Two-peak (bimodal) latency detection, ported from the web dashboard’s
detectBimodal (apps/lobe-web/lib/analytics.ts) so the CLI can flag a
fast/slow cohort split natively — without shipping samples to the LLM.
Keep the constants in sync with the TypeScript implementation.
Latency-under-contention probe: hold N requests in flight while pushing a
fixed total through, each measured with the normal phase decomposition.
Deliberately not a throughput benchmark — the output is the latency
distribution per concurrency level, not requests/second.