BF-42873 found that the following order of events is possible:

1. The validate test hook performs an insert and drop, gets the opTime of the write to use for atClusterTime
2. The writes make it to the secondaries but are not applied yet on at least one node
3. w=N write concern is satisfied so we return from the commands which allows us to run validate with atClusterTime=T (This is because majority write concern acknowledgement behavior was changed in 8.0. We acknowledge the w=N insert not when the oplog entry is applied, but when we journal it to the oplog.)
4. Validate opens the read snapshot at time T, even though we have not yet applied all oplog up until that time
5. We finally go to apply the oplog entries and hit this invariant, since we are committing behind an open read timestamp.

In the real read concern code, we call replCoord->waitUntilOpTimeForRead before setting the read timestamp to an atClusterTime timestamp. This call is not included here in the[ validate code|https://github.com/10gen/mongo/blob/92f469a3cd05ee5b9e3eba295f74966d23bdc8f2/src/mongo/db/validate/validate_state.cpp#L316-L317]. Adding this wait call should solve this issue.