Routing back-pressure alerts now name what they actually saw

When a silo had a lot of deliveries in flight at once, it raised a critical alert saying that deliveries "were not completing" and pointed at whichever address happened to be routed at that instant. Both halves were misleading. The count it reports rises whenever the machine is short of CPU — every delivery can be moving along perfectly and the number still climbs — and the address named is simply the last one dispatched, not a slow one.

That wording sent several investigations after the wrong thing, so the alert now reports only what it can actually see: how many deliveries are in flight, how many destinations they are spread across, and whether any of them are queued behind another delivery. The last of those is the one that matters — deliveries waiting on a delivery is a real blockage, while the same number spread across many destinations is just a busy moment. When the backlog clears, the silo now also says how long it lasted, which separates a two-second burst from something genuinely stuck without anyone having to inspect a running server.

Separately, a delivery that was cancelled while the silo was shutting down used to end without ever reporting that it had finished, so the slot it occupied was never given back and the queue for its destination was left standing. Those deliveries now close out properly.