> Hmm, the dateutil documentation seems to imply that relativedelta
> takes care of this:

This is all a bit moot since we don't have a "relativedelta" in the
stdlib.
I think the two questions here are:
- is this feature desirable? I think the answer is "yes"
- is the implementation correct? I don't know :-)

Talk of a magical generic constructor sounds a bit futile to me,
especially if the constructor's behaviour becomes so complicated that
it's difficult to understand for users...