The scenario that fails is described in BF-6485,
the core symptom is that the getMore sent to a secondary returns operation time that is less than the afterClusterTime set in the client request.
The reason of this is mongos parses the getMore cmdObj into the GetMoreRequest which does not include readConcern and therefore can not satisfy the request to wait.

the cursors are consistent because because the cursor is the find and getMore is an implementation detail. However the tailable cursors may need the support of readConcern