[Performance] Request hot-path benchmarks + regression gate

[fabriziosalmi]

Establish per-request latency and allocation benchmarks under representative rule sets, and gate PRs on regressions so performance cannot silently drift. Roadmap theme — benchmark harness and baselines TBD; contributions welcome.