If we collect enough of these test cases, we can reverse-engineer the bug!

My theory is there's some code that ends up processing a list of emails (even though the functionality suggests it should take a single email -- the one supplied by the user).
Rather than select the appropriate email from the list it does something like take the tail of the list.
Sometimes that ends up being the same as the email input by the user, sometimes it's another passenger on the itinerary.