Because we allow deletes during build again... There's a race condition in compute manager's _run_instance() where an instance can be deleted in the middle of an rpc.call to the network manager to allocate networks. rpc.call will raise a RemoteError and we will try to set the instance to ERROR state. The setting of instance to ERROR state fails with InstanceNotFound... and eats the original exception.
We shouldn't blow up when we try to set an instance to ERROR that has been deleted out from under us.