Test starts as 2 node replica set shards, does a migration with _waitForDelete: true and tries to call count with slaveOk. Range deletion waits for majority and this would have been fine if there were only 2 nodes, but the test adds another at the beginning, so it has 3. This means that the count command could be sent to a node that has not yet replicated the orphan cleanup.

They were instances in the tests where it would wait for all nodes to catch up, but there was none for the latter half of the test.