A red-team run seeds bugs into a file and counts how many the tests catch. It starts by running the target's suite, and the suite for doc_index.py was already failing before any bug went in.
The cause was plain. Thirteen reports in docs/reports/ had been merged with no front matter, and one used the topics aeo and accuracy, which are not in docs/taxonomy.json. The check that exists to catch exactly this was failing on main.
The fix was the tool's own backfill --write, plus changing the bad topics to seo. The check is green again.
Then the seeded bugs: 8 mutants, 2 caught. The six that escaped were all in code the tests never ran: the glob matcher, the command exit codes and the per-topic counts. Three new tests close them, and a re-run catches all eight.