perf: reranker -> bge-reranker-base (hit 88.3% in multi-query, 3x faster than v2-m3; MiniLM too lossy)

perf: lighter reranker (ms-marco-MiniLM-L-6-v2) via config; bake it in image

fix: enforce verbatim identifiers in RAG answer + synthesizer (no truncation)

deploy: SAP ERP AI Agent (HF Spaces docker)