[ioquatix]

The releases job ran releases:check only on pull requests and merge groups, so a hand edit to releases.md that got past the pull request's check, for example with a forged Release: trailer on its own commits as the review of #88 reproduced, landed on main unchecked. On pushes the job now checks against github.event.before, the tip of main before the push, so the squash commit is judged by the commits it landed; pull requests still use HEAD^1, the current base in GitHub's merge commit, merge groups their own base, and manual runs check nothing. A scripted repository with such a hand edit shows the effect: with bake-releases 0.3.11, the pull request's check passes and the push check now fails; a squash whose description carries the trailer still passes both until base-rs/bake-releases#39, which counts a trailer only on a commit that writes exactly the generated notes, is released and the lockfile updated, after which all three fail. Repositories that copy this job, as the base-project-setup and base-project-update skills say, get the change from here. CI also needed one unrelated fix: the runner's newer Cargo reports base-design and base-security as unused dependencies, which CARGO_BUILD_WARNINGS: deny turns into an error, so src/lib.rs now links them with use … as _; as it already does the other crates that supply agent context.

[ioquatix]

Independent review: no must-fix findings. I pushed one commit that requires bake-releases 0.3.12 and locks it, so this repository's own releases job uses the stricter trailer rule.

• BAKE_BASE resolves correctly for every event. Pull requests get HEAD^1, merge groups get merge_group.base_sha (before is not set there), and pushes get before. A multi-commit push judges every non-merge commit in the range. Manual runs and an all-zero before check nothing. That is right for manual runs: checking against the last tag would re-judge history that already landed, including releases made before the notes were generated. fetch-depth: 0 brings the tags that the regeneration needs.
• With 0.3.12, a forged Release: v0.1.13 hand edit on top of this branch is rejected, while the real v0.1.12 release commit abd59d7 still passes.
• The use base_design as _; and use base_security as _; lines are correct and the smallest possible fix. Both crates are context-only dependencies, linked the same way as the others. Tests pass with CARGO_BUILD_WARNINGS=deny.
• The template matches context/setup.md, which says to copy this releases job into the test-result aggregate.