homepage

While working on issue 19330, I also took a look at whether or not I could fix the @contextmanager decorator to also provide useful help information on the resulting objects. This turned out to be difficult, since calling decorated functions produces instances of contextlib._GeneratorContextManager, and help() insists on showing the docs for that rather than a __doc__ attribute set on the context manager instance itself.
This behaviour appears to be ultimately due to "pydoc.render_doc()", which has a hardcoded list of conditions it checks to decide whether to document the instance directly or to provide the class documentation instead:
http://hg.python.org/cpython/file/2b904290b3b9/Lib/pydoc.py#l1544
This first idea that comes to mind is to simply add the additional condition that if "obj.__doc__ != type(obj).__doc__", then help() shouldn't replace the object with it's type. The simple trick that I originally explored was simply setting the __doc__ attribute on the _GeneratorContextManager instance based on the docstring of the decorated function, and I believe that approach might work for property instances, too (rather than needing to hardcode the isinstance check into render_doc()). It may even handle some of the other currently hardcoded conditions in that list.

New changeset 09153a9a3bb9 by Nick Coghlan in branch 'default':
Close #19330 by using public classes in contextlib
http://hg.python.org/cpython/rev/09153a9a3bb9 

Issue 19031 is a report indicating that this is a problem for Enum instances as well.

Would it make sense to have the presence of a non-None __doc__ be a deciding factor? How often does an instance have a docstring where one would want the class info instead?

As far as I am aware, never. But "instance.__doc__" falls back to the
class, hence the != check rather than "is not None".