fix: contextual retrieval latency — max_chunks gate + prompt trim + model update

eval: v1.3.0 clean run (25 samples) — recall=0.627, relevancy=0.837; fix eval script max_tokens+print

fix: eval script — max_tokens 500→1024, max_workers 2→1, timeout 120→180 to reduce RAGAS failures

feat: add --contextual flag to eval script for contextual retrieval eval run

feat: Stage 8 — eval dashboard, mandatory web search, docs update