I think the timezone related problems are a red herring.  Aware datetime +/- timedelta arithmetics is naive - tzinfo is ignored in calculations and copied to the result:

http://hg.python.org/cpython/file/c83ce2a1841c/Lib/datetime.py#l1711

The utcoffset only will only come into play if we want to implement time - time -> timedelta, but this problem is already there in time comparisons:

http://hg.python.org/cpython/file/c83ce2a1841c/Lib/datetime.py#l1091

It is up to tzinfo subclass implementation writers to handle inability to compute utcoffset without date fields by raising an exception if necessary.  It is perfectly fine for time - time to fail with an error coming from .utcoffset().

I also don't think the fate of #13936 has any bearing on this issue.  As long as we are not trying to implement time + time -> time, we are not introducing any new notion of "zero time".