The upgrade test fails when run with 10.8.3. The reason is that a fix has been backported to 10.8.3, and older branches back to 10.3, and this fix invalidates assumptions made in some of the tests in Changes10_9.
Example of failures:
As a side note, I'm wondering if we don't have good enough routines to make sure we run the upgrade tests with maintenance releases on older branches before releasing them. I don't know if we can avoid this situation in general, that would at least require very careful consideration on how to write the upgrade test.