Thanks, Vajrasky. The v2 patch was almost correct. What you couldn't know without being as deeply enmeshed in this code as I am is that the test failures from the encoders module were actually invalid. We'd previously "fixed" them, but the fixes were incorrectly compensating for this bug in set_payload. Once this bug was fixed, those "fixes" just needed to be backed out, a 'decode=True' added to their get_payload calls, and all the tests pass.
I *think* this is the last inconsistency in the model. I hope.