The idea of splitting integration from unittests is good in theory
but leads to a lot of mental overhead. Besides whenever a tests
interacts with a database it isn't a unittest anyway.

Better to use filenames which are commonly used in the python word then
custom ones.

This commit was triggered by a recursion problem mentioned here:
8199da2100 (commitcomment-7387486)

It now unifies handling for name and description across both conditions
and benefits:

* Declaring a name property is required for any proxy class. A few
  classes had a description declared, but no name. But given that the
  description is the more verbose version, and is allowed to include
  HTML, it makes sense to still require a name. The declared
  descriptions in the models that only had one property are better
  suited for @name anyway.
* It drops the max_unaffected_items logic for Benefits. That was only
  applied if calling via the non-proxy class, and feels more like it
  should be a description anyway.
* __str__ always returns self.name. Anything more complex with that
  means we might run into problems with the python_2_unicode_compatible
  decorator.

I also added tests for the changes, and verified the tests fail without
my changes.

Now using --spec format which is more readable (most of the time).

They weren't applying to lines properly.  Fixed with test.