feat: frecency-in-ranking evidence/benchmark behind a default-off flag (informs #155)

[edheltzel]

Related: #155.

Recall computes frecency (src/lib/frecency.ts) and bumps access_count / last_accessed (src/lib/memory.ts bumpAccess), but ranking still sorts by RRF score only (src/mcp-server.ts:269-275).

Red-team caveat: do not auto-change shared ranking. #155 is needs:human and explicitly says confirm with the maintainer before changing shared ranking behavior.

This issue is evidence work:

• Add a default-off flag that blends frecency into ranking.
• Add a Suite-F benchmark using benchmarks/suites/suite-f-search-latency.ts to measure whether recency-in-ranking helps.
• Bundle a small env-gated search-timing log to feed the benchmark.

Goal: give Ed data to decide #155, not silently change production ranking.