The rsm_horizon_change.js tests that the split horizon reconfig causes the ReplicaSet to go into RSNoPrimary before going back into RSPrimary State.

Fixed a few months ago, SERVER-57893, explains how due to network events it can lead the Topology Change to the following states before going into RSPrimary State

1. RSNoPrimary or
2. It can send RSM into "RSM monitoring host in expedited mode until we detect a primary". 

But the latest error shows that drop in connections can also lead RSM monitor to stay in RSPrimaryState(by dropping connections in reconfig state and then reconnecting)

Suggested Fix: Add additional log check for "Dropping all pooled connections" to 
unknownServerDescriptionRegex