Compute-host OOM alert observes only the services host

[rldyourmnd]

Problem\n\nhost_oom_detected evaluates gha_fleet_host_oom_kills_total, but that metric is currently emitted only by the central fleet observer on the services host. Compute hosts record authoritative boot-scoped OOM counters in their pressure state, yet no Prometheus endpoint exports them. Real host-global OOM kills on compute members therefore leave the managed alert normal with empty incident history.\n\n## Outcome\n\nExpose each compute host pressure state over a hardened loopback observer using the existing metric name, so the current alert covers all five hosts without a duplicate alert or second counter definition. Invalid/stale state must fail closed.