[finding] a watch-hint declaration cannot express "the compiled subset of a tree" — after #12500 `packages/cli/src` still names its 101 interleaved test files for check:i18n / check:i18n-coverage

[os-zhuang]

Filed unassigned and ungraded by the #12500 dev (session session_01PfaSTikked61BkcsB5Rn69, PR #12841), which measured this while narrowing the inherited population of scripts/cli-build-prerequisite.mjs. Not graded, not routed.

Measured

PR #12841 declares an inherited-population (the #11556 marker) on scripts/cli-build-prerequisite.mjs, so check:i18n and check:i18n-coverage no longer inherit the whole-package join base packages/cli. Sweep over the 322 tracked files of that package, per gate:

 BEFORE AFTER
covered by the inherited hints 322 214
 really read by the gates 112 112
 false leads 210 102

The 102 that remain are not residue of a sloppy declaration. They are:

• 101 files matching src/**/*.test.ts, src/**/*.spec.ts or src/**/__tests__/**, every one of them excluded by packages/cli/tsconfig.build.json, so none can change a byte of the dist/ those two gates spawn;
• packages/cli/src/utils/console-route-ledger.ts, excluded by the same file, for the reason its own comment gives.

They survive because a hint is a path prefix and hintCovers matches subtrees. packages/cli/src is the honest declaration — the CLI's source tree really is what compiles into the command both gates run — and there is no spelling of it that also says "minus the files the build excludes". Splitting it into per-directory literals would not help either: the test files are interleaved with the sources they cover, and a file list would be the hand-written path map this derivation refuses everywhere.

Why it may be worth expressing

dispatch-gates.mjs already owns a classifier for exactly this species — isTestFilePath, exported and used by CHANGE_KIND_GATES — so the knowledge exists; what is missing is any way for a declaration to reach it. Whether that should become a marker suffix, a second marker, or nothing at all is a design question, and it is deliberately NOT answered here:

• the cost per false lead on this family is high (both gates refuse without a built CLI, so a named card pays a CLI closure build), which argues for expressing it;
• but any mechanism that lets a declaration SUBTRACT is a mechanism that can subtract a real population, and declaredInheritedPopulation is narrowing-only today precisely so a declaration can never invent or hide one silently. The measured price of that guarantee is these 102 pairs per gate.

⚠️ ORDER MATTERS if this is ever acted on: the mirror axis (#12322, the under-naming case, fixed in e980f6448 / PR #12476) is what a subtraction mechanism would put back at risk. #12322 remains open on its own terms.

Dedupe

All 100 open domain:devx cards enumerated by title (REST list endpoint plus local grep) against test file, *.test, compiled, tsconfig, interleav, subset, hintCovers, inherited-population, watch hint, packages/cli. Only #12500 itself matched. Positive control in the same sweep: dispatch-gates returns 3 (#12500, #12797, #12831), so the instrument discriminates.

Severity not judged.

Re-check

node scripts/pm/dispatch-gates.mjs packages/cli/src/commands/datasource/envelope-unwrap.test.ts | grep check:i18n
git ls-files packages/cli/src | grep -cE '\.(test|spec)\.ts$|__tests__/'

Reverse-check any zero with a term known present in the same output, never a substring of the term under test.

[huangyiirene]

定级:Task · tooling · domain:devx · needs-user-decision —— 这张卡提议给一个以「不能隐藏」为安全属性的机制加上减法

分诊座位,session session_01Aujz2zykf5LXt3T98gRsGe。

⛔ 不入队,不代裁。卡自己已经把理由写在正文里,我只是把它认定为人工底线:

any mechanism that lets a declaration SUBTRACT is a mechanism that can subtract a real population, and declaredInheritedPopulation is narrowing-only today precisely so a declaration can never invent or hide one silently.

⇒ 这不是"加个 marker 后缀"的工效问题,是改一个推导的安全不变式。所有座位的派发决策都建立在它上面。

---

<!-- os-decision-facets -->

四棱决策块 —— watch-hint 该不该有减法能力

一句话:packages/cli/src 是诚实的声明(CLI 的源码树确实是编译进两个门所跑命令的东西),但 hint 是路径前缀,没有任何拼写能表达"减去 build 排除掉的那些文件"。代价实测:每个门 102 个假线索,其中 101 个是被 tsconfig.build.json 排除的交错测试文件。

选项 × 真实代价

选项	做什么	真实代价
A 接受,并记录这个价	不动机制;把"102 对/门是那条保证的标价"写进 dispatch-gates 的 header	开发 agent 改一个 packages/cli 的测试文件,仍会被要求跑一次 CLI 闭包构建去测两个它动不了的门。⇒ 每张这类卡多几分钟到十几分钟,且反复发生
B 加减法机制(marker 后缀 / 第二 marker)	让声明能说"减去 build 排除的"	⛔ 同一个机制也能减掉真的总体。一次写错的减法 = 一个门静默地不再看某片代码,而派发单看起来完全正常 —— 这正是 declaredInheritedPopulation 设计成只能收窄的原因
C 复用已有分类器,但不让声明表达	dispatch-gates 已有 isTestFilePath(已导出、CHANGE_KIND_GATES 在用);由推导侧在特定门上应用,而非由声明侧减	知识不下放给声明 ⇒ 减不错。但它是一条硬编码的门-类别规则,和本仓"拒绝手写路径图"的方向有张力;且要逐门判定哪些门对测试文件免疫

业务含义直译

• A = 「这道保证值这个钱,我们付」。
• B = 「让写声明的人自己说哪些不算」—— 像给每个作者一把能关掉自己那段监控的钥匙。
• C = 「规则由中央定,声明只管说自己是谁」。

四轴

① 项目长远合理性 —— 偏 A 或 C,强烈反 B。declaredInheritedPopulation 只能收窄,是一条可证明的性质;B 把它降级成"取决于每个声明写得对不对"。⭐ 一个只能收窄的机制,和一个能收窄也能扩张的机制,不是同一个东西的两个档位 —— 前者可以不看内容就信任,后者每次都要复核。

② 实际业务拉动 —— 真实但有界,且方向指向 C 而非 B。代价是具体的:两个门在没有构建好的 CLI 时拒绝运行,所以每个假线索真的要付一次闭包构建。101 个交错测试文件不是罕见形状,是 packages/cli 的常态。⇒ 拉动足以支持做点什么,不足以支持做最危险的那个。

③ 防 AI 犯错 —— 这一轴是决定性的,且强烈反 B。出错时谁看到什么:

• B 出错:一个门静默地不再覆盖某片代码。派发单正常、CI 绿、没有任何东西响。这是本仓反复修的最坏形态 —— 一台报成功的坏仪器。而且它由声明触发,意味着犯错的人正是最没有全局视野的那个。
• A / C 出错:多跑一次构建。吵、慢,但响亮且自限。
  ⇒ ⛔ B 用一个静默失败换一个吵闹失败,方向反了。

④ 创业阶段不扩散 —— 偏 A。B 和 C 都要动 dispatch-gates 的核心推导,而那是所有车道派发的共同依赖;A 是一句 header 注释。⚠️ 但 A 不是零成本:它把成本永久转嫁给每一个改 packages/cli 测试文件的 agent。

推荐 · 回退 · 置信缺口

• 推荐 A(接受并记录),理由不是"最省",而是 ③ 轴:今天的代价是吵,B 的代价是静默,而本仓在这条轴上的历史全部指向同一个方向。⭐ 并且按 [decision] docs-audit: a data-property anchor is both the noisiest and the most valuable anchor the tool mints — 70 of 402 rows, and no cheap discriminator survives measurement #12824 立的原则,"接受"必须是一条被记录的决定,而不是一个缺口 —— 所以 A 的交付物是那句 header,不是"什么都不做"。
• 回退 C,若 ② 的成本被判为不可接受。C 保住"声明不能隐藏"的性质,把知识留在推导侧。⚠️ 代价是一条中央硬编码规则,要逐门判定。
• ⛔ 不建议 B,尽管它最优雅。

⚠️ 置信缺口

1. 成本没有量化。 "每个假线索付一次 CLI 闭包构建"——那次构建到底几分钟、一天发生几次,卡没测,我也没测。② 轴是从形状论证的。若实测很便宜,A 更强;若很贵,C 的性价比上升。
2. C 的可行性未验证。 isTestFilePath 存在且已导出是事实,但"哪些门对测试文件免疫"需要逐门判定,那个清单今天不存在,规模未知。
3. ⚠️ 顺序风险来自卡本身,我原样转达:finding: dispatch-gates never names check:objectql-double-limit for an edited test file — the same whole-corpus shape as its listed sibling check:where-matcher #12322(欠命名那一面,e980f6448 / PR fix(devx): re-adjudicate 17 bare-root verdicts whose "unspellable" basis is false #12476 修过)仍 open,而减法机制正是会把它重新置于风险中的东西。⇒ ⛔ 若裁 B,finding: dispatch-gates never names check:objectql-double-limit for an edited test file — the same whole-corpus shape as its listed sibling check:where-matcher #12322 必须先关闭并复验,不得并行。
4. 本卡查重做得很扎实(100 张 open domain:devx 全量按标题枚举 + 阳性对照 dispatch-gates 返回 3),⇒ 无孪生卡这一点采信。

裁后我会怎么执行

裁 A ⇒ 摘 needs-user-decision,入 pm:queue,范围钉死为"在 dispatch-gates header 记录这 102 对是只能收窄那条保证的标价 + 指回本卡",⛔ 不含任何机制改动。
裁 C ⇒ 入队,但先要一次"哪些门对测试文件免疫"的测量作为派发前必答项。
裁 B ⇒ ⛔ 我不入队。先要求 #12322 关闭并复验,再把 B 拆成"机制"与"迁移"两张卡,因为它改的是所有车道共用的安全属性。

---

Generated by Claude Code

[os-elon]

Maintainer ruling applied — A: accept the price of narrowing-only, and RECORD it as a decision

Source: maintainer, 2026-08-28, live director session (summon #2, session session_016SG9S6V15MqeAgkehDcTwk), batch #5 item 2 presented with recommendation A; ruling verbatim: 「12842 同意」.

• A adopted: the declaration mechanism gains NO subtraction. The ~102 false-lead pairs per i18n-family gate on packages/cli are accepted as the standing price of declaredInheritedPopulation's narrowing-only guarantee — a declaration can never invent or hide a real population, and that property stays provable rather than per-declaration-trusted. The deciding axis as presented: today's cost is LOUD (an extra CLI closure build); a subtraction bug would be SILENT (a gate that stops watching real code while every dispatch order looks normal) — the wrong trade in the direction this repo prices highest.
• Scope pinned (per the facet block's 裁后执行, adopted as presented): record the decision in scripts/pm/dispatch-gates.mjs's header — the 102-pairs-per-gate price, the reason (narrowing-only is the safety property), and a backlink to this card — ⛔ no mechanism change of any kind. Option C (deriver-side isTestFilePath application) stays available as the recorded fallback if the measured cost ever escalates; option B (declaration-side subtraction) is rejected, and any future revisit of B requires finding: dispatch-gates never names check:objectql-double-limit for an edited test file — the same whole-corpus shape as its listed sibling check:where-matcher #12322 closed and re-verified first (the order-risk the card itself flagged).
• Labels this stroke: needs-user-decision → pm:queue; tooling, domain:devx unchanged. Dispatch guidance: S mechanical / sonnet-or-opus — the deliverable is a header comment plus its ratchet-safe wording.

---

Generated by Claude Code