ci: report release-template skips as SKIP, not OK

[justinjoy]

Refs #1288. Does not close it — two of its four acceptance criteria remain open, see below.

The defect

scripts/ci/check-release-template.sh had five exit 0 paths: one real pass and four skips. Meson reads exit 0 as a pass, so the gate reported abi / release_template OK on every run while asserting nothing.

After this change:

1/2 abi - wirelog:release_template          SKIP  exit status 77
2/2 abi - wirelog:release_template_selftest OK

Meson is the only caller (tests/meson.build), so nothing sees 77 under set -e. A bare workflow run: step would fail on it, and the header now warns against that.

What this does not fix

The gate enforces nowhere in CI, and this PR does not change that. Two independent blockers:

1. no job supplies a GH_TOKEN, so gh release view returns nothing;
2. even with a token, the release does not exist when the Tag / ABI job runs. Nothing in this repo publishes one — a maintainer runs gh release create by hand (docs/RELEASE_PROCESS.md steps 4–5; automation is Route release artifact publishers through the release-tag verification gate #1152). If feat(release): sign, publish, and document release artifacts #1289 lands, publication moves into a job that needs: release-verification, which needs: abi — so the gate still runs first.

PR #1289 fixes neither. It adds GH_TOKEN only to the artifacts job, and its own RELEASE_PROCESS.md text tracks the gate back to #1288. An earlier draft of this PR claimed #1289 would wire it; that was wrong and is corrected. Do not assume a token alone switches enforcement on.

Acceptance criteria

#	criterion	status
1	a differing release body fails CI	open — needs release-tag.yml (#1289 owns it)
2	genuine skips report SKIP, never OK	done
3	docs/RELEASE_PROCESS.md describes actual enforcement	open — #1289 owns the file
4	a test covers the mismatch case	done

The self-test

Copies the real gate into a fixture repo and drives all seven terminal outcomes, asserting exact statuses — ! cmd would conflate 1 with 77, which is the distinction at issue. The mismatch case and the missing-CHANGELOG failure had never executed; demoting either to exit 0 previously left the suite green.

Three source-level assertions cover what behaviour cannot reach: exactly one exit 0, it is the success path, SKIP_EXIT is 77. They deliberately do not key on skip wording or count echo-vs-exit sites — that shape false-failed a refactor to the skip() helper run-doop-perf-gate.sh already uses, while a differently worded skip escaped it.

Fixture isolation took more care than the fix. GIT_DIR/GIT_WORK_TREE override git -C rather than supplementing it, and git exports them inside hooks, git bisect run and git rebase --exec — an earlier draft committed the caller's uncommitted work to their real repository and planted a v9.9.9 tag, a well-formed vX.Y.Z that would put this gate into enforcement against a nonexistent release. Reproduced on a victim repo, then fixed. The fixture also survives commit.gpgsign, core.excludesFile, core.hooksPath (--no-verify skips pre-commit, not reference-transaction), init.templateDir, core.autocrlf+core.safecrlf, and the identity variables — each of which otherwise fails a correct tree naming neither the gate nor the issue.

Known, deferred

A CHANGELOG blank line carrying trailing whitespace survives the command substitution that strips trailing newlines, so the extracted section gains a line the body never had. Pre-existing, fails in the safe direction, and fixing it means perturbing comparison logic an exit-code change should leave alone. Recorded on #1288.

Validation

• self-test: 22 assertions
• meson test -C build --suite abi: 41 Ok / 0 Fail / 1 Skipped
• full suite: 304 Ok / 0 Fail / 13 Skipped
• mutation-tested: SKIP_EXIT=0, reworded skip at exit 0, both 7th-outcome demotions, v[0-9]*→v*, norm→cat, asymmetric norm all fail; a genuine skip() refactor and a comment inserted between the OK echo and its exit both still pass
• portability: verified on a PATH farm without tac (GNU-only, absent on macOS/BSD)

Three independent review rounds; final verdicts Architect approved, Critic approved, Reviewer approved.