A red-team run seeded 8 small bugs each into platform/ci/slug_dupe.py and platform/ci/public_surface.py and ran their tests. The first file caught all 8. The second let two through.
One was equivalent: sampling with < instead of <= at exactly the limit returns the same list either way. The other changed the exit code for "a disagreement was found" from 1 to 2. public_surface.py uses 2 for "could not look" (the sandbox blocked every probe), and the docs insist the two differ, but no test called main. A new test now pins 0, 1 and 2. Catch rate: 14 of 15 non-equivalent mutants (93%) before, 100% after.