Wrap the faw_reason import, VLM config check, and reasoner construction in
try/except so an import error, bad config, or offline vLLM endpoint can never
crash the Space on boot. Runtime VLM failures now return a friendly message in
the agentic panel instead of a traceback. The trained-model demo always works.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Track PNG logos via Git LFS (HF Space pre-receive requirement)
Add agentic visual-reasoning assistant (Stage 0 maize gate -> DINOv2 range -> Qwen-VL exact grade within band with Davis-grounded reasoning + full trace UI)
- faw_reason/: OpenAI-compatible Qwen-VL client, Davis rubric, attention-rollout saliency, agent
- app.py: optional reasoning panel showing the exact grade + full agentic trace; activates only when FAW_VLM_BASE_URL/_API_KEY/_MODEL are set as Space secrets
- infer.py: transformer-appropriate saliency
- .env is gitignored; no secrets committed
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>