removed use_safetensors=True flags
- Added documentation comments explaining the choice

Expected result:
✅ CLIP model loads from preloaded pytorch_model.bin
✅ Text-to-LEGO functionality works correctly

code/clip_retrieval.py

@@ -141,7 +141,10 @@ class CLIPRetriever:
                 model = CLIPModel.from_pretrained(
                     self.model_name,
                     cache_dir=HF_CACHE_DIR,
-                    use_safetensors=True,  # Force safetensors to bypass CVE-2025-32434
+                    # NOTE: Not using use_safetensors=True because openai/clip-vit-base-patch32
+                    # only has pytorch_model.bin in main branch (model.safetensors exists in
+                    # revision d15b5f2 but not merged). Using pytorch_model.bin is safe for
+                    # official OpenAI model with local_files_only=True (prevents malicious replacements)
                     torch_dtype=torch_dtype,
                     local_files_only=True  # Use pre-downloaded model from build
                 ).to(target_device)
@@ -149,7 +152,7 @@ class CLIPRetriever:
                 processor = CLIPProcessor.from_pretrained(
                     self.model_name,
                     cache_dir=HF_CACHE_DIR,
-                    use_safetensors=True,  # Force safetensors to bypass CVE-2025-32434
+                    # Processor doesn't have weight files, use_safetensors not applicable
                     local_files_only=True  # Use pre-downloaded model from build
                 )