So, after getting in contact with Atlas, indeed they do document the fact that they delete records older than 24hrs.

https://www.mongodb.com/docs/atlas/cluster-additional-settings/#:~:text=Atlas%20enables%20auto%2Dscaling%20by,removes%20them%20from%20the%20oplog.

This is subtle though, and while it is linked to storage.oplogMinRetentionHours. That documentation explicitly says BOTH oplog capacity AND minRetention must be exceeeded prior to oplog logs being removed. There is no way my oplog with 3 records in it exceeded any oplog capacity settings (unless Atlas sets it to 0 by default) so I feel like this confusing at best.

There is no way for shared (M0/M2/M5) plans to set the oplogMinRetentionHours value either, which was a bit more obvious in the documentation.