This is because remove() only removes the (shared_ptr) reference to the ReplicaSetMonitor from the ReplicaSetMonitorManager: it does not cause the ReplicaSetMonitor's destructor to be called.

Is there any way on ReplicaSetMonitor::remove() to interrupt ongoing refresh cycle(s) from concurrent threads with a ShardNotFound or even better, a ShardRemoved-esque error?

Here's an example of refreshes continuing after we attempt to remove the ReplicaSetMonitor:
https://logkeeper.mongodb.org/build/45b90edca90e143c2b59e6e42ad34c9d/test/5810e98cbe07c45f730ab605#L1120

from this Evergreen run on October 26, 2016:
https://evergreen.mongodb.com/task/mongodb_mongo_master_enterprise_rhel_62_64_bit_sharding_WT_67994f33a88f4aa70283e155c75f48ce997ccdc3_16_10_26_16_05_19