For cluster-level CheckMetadataConsistency, the dbPrimary shard of a given database is responsible for coordinating the metadata checks for that database (e.g. dispatching the check to all participants). To do so, it first takes the DDL lock, which requires a dbVersion check. If the dbPrimary shard happens to have corrupted database metadata, the dbVersion check will fail and the database metadata consistency won't be checked.

When this happens on a cluster-level CheckMetadataConsistency run, CMC silently skips the database and reports no inconsistencies for it.
When this happens on a database-level CMC run, the command at least fails with StaleDbVersion, though it does not report a metadata inconsistency.

So, ironically, CMC malfunctions when there is a database metadata inconsistency.

This ticket should fix this so that database metadata inconsistencies are always reported by CMC.