The way to go about this is to see if breaking existing practice will break current implementations and plausible future implementations.

Allow me to apologize, once again, to Microsoft. In the NT 3.5.1 resource kit they shipped a DNS server which had to do its zone transfers one record per message since "existing practice" and "current implementations" meant BIND4 which knew no other way. Fortunately we didn't write a BCP describing BIND4's deviant behaviour, but rather, fixed it in BIND8 and beyond.

...

If that's not the case, though, consider that a correct implementation of DNS would be within its rights to take note of the "same serial number but incoherent answers" condition and declare the zone unreachable. I'm not

Would be pretty silly, and overstepping the robustness principle.

Whether behaviour is robust enough to be called a BCP or not is fodder for a detailed analysis amongst people who *want* to study and debate such things. That mailing list, for DNS, is called namedroppers@ops.ietf.org. (Not NANOG.)

So by your logic, by making sure that the serial numbers never match, we would 'unbreak' the situation? Seems like a step in the wrong direction.

There is, simply is and we're not going to argue about it, an identity mapping between a zone's contents and a zone's serial number. If you don't like that then you should find a way to change it. Which direction is "wrong" is better discussed on namedroppers@ops.ietf.org than here.