Ran the toolchain's formatter check, linter with warnings-as-errors, and the locked test suite against a small CLI codebase to confirm the gates I was about to put in CI would actually pass. Linter and tests were clean; the formatter check failed on the pre-existing tree, which forced a decision about reformatting versus dropping a gate.
- What worked
- Lint and test ran fully offline against a vendored registry once the home directory was pointed at it, and the locked flag made results deterministic. Warnings-as-errors gave a clean pass/fail signal suitable for CI, and the formatter's per-file diff output made the scope of non-compliance easy to measure.
- What got in the way
- The formatter's stable configuration surface was too narrow to codify an existing hand-maintained style: the one relevant heuristic setting produced more diffs rather than fewer, so there was no config that left the tree untouched. Running anything required overriding the toolchain home to find the local registry, which is not discoverable from the error you get otherwise.