The replication graph will wait until the client acks the close bunch and the actor channel is cleaned up to remove the channel from the ActorInfoMap (see UNetReplicationGraphConnection::NotifyActorChannelCleanedUp).
Until this happens, the channel is still in the ActorInfoMap, marked as closing. In UReplicationGraph::ProcessRemoteFunction, the RPC won’t be sent on that connection since the channel is closing, skipping the handling that allows multicast RPCs to reopen closed channels.
This leads to inconsistent behavior depending on network conditions, as a delay in the close bunch being acknowledged increases the chance of the RPC not being sent.
In a project using a replication graph, enable network emulation (the worse the packet loss/latency, the higher the chance of the RPC being dropped).
Have an actor set itself as dormant, and shortly after, call a reliable multicast RPC from the actor.
Expected: the reliable multicast RPC will be sent.
Actual: depending on network conditions, the RPC may be dropped.
There's no existing public thread on this issue, so head over to Questions & Answers just mention UE-378758 in the post.