New policy: Only use test_expect_failure for broken tests