Whilst I agree in principle... The trouble is that when you do this:

    python -m unittest test_pk

What you are saying is "run all the tests from the test_pk module". You *aren't* launching discovery. 

This should work:

    python -m unittest discover -t . test_pkg

This is more verbose than is ideal. Suggestions for improvements welcome. Having unittest revert to discovery when it is passed a package name that turns out to be empty seems a bit magical (and complex in terms of implementation).

Yes, calling loader.discover inside a load_tests function will mutate that loader - so having discover restore _top_level_dir on exit would be better. Can you post that as a separate issue?

I think there is a separate issue for improving the test failure name (including module) reporting. I'll try and dig out the issue number.