So, we are talking about adding a feature that could cause problem whether 
cleanup is performed before tearDown or after tearDown. Don't we risk 
confusing developers who are not familiar with the cleanup order?

Do we really want to add this feature? The added functionality doesn't 
seem so attractive compared to the risk of confusion. We might as well 
continue to let the developers implement this feature in their own 
TestCase subclasses (for example, my own subclass already implement a 
TestCase.create_tmpdir(name) (which does cleanup, of course), which is 
easier to use than the rmtree example that was given as a use case). At 
least, when they do, they know about the cleanup order.