Catalog-level data corruption is only detected across restart (or rollback to stable). While a collection validation does read collection and index data from disk, and compares some aggregates to in-memory collection objects (e.g: multikey), the collection object is never checked.

We should add some basic checks to collection validation. Most importantly, verifying the on-disk document exists. Other possibilities (not all may be possible, particularly for background validation):

• Comparing in-memory multikey values to disk ones
• collection + index ident mappings
• UUIDs
• Namespace value

This check will help catch some errors that are currently only observable on tests that intentionally crash/restart, but may be occurring much more frequently, leading to easier reproducers.