@billyvg spotted that:

For on error sampling, we moved sampling decision from the beginning of the process, to right before we flush the replay, but some of the tagging logic was not updated to reflect this. When we get an error (for on error sampling), we:

• always tag the error event with the replay id
• send event
• add error id to replay context
• SAMPLE ROLL
• flush

So if it's unsampled, we will have incorrectly tagged the event and replay