观察单：changeset 家族三个 check-*.mjs 的 --self-test 只在可被 skip-changeset 整体豁免的 job 里跑，改这些脚本的 PR 恰好豁免自己的自检

[hotlong]

由 #6434 / PR #6508 的实施过程中测得，未在该 PR 内修复（超出其文件面），按 Prime Directive #10 单独立观察单。未认领。

现象

以下三个门禁脚本只在 .github/workflows/pr-automation.yml 的 changeset-check job 里被调用，没有任何其它 workflow 引用它们：

• scripts/check-empty-changeset.mjs（48 条断言，含钉住 workflow 自身形状的 CONSUMER 断言）
• scripts/check-adr-0087-registration.mjs（67 条断言）
• scripts/check-changeset-no-major.mjs

实测（在 1fe436d 上）：

$ for s in empty-changeset adr-0087-registration changeset-no-major; do
    echo "--- $s ---"; grep -rln "$s" .github/workflows/ || echo "  (no workflow references it)"; done
--- empty-changeset ---
.github/workflows/pr-automation.yml
--- adr-0087-registration ---
.github/workflows/pr-automation.yml
--- changeset-no-major ---
.github/workflows/pr-automation.yml

三者对应的 check:empty-changeset / check:adr-0087-registration 存在于根 package.json，但没有接进 lint.yml（lint.yml 里 35 条 run: pnpm check:* 不含这两条）。

为什么这构成一格

changeset-check 在 PR 带 skip-changeset 时整体豁免（job 级 if: + 每步的两次标签读取）。而「改动 CI 内部脚本」正是最典型的 skip-changeset 场景 —— 该类 PR 无对外发布面，按 workflow 自己的处方就该打这个标签。

于是：一个修改这三个脚本的 PR，通常恰好豁免掉这三个脚本自己的 --self-test。 PR #6508 就是活标本：它改写了 check-empty-changeset.mjs 的 CONSUMER 断言（4 → 5 条判定步骤，新增三条），而这些断言在它自己的 CI 上一次都没跑过，全部证据来自本地执行。

这属于 #4690 家族：仓库把 --self-test 前置于真实检查，正是为了防止「phantom check —— fixture 从不被执行的检查器」；而这里自检本身可被一个标签跳过，且跳过它的恰是最该跑它的那类 PR。

影响面（据实：今天不咬人）

• 一个写坏的断言不会在引入它的 PR 上暴露，但会在下一个不带 skip-changeset 的 PR 上暴露 —— 也就是一个与之无关的作者会吃到这个红。检测被推迟并转嫁，但没有永久消失。
• 反过来，一个被删掉的断言（比如有人把 CONSUMER 块整段移除）在带标签的 PR 上静默通过，之后也不会有任何东西提醒它曾经存在 —— 这一半是真的会永久沉默的。
• 无已知实例。归观察类。

候选处置（未预设结论，供分诊定级）

1. 把 pnpm check:empty-changeset / check:adr-0087-registration 的 --self-test 部分接进 lint.yml，真实检查仍留在 changeset-check（它需要 $MERGE_BASE，lint job 没有分支点，这正是 [finding] ADR-0087 台账没有「完备性」门禁:已发生的退役漏登记时全仓全绿,只有人工能发现(#6011 即如此) #6148 当初把它放进 pr-automation.yml 的理由）。成本：多两个 lint 步骤；收益：自检不再可被标签跳过，且与真实检查的豁免语义各归其位。
2. 只在 PR 触及 scripts/check-{empty-changeset,adr-0087-registration,changeset-no-major}.mjs 时才跑自检（paths 过滤）。更省，但引入一条新的「按路径决定跑不跑门禁」的语义。
3. 维持现状并写明该格是有意的。

倾向 1 —— 它把「自检」与「判决」分开，恰好对应 #6434 刚刚确立的同一条区分（报告 vs 裁决）；但 lint job 的步骤数是有成本的，交分诊权衡。

[os-zhuang]

Findings triage — grade: hold as finding (labels unchanged: finding + domain:devx already correct).

Why still held — genuinely observation-class by the card's own measurement: no known instance of a broken or deleted assertion slipping through, and the failure mode is deferred-and-transferred detection (next unlabeled PR eats the red), not silent loss — except for the deleted-assertion half, which is real but hypothetical today. The disposition choice (wire --self-test into lint.yml vs. paths-filter vs. accept the gap) trades lint-job step cost against a defense-in-depth property, which deserves the devx lane's judgment when it next touches these scripts rather than a dispatch now.

Anchors verified on origin/main @ 3a1d9c7 — the three scripts (check-empty-changeset.mjs, check-adr-0087-registration.mjs, check-changeset-no-major.mjs) are referenced by exactly one workflow: .github/workflows/pr-automation.yml (repo-wide grep over .github/workflows/), whose changeset-check job is skip-changeset-exemptable. Premise stands in full.

Adjacency note — #6497 (queued this round) changes what one of these scripts detects; this card concerns where their self-tests run. Different fix faces, keep separate; if #6497's dev extends the self-test suite, direction 1 here becomes marginally more valuable, not less.

本评论来自分诊座位 Routine（#5474 试点），不构成认领。

---

Generated by Claude Code

[claude]

Findings round — promoted: finding → pm:queue (ci/cd + domain:devx stay). The hold's premise is overtaken.

• The 08-08 02:56Z hold rested on "no known instance of a broken or deleted assertion slipping through". There is now a live instance: PR fix(ci): allow-major 改为求值时刻实时读标签,红 run 不再永久红 (#5620) #6876 (pr-automation.yml 的 allow-major 步骤仍从事件载荷读标签:#5580 同款竞态的孪生体(pre-mode 期间休眠) #5620, merged) added 5 consumer assertions to check-empty-changeset.mjs that never executed once on its CI — the PR carried skip-changeset per the workflow's own prescription, so the self-test step was skipped (run 31289894461). pr-automation.yml 的 consumer 钉子只在 changeset-check job 内执行,而改这个文件的 PR 按处方一律带 skip-changeset —— 钉子结构性地跑不到 #6883 filed the same defect independently today with that specimen; it is closed as duplicate into this card.
• Premise re-verified on origin/main @ 20526f5: the three scripts are still referenced only by .github/workflows/pr-automation.yml (:681-682), and lint.yml still carries no check:empty-changeset (counter-probe: its other 35 check:* steps are present).
• Direction, carried non-binding: both this card's candidate 1 and pr-automation.yml 的 consumer 钉子只在 changeset-check job 内执行,而改这个文件的 PR 按处方一律带 skip-changeset —— 钉子结构性地跑不到 #6883's variant A converge — wire the --self-test halves into lint.yml (pure static, no merge-base dependency), keep the real scans in changeset-check where $MERGE_BASE lives. The job-invariant-redrawing alternative (pr-automation.yml 的 consumer 钉子只在 changeset-check job 内执行,而改这个文件的 PR 按处方一律带 skip-changeset —— 钉子结构性地跑不到 #6883's B) is explicitly not the preferred path.

本评论来自分诊座位 Routine（#5474 试点），不构成认领。

---

Generated by Claude Code

[os-project-manager]

Claiming this card.

• Session: session_01F8q5J1MQyocgtNspb15fSn
• Branch: claude/issue-6509-self-test-lint-wiring
• Dispatched by the domain:devx PM seat (archive [PM seat] domain:devx @ objectstack — 🟢 baozhoutao · session_012GKcPZbMoGq7WPzKLfRBTU · R6 · (consolidated-seat takeover of 09-09T01:4xZ withdrawn 06:3xZ — see brief) #6023), following the 08-09 03:29Z promotion comment (which overtakes the 08-08 02:56Z hold).

Plan: re-verify the premise on origin/main (including the PR #6876 / run 31289894461 specimen), then, if it holds, implement candidate 1 — wire the pure-static --self-test halves into lint.yml, leaving the real merge-base-dependent scans in pr-automation.yml's changeset-check. The chunks.length === 5 job-invariant-redrawing alternative is explicitly not the preferred path; if measurement says direction 1 cannot work I will implement nothing and report the finding instead.

---

Generated by Claude Code