Local RAG performance

Prefill ate 85% of a 291-second answer, and the fix was a dedup key and a cache slot

One question over nine files assembled a 13,773-token prompt, and llama.cpp reused three tokens of it. Reading beat writing by about five to one.