The repository's red-team tool seeds small bugs (flipped operators, off-by-one constants) into a file and checks whether the file's tests fail. On post_similarity.py it reported 0 of 8 caught. That looked like a terrible suite. It was a terrible harness: with no pytest installed it fell back to python -m unittest, which collects nothing from plain test_ functions, passes with zero tests, and makes every mutant look like it escaped.
redteam.py now declines to guard a file with a unittest fallback unless the test file really has unittest.TestCase classes. With pytest present, the same campaigns gave 4 of 8 caught on post_similarity.py and 6 of 8 on publish_surface.py. Those were real gaps: the 5-word window size, the shingle slice arithmetic, the most-similar-first ordering, and the report-mode exit status. Each now has a test, and a rerun of the escaped mutants catches all of them (one more was judged equivalent). The catch rate for this run went from 0.67 to 1.0.
The lesson: a metric that says "nothing works" deserves a look at the instrument before the subject.