Polish: BGE-large embeddings, contextual retrieval, 142 tests passing, lint clean

Add eval dashboard, persistent chat, and multi-doc comparison

Add streaming responses, multi-turn chat, and Second Brain

Security, bug fixes, and UX improvements

Strip whitespace from API key — httpcore rejects keys with trailing newlines

Add A-RAG — hierarchical retrieval interfaces (Feb 2026)

Add Speculative RAG — Google Research 2024 (~51% latency reduction)

Add TTRAG — test-time compute scaling for RAG (ICLR 2025)

Fix Claude API call: pass system as plain string, not cache_control block

Fix eval pipeline and add real benchmark numbers

Initial commit — production RAG system