A restarted browser session or disconnected worker no longer costs the server a failed delivery per change

When a portal restarted, or a script worker connected over gRPC went away, every node those sessions had open kept sending its changes to the address that was no longer there. The delivery was refused each time, an error was logged each time, and because the refusal could not be delivered back to the node that sent it, nothing ever told that node to stop — one production portal spent hours refusing tens of deliveries a second (over sixty thousand in half an hour), burning CPU and log budget that real work competes for.

The refusal now carries the cluster's verdict that the address is genuinely unserved, that verdict reaches the sending node, and the node drops the dead subscription on the spot. A session that is still alive is unaffected: it re-subscribes on its own, exactly as it does after the node it watches is recycled. The routing log also stops paying per delivery — a known-dead address earns one full error line per minute carrying a count of what it absorbed, so an incident stays visible without burying everything else.