fix: background thread init so uvicorn binds port 7860 immediately on HF Spaces

fix: set HF_HUB_OFFLINE=1 at runtime to prevent model network calls on startup

fix: multi-model fallback cascade (4 models x 3 keys = 12 combos) + FastAPI root redirect

fix: bump langchain-core to 0.3.29 (langchain-groq 0.2.3 requires >=0.3.29)

fix: bump pydantic to 2.10.6 (langchain 0.3.7 requires >=2.7.4)

fix: pin all requirements_api.txt versions to stop pip backtracking

fix: add HuggingFace Spaces README.md with sdk: docker metadata

fix: lazy GeminiKeyManager init + safe CORS default

feat: multi-key rotation for Gemini (3 keys) and Groq (3 keys)

feat: full optimization - Groq LLM, disease cache, deploy configs

Push all code to huy_xau_trai repository

Add RAG integration: Flask API server, disease selector, evaluation system with improved case generation