If a StaleShardVersion is encountered while running a BatchWriteExec::executeBatch, we will call ChunkManagerTargeter::noteStaleShardResponse that will populate the _remoteShardVersions map  with the stale error info.
Then a refresh of the targeter is attempted that will trigger a refresh of the routingInfo , this can throw a ShardNotFound exception (catched here). At this point we will loop again in the executeBatch function and we will call ChunkManagerTargeter::noteCouldNotTarget, now since the _remoteShardVersions map is not empty we will hit this dassert.

The bug has been triggered by the jstests/concurrency/fsm_workloads_add_remove_shards/clusterwide_ops_with_add_drop_shards.js test that is manually dropping the `config.shards` collection of the mongos causing the ShardNotFound error. It is only happening in the Enterprise RHEL 6.2 DEBUG Code Coverage and the

SAN Enterprise Ubuntu 18.04 DEBUG both in 4.4 and in master.