This was indeed a bug and has been fixed.
It’s benign, in the sense that the error reporting is wrong after the first violation is reported, but the end result - that stable-compatibility does not hold - is still valid.
Thanks alot for reporting it and for sending me a repro!