No more failure storms when a node or pod goes away

When the mesh could not deliver a message, it replied to the sender with a delivery failure. That is right for a real request someone is waiting on, and wrong for two kinds of traffic: background keep-alive pings, which nobody is waiting for, and failure replies themselves, which produce another failure reply and go round in circles.

Routing had always been written to skip both — but the check was looking at the wrong thing and never actually matched, on every routing path in both the single-process and clustered hosts. So a node that had been deleted, or a pod that was shutting down, generated a burst of pointless failure traffic at exactly the moment the portal had least capacity to carry it: pages that hung, background work that stalled, and in the worst case a restart.

The check now travels with the message instead of being re-derived after the fact, so it works everywhere, and both hosts genuinely behave the same way. Real requests still get their answer immediately, so genuine errors surface just as fast as before.