fix(infra): pods list live by default, and a terminated pod is "Ended" not "Down"

[Makisuo]

The Kubernetes pods view painted most of an autoscaled fleet red — hundreds of rows marked Down on a fleet where nothing was actually failing. Two causes, both fixed here.

Why "Down" fired on the normal case

deriveHostStatus returned "down" — styled --severity-error — for any resource whose newest datapoint was older than 5 minutes. But lastSeen is max(TimeUnix) over a whole query window (12h by default), so "down" really meant "this pod's metric series ended". On an autoscaled fleet that is the normal end of a pod's life: HPA scale-in, a rolling deploy, a replaced Fargate task, a reclaimed spot node.

Continuity alone cannot prove a resource is down, so the state is now ended and reads neutral (text-muted-foreground, bg-muted-foreground/40 dot) instead of error red. A genuine down signal needs an expectation — k8s.pod.phase, or a workload's available-vs-desired replicas — and belongs beside these states, not instead of them.

Why the list was full of dead pods

The list returned the union of everything that reported anywhere in the window, with no notion of who was still running. Dead pods inflated the "N pods in scope" denominator and outranked live ones in the default saturation sort.

PodLifecycle ("live" | "ended" | "all") now defaults to live, cutting at the same 300s-before-endTime threshold the badge uses — so the predicate that selected a row and the badge on it cannot disagree. It rides in listPodsFilters rather than with the scope, so listPodsCount narrows identically and the page can't read "50 of 656".

What changed

• HostStatus is "active" | "idle" | "ended"; down is gone. Shared policy, so nodes, hosts and workloads get the same treatment — an ASG or Karpenter scaling a node out is the same non-event.
• listPodsQuery applies the lifecycle outside the grouping, alongside the scope predicate.
• listPodsSummaryQuery returns livePods/endedPods as absolute counts, with the three saturation buckets counted within the requested lifecycle so the band stays a valid denominator for the list beside it.
• The old stale scope was the same idea under a worse name — it is now the ended lifecycle, and the band's fourth cell. Clicking it swaps which fleet is on screen rather than narrowing the live one.
• The band header reads "N live pods in scope · share of the live fleet by peak utilization".
• The unfiltered empty state no longer says "install the Helm chart" when the fleet has simply all ended (a finished job, an environment scaled to zero) — it points at the Ended scope with the count.

Reviewer notes

• Breaking URLs: shared links carrying ?scope=stale (pods) or ?status=down (nodes) will now fail validateSearch. I took the clean rename over carrying both literals — flag it if you'd rather keep the old values decoding.
• Not verified against real data. The local dev org has no Kubernetes metrics, so this is covered by SQL-shape tests but the visual pass on the screen is unverified. Worth a look on an org with a live fleet.
• Follow-up left out: facet counts in the filter sidebar are still window-wide rather than live-scoped, so the sidebar can offer a namespace that yields zero rows. Bounded the diff deliberately.
• The natural next step is the expectation signal — k8s.pod.phase, k8s.pod.status_reason, k8s.container.restarts, and k8s.deployment.desired vs .available are all collected by default in deploy/k8s-infra but unqueried. That is what would let "Down" mean something again, at the workload level where replicas missing (rather than churning) is the actual incident.

bun typecheck, bun run lint, and the query-engine + web infra test suites pass.

🤖 Generated with Claude Code

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}

[github-actions]

🍁 Maple PR preview

Warning

Preview cleanup could not be confirmed. The Alchemy teardown outcome was skipped.

Final commit 0e69bd6 · View workflow run