Your last point there has been how I've tackled it. My test methods/classes are all well named and broken up into particular chunks so that even if the test failed message isn't very useful, the context it failed in answers it 90% of the time :)