feat: multi-step ReAct diagnosis

[Goodnight77]

Closes #21

diagnose() now runs a reason/act loop (LangGraph create_react_agent): search manuals, search history, reason, and re-search with a refined query if needed before answering. Output stays the Diagnosis schema.

• Tools = search_manuals/search_history (MCP-first + direct fallback, unchanged)
• Model: agent_model (deepseek-chat, needs tool-calling) via ChatOpenAI
• Any agent failure degrades to the previous single-shot path
• Verified end-to-end against DeepSeek + Qdrant (multi-step retrieval, grounded Diagnosis)