“AI models in the MirrorCode benchmark particularly struggled with the `ruff` Python linter, the `giac_subset` mathematics package, and the `mailauth` email authentication library.”