fix(opencost): usage metrics are absent for every namespace, so efficiency and rightsizing are unmeasurable #2782

🤖 Generated by the Daily AI Engineer

Why

OpenCost reports cpuCoreUsageAverage: 0 and cpuEfficiency: 0 for all 34 namespaces on prod. Requests are populated correctly (e.g. kube-system 5.32 cores), so the allocation side works — only the usage side is empty.

The consequence is worse than a missing feature: a broken sensor and a genuinely idle workload look identical. Anything reading efficiency would conclude the entire cluster is 100% idle and "rightsize" it on an artefact. Cost attribution by CPU/RAM still works; efficiency and rightsizing do not.

Reproduced 2026-07-21 via the allocation API (svc/opencost:9003, window 3d). OpenCost is configured to read Coroot's bundled Prometheus at coroot-prometheus.observability.svc.cluster.local:9090, so the likely cause is that query path — the metric names it expects, or reachability.

What

Restore usage metrics so efficiency is real, then confirm a known-busy namespace reports non-zero usage.

Acceptance criteria

• cpuCoreUsageAverage is non-zero for namespaces that are demonstrably doing work
• cpuEfficiency is plausible (not uniformly 0, not uniformly 100%)
• Root cause named — metric names, query path, or reachability
• A check that fails if usage goes uniformly zero again, so this cannot regress silently

Part of #2355 (FinOps agent) — its rightsizing analysis is blocked on this and currently refuses to emit candidates rather than emit wrong ones.