Measure startup and recurring cost of the default health sampler #3335

Measure the startup and periodic cost of #372's default-on health polling before choosing a lighter backend. The module stays cold until startup readiness, but with health checks enabled that first use happens on every boot. CPU polling imports systeminformation; unconstrained memory polling also uses it, at the default 60-second interval.

Source inspection of the installed 5.x package confirms platform-specific subprocess work for memory/swap data (PowerShell on Windows and vm_stat/sysctl on macOS), including information the health signal does not consume. This is source evidence, not a measured native-platform regression. No throughput, heap or latency win is claimed yet.

Compare startup heap/time, poll duration and child-process count on Linux, Windows and macOS. Assess os.totalmem/os.freemem and os.cpus time deltas where they preserve the intended semantics. Node24 Linux os.freemem() matched MemAvailable in the investigation; macOS available/reclaimable semantics need a separate check. Preserve cgroup working-set behavior, deadlines, one-pending-reader limits, stop ownership and startup-only configuration. If recommending a backend replacement, show the before/after measurements and native-platform correctness tests.

Priority P3: off-roadmap optimization/measurement task with no established customer impact or measured regression. Effort Medium: known collector paths, but cross-platform measurement and semantic checks. Parent #1679 (runtime portability/dependency slimming), whose P2 is one tier above this unmeasured optimization. No release commitment/milestone. Customer Impact and Security Notes inapplicable.

Review-Deferral: a46004c5649de4c3 severity=minor category=performance

_{🤖 Generated by OpenAI Codex; posted via @kriszyp.}