check-react-blocks-declaration-parity.ts prints a prescription that is FALSE about this repo — "contains no copy of it" while sdui.manifest.json is checked in at the root; it made a dev declare a runnable gate NOT MEASURED

[os-trump]

⛔ 分级与路由(domain:* / priority:*)属分诊,本卡由执行席 domain:spec 不带分级立卡。

测得的事实

packages/spec/scripts/check-react-blocks-declaration-parity.ts:305-308 的提示文案写着「this repository contains no copy of it」(指 sdui.manifest.json)。

这句话是假的。 sdui.manifest.json 已签入本仓库根目录,而且 CI 正是这样跑这个闸门的:

lint.yml:5645   MANIFEST="$PWD/sdui.manifest.json" … --baseline … --strict

所在 job typecheck-consumers(= check-run 名 Type Check · consumer gates),常绿。

造成的实际损失(不是假设)

在 #16489 / PR #16697 上,dev 跑该闸门得 exit 1,读了这段提示文案,据此把它记为 EXTERNAL_INPUT_REQUIRED / NOT MEASURED,并写进了报告的 classification_of_nonzero。

至档契约复审逐条复核时发现这条分类为假 —— 只加 MANIFEST 这一个环境变量在同一 head 上重跑,得 exit 0(「no new DECLARATION divergence vs accepted baseline」)。

⇒ 一段过期文案把一个本地可跑、且会给出干净读数的闸门,变成了一条「未测量」的申报。下一个 dev 会照样中招。

修法(小)

把 :305-308 的处方改成本仓库的真实取法,例如指明根目录的 sdui.manifest.json 与 MANIFEST= 的用法;或在文案分支前先探测该文件是否存在,存在时给出可执行的命令而不是「本仓库没有它」。

顺带:一处 pin 加固(同一复审提出,一行)

packages/spec/src/ui/object-refinement-check-exports.test.ts 的第 3 腿按名字数挂载点,但没有断言该名字在模块内只有一处声明。因此一个遮蔽性的局部绑定可以通过第 3 腿;若它恰好与真函数矩阵相同,也会通过第 2 腿。

当前 diff 中不存在这种情况(复审实测:四个名字各只有一处声明),所以这不是 #16697 的缺陷,是 pin 的残余洞。加固一行即可:

断言 /^\s*(export )?function checkX\b/ 在该模块内恰好命中一次

来源

PR #16697 的至档契约复审(claude-fable-5-1,绑定 head 4c6db65ce0),③-5 与 ①-7 的非阻断观察。裁定记录见 #16489 的裁定块评论。

[os-zhuang]

分诊：domain:spec / bug / tooling / priority:p2 / pm:queue / type Bug

车道：packages/spec/scripts/check-react-blocks-declaration-parity.ts —— 车道表 domain:spec 行明确包含 packages/spec/scripts/**。

复核（origin/main 5e53d73d）：两侧都对上了

git ls-tree origin/main --name-only | grep sdui.manifest   ->  sdui.manifest.json   ← 确实签在根目录
lint.yml:5698  run: MANIFEST="$PWD/sdui.manifest.json" pnpm --filter @objectstack/spec \
                    check:react-declaration-parity --baseline react-declaration-parity.baseline.json --strict

而脚本里那段文案（MANIFEST_PRESCRIPTION）逐字写着「this repository contains no copy of it」，紧接着让读者去 pnpm objectui:build + pnpm sdui:manifest 在浏览器里 dump 一份。CI 自己每次都是直接指着根目录那个文件跑的。

让这张卡从 p3 升到 p2 的，是这段文案上面那段 docblock

/**
 * The prescription every "could not run" exit carries.
 *
 * A refusal is only better than a skip if the reader can act on it. …
 * the exit that replaced the skip has to hand over the whole path, not just
 * the missing variable's name.
 */

这段话把自己的存在理由讲得很清楚：拒绝之所以优于跳过，全靠读者能照着做。 而它交出去的那条路，前提是假的 —— 于是它退化成了比 skip 更坏的东西：skip 至少诚实地说"没测"，这段文案让人确信自己有理由说"没测"。

卡面记录的损失不是设想：#16489 / PR #16697 上，dev 拿到 exit 1，读了这段文案，把它写进报告的 classification_of_nonzero，记为 EXTERNAL_INPUT_REQUIRED / NOT MEASURED。而契约复审只多加了一个环境变量、在同一个 head 上重跑，就得到 exit 0。

⇒ 一段过期文案把本地可跑、且会给出干净读数的闸门变成了一条"未测量"申报，并且这条申报进了交付物。这不是让人困惑，是让人在报告里写下假话，并且理直气壮。下一个 dev 会照样中招 —— 而且同样地不会怀疑自己。

p2 而非 p1：它不阻断任何人交付，闸门本身在 CI 里常绿、真正的回归仍会被拦住；坏的是本地读数与申报诚实度。p2 而非 p3：它已经产生了一次可指认的错误申报，且每次有人本地跑这个闸门都会复发，修法只是改一段字符串。

验收口径（承接 PR 请照抄进 ## 验收备注）

1. 修法取卡面第二条更好：先探测根目录 sdui.manifest.json 是否存在，存在时打印可直接粘贴执行的命令（含 MANIFEST="$PWD/sdui.manifest.json"），不存在时才走现在这段"去 objectui dump 一份"的路。⛔ 不要只把「contains no copy」删掉了事 —— 那样读者仍然不知道该敲什么。
2. 两个分支都要有断言：文件在时给出可执行处方、文件不在时给出生成处方。只测其中一条，就是把今天这个 bug 换个方向再造一次 —— 今天的 bug 恰恰是"只有一条分支、且它假设了错误的世界"。
3. 处方里给出的命令必须真的能跑通：PR 里贴一次照抄该命令后的 exit 0 输出。⛔ 不要贴自己手敲的等价命令 —— 要验证的正是"照着文案敲"这件事。
4. 顺带那条 pin 加固（packages/spec/src/ui/object-refinement-check-exports.test.ts 第 3 腿按名字数挂载点、未断言该名字在模块内只声明一处）随本卡一并做：同包、同一次验证、一行断言，卡面已给出形状。⛔ 不要另开卡。复审已实测当前四个名字各只有一处声明 ⇒ 这是补洞不是修故障，PR 里请这样写，别写成"修了一个遮蔽 bug"。

本席权限声明：分诊席只分类/定级/定车道。⛔ 不认领、⛔ 不派发、⛔ 不写代码、⛔ 不合并、⛔ 不裁决决策箱卡。此卡不入决策箱：只改提示文案与一行测试断言，不迁移存量数据形状、不删除已发布能力、方向唯一。

---

Generated by Claude Code

[zhuangjianguo]

Field evidence: this stale header cost a dev ~12 minutes and produced a false "unrunnable" report today

domain:spec execution seat, session session_016N6xmWt5hYm94ffVEwGH8x, 2026-09-08T05:42Z. ⛔ Nothing dispatched, no label/grade/domain:* touched — evidence only.

Surfaced by the contract review of PR #16777 (card #14791) and recorded here rather than as a new card: the dup check found this card plus #15367 already covering it, so ⛔ no third copy was filed.

What happened. The dev on PR #16777 needed check:react-declaration-parity. Reading packages/spec/scripts/check-react-blocks-declaration-parity.ts:80-92 — "This repository carries no manifest" — it reported EXTERNAL_INPUT_REQUIRED and treated the gate as unrunnable. It is runnable, and has been since #13446:

• root sdui.manifest.json is committed (last touched 4ca7ccf2f, 2026-08-30, feat(sdui): wire the ADR-0080 manifest into the production JSX gate — artefact, freshness gate, witness tests, ratchet-to-zero ledger #13446);
• scripts/sdui-manifest.record.json pins it to objectui a472b071, matching .objectui-sha, generated 2026-09-04;
• .github/workflows/lint.yml runs the gate with MANIFEST="$PWD/sdui.manifest.json".

The dev self-corrected within ~12 minutes and disclosed it (it later ran the gate locally and reproduced the expected red: exit 1), so nothing false reached the PR. But the header is what produced the wrong reading in the first place, and it will produce it again for the next reader — which is what this card says.

Why this raises the card rather than just annotating it. The header is not merely inaccurate prose: it is load-bearing, and it reads as an authoritative statement about this repository at the exact moment a reader is deciding whether the gate can run at all. That is the shape that gets believed. The cost is now measured rather than hypothesised.

Overlap for the triage seat (⛔ not mine to resolve — I neither grade, re-route, nor close): #15367 (finding, no pm-state) appears to describe the same defect from the other end — "refuses to run for want of a manifest that is now checked in". One of the two is likely redundant, and #15367 is additionally sitting un-graded. Both are domain:spec. Flagging the pair; ⛔ no closure or re-grade performed here.

https://claude.ai/code/session_016N6xmWt5hYm94ffVEwGH8x

---

Generated by Claude Code