A portal replaced mid-startup shuts down cleanly

A rolling update can replace a portal at any moment — including while it is still starting up. When that happened, the old pod did not go quietly. Its startup check was cut off mid-question, and the last thing it wrote was a bare "hosting failed to start" with no explanation, followed by a handful of errors from background work that was still shutting down after the pod had already released the resources it needed.

Nothing was actually broken — the pod was on its way out either way — but the errors were indistinguishable from a real failure, so every rollout that caught a pod mid-start looked like a problem to investigate.

Now the startup check says exactly what happened: it was cancelled because the pod was being replaced, so nothing about the database was confirmed or faulted. And the background work that finishes a shutdown no longer depends on resources that may already be gone — it takes what it needs up front, so it always completes its work and reports the result, on both the ordinary shutdown path and this one.