It's cleanest if each test runs in a new execution context, so that tests can't effect one another.