This is an issue for tracking progress towards 100% test coverage. I believe that, as measured, #520 will be a prerequisite for this.

A lot of the uncovered lines according to Codecov that aren't related to the "no builds on Windows" stuff seem to be error conditions and repr values for various private and/or public classes.

For private class reprs, we can "smoke test" them in test_internals.py, which is where we put tests that aren't intended to document the public interface, but just to exercise the various lines that otherwise would go under a nocover pragma.