Both the old and new approaches are deeply unsatisfying to me, but
without reworking how these test commands are run, I think this is
about as close as we can get to enforcing that this specific call
raises ValueError.