This will stress-test at least the following sharding logic:

• readAfterOpTime of the configOpTime (which should be sent on all reads to config server nodes)
• readConcern in metadata commands (reads can use the majority snapshot even if readConcern on opCtx is 'local' if a prior read, such as overtaking a distlock or loading the ShardRegistry, set the majority timestamp on the RecoveryUnit)
• causal consistency of metadata changes (see SERVER-33954)

This could be a potential intern/new grad project - it's useful since it's test-only, but relatively low-risk.