Liveness probe — pod runtime snapshot

Always returns HTTP 200. On the public API the body is { status: "OK" }. On the internal instance it carries the pod's rolling 60s runtime window — mean cores against the cgroup quota, responses served, RSS against the memory limit, worst event-loop drift. Nothing here fails the probe yet: the thresholds that would identify a GC-bound pod are not calibrated against production traffic, so this ships as telemetry first. Every window is also logged as [RUNTIME] window.