feat: agentic RAG for search_memory

[chinkan]

Summary

Opt-in agentic retrieval pipeline for search_memory: rule-based query classify → dispatch (facts / knowledge / messages / hybrid) → optional LLM self-eval → max N iters.

Why deferred

Flat hybrid (FTS + vector RRF) already ships. Extra LLM round-trip per search is latency tax until we prove recall gaps.

Spec source

/tmp/rustfox/PLAN-2026-09-01-memory-upgrade-comprehensive.md Step 4; tooltimeout plan Part 3.

Depends on

Temporal Facts + Knowledge History (this branch / related work).

Acceptance (when picked up)

• MemoryConfig.agentic_rag_enabled default false
• agentic_rag_max_iters default 2
• classify + dispatch unit tests without LLM
• fallback to flat search when disabled or LLM missing