During global process shutdown, we kill all operations and return a shutdown error over client sockets. But this doesn't actually make a lot of sense, as the host is unlikely to be a good target for more operations over those sockets and network errors are already retriable. Just closing the socket would probably produce better results than what we're doing today.

This should be a reasonably simple change (a check, probably somewhere in the service state machine / entry point, to close the socket rather than return a result if we're in global shutdown and we've been killed with a shutdown error)