I think some variant of these patches was rejected before and part of 
these reason had to do with subtle semantic changes when switching from 
an old-style class (inheriting from UserDict) to a new-style class 
(inheriting from dict).  There were also some concerns that it could 
break code currently relying on isinstance(x, UserDict).

IMO, this just isn't work it.  The os.environ use cases are not 
typically the performance critical part of a script.  This is a false 
optimization.  The primary motivation seems to be that the OP doesn't 
like UserDict.

Looking at the patch, I find the existing code to be more self-
evidently correct and maintainable than the proposed new code.  Though 
this is probably a matter of taste.