> how that's acceptable?

For docutils, we'll most likely propose some variant of Stéphane Wirtel's script to test semantic equivalence for docutils.  For other cases, Serhiy is working on a C14N canonicalization tool which is specifically designed for the task of creating reproducible output, in a cross-language standards compliant way.

As Stefan Behnel clearly articulated, there are multiple reasons why Python should not guarantee byte-for-byte serialization across point releases.  That said, we'll likely make the guarantee across micro-releases.  That will make it possible a third mitigation strategy of generating new baseline files for a new point releases and adding a version check to decide which baseline to test against.

FWIW, we had a similar discussion regarding hash randomization.  While there are a number of significant differences, the outcome is relevantL  User tests that depended on non-guaranteed implementation details had to be fixed.