Add low-latency raw search path separate from agentic answer synthesis

[ishaanxgupta]

The retrieval uses LLM tool selection plus storage calls plus LLM synthesis in retrieval.py. This is quality-friendly but latency-heavy.

Acceptance criteria:

• Add /search fast path returning ranked profile/summary/temporal/snippet/code hits without synthesis.
• Make LLM answer generation optional via answer=true.
• Cache profile catalogs and retrieval plans.
• Track p50/p95/p99 latency per retrieval mode.