Given this sequence:

1. mongos RSM cannot reach all shard replicas
2. Shard adds new node, which in principle should be reachable.
3. config.shards was updated with the latest connection string by the shard primary.
4. mongos refreshes from config server, and fetches the connection string.
5. The ShardRegistry usually re-uses the RSM instance, so it will keep being disconnected unless we can tell the RSM that there's something new.

Although this scenario is unlikely (if all shard nodes were unreachable, it seems likely that any new nodes will also be unreachable), it posses a problem in a scenario where the ShardRegistry can still connect to the CSRS, gradual ShardRegistry refreshes would never recreate the RSM. If the network is eventually re-established, the RSM will still be trying to track the old nodes.