When a mongod is started with --repair and data or metadata is modified by the operation, the instance will be unable to rejoin its original replica set without a full resync. The only way to add the instance back to a replica set would be to delete the local.system.replset collection or wipe the data directory entirely.
Additionally, if a repair operation fails to complete for any reason, the instance will not be able to start again without the --repair option.
Engineering Ticket Description:
When MongoDB is started with --repair, the repair process will attempt to salvage data at the expense of potential data loss so that MongoDB can be started normally.
If a user is unaware they have lost data, they may attempt to add the node back to its original replica set. In the worst-case scenario, this node will become primary and data will appear to go missing.
There are currently no measures in place to prevent a secondary with data modified offline from re-joining its original replica set (even without repair).
Solution: If repair modifies collections (not indexes), repair will add a top-level field "repaired: true" to the local.system.replset configuration document. This will invalidate the document and prevent the node from re-joining as a member of the replica set. When the server starts up again normally, a warning will be printed with instructions for performing a re-sync.