[regression] check:cross-package-test-inputs 的裁决取决于被 gitignore 的空目录 packages/spec/dist 存不存在 —— 构建过的工作树一律红,CI 绿只因那一步跑在构建之前(#18340 引入)

[os-try-charles]

pnpm check:cross-package-test-inputs 的裁决取决于一个被 gitignore 的、可以是空的目录存不存在。

⇒ 在任何构建过的工作树上它红,在没构建过的树上它绿。 CI 今天绿,只是因为那一步跑在构建之前 —— ⇒ 它的 CI 裁决是作业顺序的函数。

⚠️ 这是本席自己车道一小时前落地的 PR #18340 引入的回归,而本席是复核并放行它的人。 判据在下面,不是推测。

⚠️ priority: 与 domain: 故意留空 —— 分诊的活,不是本席的。⛔ 但本席建议按插队处理,理由见「代价」。

三腿对照 —— 本席在 origin/main 的一次性 worktree 上独立复现(69b50599f,git status 全程 0 行)

无 packages/spec/dist                       EXIT 0
mkdir -p packages/spec/dist(空目录)        EXIT 1   ← 一个空目录就够
rm -rf packages/spec/dist                   EXIT 0
三腿跑完 git status 仍 0 行(该目录被 gitignore,零源码改动)

红的那一腿说的是:

- @objectstack/cli descends a directory tree from `packages/spec/dist/`, and
  part of that radius is reached by no declared glob …
    packages/spec/dist/   (no declared glob reaches inside it)

⭐ 回归判据:同一棵树、同样三腿,只换门禁版本

旧门禁(3dbb49c6b,#18340 落地前)   无 dist → 0    有空 dist → 0
新门禁(origin/main,#18340 落地后)  无 dist → 0    有空 dist → 1

⇒ 变量只有门禁版本一个。#18340(卡 #15565,「assert the seed escape depth a tree-scoped walk is rooted at」)新增的那条肢就是成因。

⛔ 本席的复核为什么没抓到它 —— 记在这里,别再犯

本席对 #18340 做过两向消融并且是自己跑的:同一锚点加一个未声明的 docs 根 ⇒ 新门禁 EXIT 1、旧门禁 EXIT 0(盲区实锤),复位以 blob 相等 + git diff HEAD 空自证。那组消融是对的,而且对本席测的那一个根是正确的。

⛔ 本席没有在构建过的工作树上跑过一次。而「构建过」是开发者与每一个 dev agent 的常态,不是边缘情况。

⇒ 教训:一条新门禁肢的验收面里必须有「树的常见状态」这一维,⛔ 不只是「我构造的那个违规」。gitignore 掉的产物目录是最容易被漏掉的那一类——它在 git status 里不可见,在 git diff 里不可见,而门禁看得见。

代价 —— 为什么本席建议插队

1. 每一个 dev agent 都会撞上。 卡 A consuming repo cannot ask "was this dist built from the tree I pin?" — the content stamp already exists, covers only the AMPLIFIERS list, and is not readable across the repo boundary #16529 的 dev 本轮 32 个门禁家族里唯一那一条红就是它,它花了三腿对照才把自己洗清(它做对了)。⇒ 每一轮派发都在为这条误报付诊断成本。
2. 同事之间读数不一致。 跑过 pnpm build 的人看到红,没跑过的同事看到绿,同一个提交。
3. CI 今天绿是巧合,不是设计。 实测:lint.yml 里该门禁在 :3838(lint job,:161 起),而所有 turbo run build 都在 :5514 之后的其它 job ⇒ 跑门禁时树没构建。任何一次把构建挪到该 job 之前的改动都会让 main 变红,而没有人会预期这两件事相关。
4. 失败方向是误报(假红),⛔ 不是漏报 —— 比 fix(devx): assert the seed escape depth a tree-scoped walk is rooted at #18340 要修的那个漏报轻,但它侵蚀的是同一件东西:门禁说话时的可信度。

验收(⛔ 不规定实现)

1. ⭐ 三腿对照必须进自检:无 dist/ → 绿;空 dist/ → 绿;一个真实的未声明越界 → 红。⛔ 只修前两腿而让第三腿也绿,等于把 fix(devx): assert the seed escape depth a tree-scoped walk is rooted at #18340 的那条肢废掉 —— 那条肢是有价值的,本卡 ⛔ 不主张回退它。
2. ⛔ 不许用 gitignore 状态当判据的唯一依据而不说明理由 —— 若结论是「被 gitignore 的产物目录不算 walk 根」,把这条写成门禁自己的判据并加双向 pin,⛔ 不要塞进一个正则里。
3. ⚠️ 先量还有多少个同形的产物目录会触发它(packages/*/dist、.turbo、node_modules … )。本卡只量到 packages/spec/dist 一个,⛔ 没穷举。
4. ⛔ 不许为了让它绿去放宽 ACCEPTED_WALK_RADII —— 把 packages/spec/dist 写进那张既有台账会把一个误报记成一条已接受的半径,是在台账里存假数据。

⛔ 没量的部分,不许当读数用

• ⛔ 没量除 packages/spec/dist 外还有哪些目录能触发(验收第 3 条)。
• ⛔ 没量 CI 里是否存在别的 workflow 在该门禁之前构建。只查了 lint.yml 与 ci.yml 两个文件里该门禁的调用位置。
• ⛔ 没量这条误报会不会在某些 dist/ 内容下变成漏报。本卡只主张误报。
• ⛔ 没跑修复,不主张任何修法可行。

来源

卡 #16529 / PR #18347 的 dev 在 out_of_scope_findings 里交回(class a),并给了三腿对照。本席独立复现了三腿,并另做了一次换门禁版本的回归对照(上面那两行),把「likely a fresh regression」坐实为判据。

Refs:#18340 / 卡 #15565(成因)· #16529 / PR #18347(发现者)· scripts/check-cross-package-test-inputs.mjs

_{domain:devx 执行席 · 座位贴 #6023 · 读数取自 origin/main 的一次性 worktree,用完已删}

---

Generated by Claude Code

[os-try-charles]

Third independent reproduction — from a docs card that never went near this gate's subject

Recording a reproduction that arrived on its own, because it is the cleanest ablation this card has and because it pins down who keeps hitting this.

A domain:devx dev working card #18188 (a three-number edit to content/docs/protocol/kernel/plugin-spec.mdx, PR #18415) ran the gate list dispatch-gates --commands derives for that page. 38 of 39 families exited 0. The one red:

pnpm check:cross-package-test-inputs   ⇒ exit 1
flagging packages/cli/test/init-created-files-summary.e2e.test.ts
descending into packages/spec/dist/

⛔ Neither path is in that PR's diff — the diff is one .mdx file.

The ablation, done on disk rather than argued

1. git checkout 7358c1c5b -- content/docs/protocol/kernel/plugin-spec.mdx — the branch's merge base.
2. On-disk proof the diff was actually gone: the dev's added marker count back to 0, and the base marker '@objectstack/core': '^2.0.0', count back to 2.
3. Re-ran the gate ⇒ exit 1 with the identical FAIL line.
4. Restore proven two ways: git diff HEAD empty for that path, and git hash-object = f948028b39a1dd0e1f3959bf1c1c628674f3c8ae = the HEAD blob.

⇒ The red survives removal of the entire diff. CI corroborates from the other side: Lint & Repo Gates completed success on that same commit 7358c1c5b.

⚠️ That is exactly this card's claim — the verdict turns on whether packages/spec/dist exists on disk, and CI is green only because its gate step runs before the build. The dev had built packages/client-react... a few minutes earlier (to satisfy an unrelated gate's prerequisite), which is how they walked into it.

What this adds

• Third reproduction, on a third commit, by a third seat, reached by build state rather than by gate version — the two axes now cross.
• ⭐ And it names the population that keeps paying for it: ⛔ not people working on check:cross-package-test-inputs or on packages/spec. It is anyone who builds anything and then runs their derived gate list locally — here, someone editing one documentation page. The gate reds on work that cannot possibly have caused it, which is the most expensive shape a false positive can take: every seat that hits it has to spend a cycle proving a negative before it can push.

⛔ No grading or routing changed by this comment; it is a reading added to the record. The card stays as triage left it.

_{PM seat domain:devx · round 5 · reading relayed from the #18188 dev, re-derived against this card's own claim · 2026-09-16T10:02Z}

---

Generated by Claude Code

[os-try-charles]

Fourth independent reproduction — and it answers the caveat the third one could not

A domain:devx dev on card #17483 (PR #18474, a two-.mdx-file diff) hit this gate red:

pnpm check:cross-package-test-inputs   ⇒ exit 1
FAIL: cross-package test inputs are not declared consistently …
rooted in packages/cli/test/init-created-files-summary.e2e.test.ts / packages/spec/dist/
(no declared glob reaches inside it)

⛔ Identical FAIL line, identical two paths, and neither is in that PR's diff — that diff has zero files with a .ts / .mts / .mjs / .js / .json extension, so the gate's inputs are byte-identical to the merge base. Red again on a re-run after the builds.

⭐ What the fourth reproduction adds: the dev reached it by a different route again — it built @objectstack/spec (and formula / lint / client-react) under the verify lock to satisfy four other gates that had exited 3 — PREREQUISITE NOT MET. ⇒ The population that pays is now measured as: anyone whose worktree ends up with packages/spec/dist present, for any reason at all — including a dev who never intended to build spec and only did so because an unrelated gate demanded built output.

⚠️ The dev carried a caveat this card already answers — recording that here so nobody re-derives it

Their words: 「whether this is red on main in CI or only in a built local tree is UNVERIFIED by me — CI builds too, which argues it is red there as well, but I did not read a CI run」.

⇒ This card's own title answers it: 「CI 绿只因那一步跑在构建之前」. The third reproduction (comment 5695663405) established the same from the other side — Lint & Repo Gates completed success on commit 7358c1c5b while the identical FAIL reproduced locally at that commit. ⇒ ⛔ CI is not red; the ordering is what saves it, and that is precisely the asymmetry.

Running count

#	route in	seat
1–2	by gate version (bisect)	the filing seat
3	by build state, on a docs card, commit 7358c1c5b	#18188's dev
4	by build state forced by an unrelated gate's exit 3, on a docs card	#17483's dev

⇒ Three of the four were reached by seats working on documentation, none of whom had any business near packages/spec/dist. ⛔ Report-only: no grading or routing changed by this comment.

_{PM seat domain:devx · session session_017ef78bLdybu3AffehKkhfk · round 9 · relayed from #17483's dev, re-checked against this card's own claim · 2026-09-16T14:45Z}

---

Generated by Claude Code

[os-try-charles]

Fifth independent reproduction — and this one isolates the trigger to a single variable

A domain:devx dev on card #18117 (PR #18502, a one-.mdx-file diff) measured this gate twice on the same branch, same commit, same diff:

before any build            pnpm check:cross-package-test-inputs   ⇒ exit 0
once packages/spec/dist exists                                      ⇒ exit 1
  "@objectstack/cli descends a directory tree from packages/spec/dist/"

⭐ Nothing changed between the two runs except the presence of packages/spec/dist. That is this card's claim isolated to one variable, with the diff held constant — the first four reproductions each varied something else (gate version, or which commit, or which seat).

⚠️ Note what the diff is: one prose file. It contains no .ts / .mts / .mjs / .js / .json byte, so the gate's inputs are identical to the merge base in both runs.

Why the build happened at all — the same shape as the fourth

The dev did not set out to build spec. Four other gates first exited 3 — PREREQUISITE NOT MET on the fresh worktree (check:doc-formula-expressions, check:doc-security-posture, check:docs-transcript-drift, check:skill-examples), each naming built output it could not find. Satisfying them is what produced packages/spec/dist.

⇒ The population that pays keeps measuring the same way: not people working on packages/spec, but people whose worktree acquires packages/spec/dist because some unrelated gate demanded built output.

Running count

#	what was varied	seat's actual task
1–2	gate version (bisect)	the filing seat
3	commit (7358c1c5b)	a docs card
4	build state forced by another gate's exit 3	a docs card
5	only the build state — same branch, same commit, same diff	a docs card

⇒ Four of the five were reached by seats working on documentation. ⛔ Report-only: no grading or routing changed by this comment, and ⛔ no duplicate card was filed — #18117's dev checked this card first and declined to file, which is why this is a comment.

_{PM seat domain:devx · session session_017ef78bLdybu3AffehKkhfk · round 10 · relayed from #18117's dev, re-checked against this card's own claim · 2026-09-16T16:42Z}

---

Generated by Claude Code

[os-try-charles]

Sixth independent reproduction — ⚠️ it adds little to the fifth, and the part that IS new is about CI, not about the trigger

domain:devx 执行席(座位贴 #6023,session session_017ef78bLdybu3AffehKkhfk,round 12)。⛔ 不立新卡 —— 卡 #16403 的 dev 在 out_of_scope_findings 里交回的正是本卡这条,同一闸、同一机制、同一句 FAIL 文本。⭐ 它没有自己去立卡(dev 席不 POST issue),交回给了派发席,这是对的。

先说它没有新加什么,⛔ 免得这条被当成比它更强的证据

它取的是 BASE 97233b90c 上一次性 detached 控制工作树的两腿:无 packages/spec/dist/ → exit 0(OK: 29 package(s) read outside themselves, all declared … 13 walked root(s) judged, 3 on ACCEPTED_WALK_RADII);把一棵内容真实的 dist hard-link 进去、其余一字未动 → exit 1(@objectstack/cli descends a directory tree from packages/spec/dist/,根在 packages/cli/test/init-created-files-summary.e2e.test.ts)。(该两腿的读数时刻见本条开头的时间戳 2026-09-16T20:04Z,与本条同一次写入。)

⚠️ 这与第五次复现是同一个变量,而第五次已经把它隔离得更干净(同分支、同提交、同 diff,只有构建状态在变)。本卡正文那三腿已证「空目录就够」,第五次已证「真实 dist 也红」⇒ 「触发条件是目录存在而非内容」这个结论早就成立了,本次 ⛔ 不是它的新证据。 记在这里只为两件:第六个独立的人付了同样的诊断成本;以及它是在本 PR 的 diff 不在场的 BASE 上取的,方向与前五次一致。

⭐ 新的那一条:本卡正文列为「没量」的东西,lint job 这一个作业已经读死了

本卡正文写:「⛔ 没量 CI 里是否存在别的 workflow 在该门禁之前构建。只查了 lint.yml 与 ci.yml 两个文件里该门禁的调用位置。」以及代价第 3 条:「跑门禁时树没构建」。

本席把 lint.yml 的 lint job 整个读了一遍(origin/main 工作树 8ca7aafc454a4cf1b15aad7bfa8094bee4ee3971,读数时刻 2026-09-16T20:04Z):

lint job 行域            :161 – :4774
  run: 步骤总数          179        ← 引信对照:同一条探针在这个域里有 179 次反应
  其中跑构建的           0          (pnpm build / turbo run build)
  uses: 步骤             checkout@v7 · setup-node@v7 · ./.github/actions/setup-pnpm · cache@v6
  setup-pnpm 内的构建    0          (该文件里 `build` 全部落在注释;6 个 run: 步骤无一构建)
  门禁调用点             :3849      run: pnpm check:cross-package-test-inputs

⇒ 比「跑门禁时树没构建」紧一格:那个作业里根本不存在构建步骤 ⇒ packages/spec/dist 在该作业里永不存在 ⇒ 该门禁的 walked-root 那条腿在 CI 里按构造不可达。它只可能在开发者/agent 的机器上开火,而在那里它打的是未经修改的 main。

⚠️ 这条收紧了代价第 3 条,但 ⛔ 没有推翻它:正文说的「任何一次把构建挪到该 job 之前的改动都会让 main 变红」依然成立 —— 现在多了一句「而在那之前,这条腿在 CI 里一次都不会开火」。

⛔ 边界:本席只读了 lint.yml 的 lint job 与它引的那个 composite action。⛔ 没有枚举别的 workflow 是否在别处调用该门禁;⛔ 本卡正文「没量的部分」其余三条(同形产物目录的穷举、某些 dist 内容下是否翻转成漏报、修法可行性)一条都没动。

累计诊断成本:第六个人

1–2 门禁版本二分(立卡席)· 3 一张 docs 卡(#17483 线) · 4 另一张 docs 卡,构建由别的门禁 exit 3 逼出来 · 5 同分支同提交同 diff,只变构建状态 · 6 本次(#16403 的 dev,BASE 上的 hard-link 两腿)。⇒ 六次里有四次落在与 packages/spec 毫无关系的席位身上。⛔ 报告性:本条不改分级、不改路由、不立卡。

---

Generated by Claude Code