A pod shutting down is no longer an error on the second path either

When a portal replaces its pods, some nodes are inevitably waking up on a pod that is already on its way out. The work is abandoned, nothing is written, and the next person to open that node gets it from a live pod. It is the most ordinary thing that happens during a deployment.

Two weeks ago the platform stopped reporting that as a failure. Before, it could not tell the difference between "this pod is going away" and "this node's configuration is broken" — one message named both possibilities and committed to neither — so every rollout raised an alarm about a condition the message itself described as expected.

That fix was correct and incomplete, in a way worth writing down. A node waking up runs through two steps, and each has its own place to report a problem. Only one of them was taught the difference. The other went on reporting every failure at alarm level, including the shutdown.

What makes that more than a missed line is how alarms are grouped. They are deliberately grouped by the fault, not by who reported it — so that one problem printed by two different parts of the platform raises one alarm instead of two. That is the right rule, and it has a consequence nobody had drawn out: quietening one reporter changes nothing while another still shouts. The alarm the first fix closed reopened a fortnight later, reporting the same shutdown from the second step, and looked for all the world like the fix having come undone.

Both steps now agree: a node abandoned because its pod is tearing down is recorded as the routine event it is, and says so in a sentence instead of listing possibilities.

Nothing became quieter than that. A node that genuinely cannot start — a configuration that throws, a node that never resolves, a step that times out — still raises the alarm, with the same wording as before, and with no change at all to what actually happens: the failure is still recorded, still reported to whoever was waiting, and the node still tries again on the next access. The only thing that changed is whether a deployment doing its job wakes somebody up.

The general lesson is recorded with the machinery: when an expected condition is downgraded, the thing that keeps the alarm alive is the reporter nobody thought to look for.