We recently added MemberDowned event and that can be used in singleton and sharding to stop earlier rather than waiting for removal.

ClusterDaemon is shutting down itself when it sees itself as Down but it waits until other reachable nodes have marked the state as seen. The reason for not shutting down immediately is to give the gossip a chance to spread the downing information to other downed nodes, so that they can shutdown themselves. That is good but maybe we should introduce an upper bound on that delay.