Make the tests use nose instead of unittest, and use properly
parameterized arguments. This makes it a lot easier to see
what fails.

To get commands, env vars etc, we need the image's defaults to
do a diff.

I knew I'd forgot something.