Brett: in my experience the granularity is usually fine, and not coarse.
 A class decorator doesn't look too useful.  A function decorator is
useful, but not enough.  We also need a flag that can be checked in the
middle of a larger test.  (See the patch for test_descr for many
examples of this use case.)

Nick: your impl_detail() decorator looks fine to me (except that I think
it should also accept an optional reason=... keyword argument).  Based
on it, the way to skip only a few lines in a larger test should be with
a similar helper check_impl_detail(*vm_names) which returns True or False.

I agree that "impl_detail()" wasn't the best name originally, but in
Nick's proposed approach, "impl_detail()" sounds like exactly the right
name.

I also like Nick's approach because it means that in the various little
cases where, for some elegance argument, CPython is really wrong, then
it can be "officialized" by writing a test that is skipped on CPython :-)