Honest eval narrative + query rewrite off by default

Add query rewriting + corrective RAG + 3-stage RAGAS ablation

Add reranking + RAGAS evaluation

Remove Gemini 3 Flash: persistent 503 unavailable

Cap LLM history at MAX_HISTORY_TURNS=4 to avoid context overflow

Fix HF auth: explicit login from HF_TOKEN before gated model download

Deploy: auto-ingest, hybrid RAG, streaming, UMAP viz, 16 LLMs