Currently a master with HA replication configured to check-for-live-server will check the presence of other nodes in the cluster with the same node ID by:

• configuring an HA cluster connection using the configured cluster connector list with no reconnection attempts
• register for the topology on such connection
• if any member of the topology with the same id is found, trying to create a ClientSession to it

Spit brain could be exprienced in different cases:

• if the node we're connecting to receive the topology doesn't have a correct topology (eg slave still becoming live after a previous crash of master)
• if the node we're connecting to receive the topology is not reachable: master will assume to be the only live

A possible improvement could be to not rely on the cluster connection topology as the source of truth. but perform a(nother) check requesting a quorum vote to know if it's correct to became live.