The #12981 swallow census has controls that nothing ever runs — wire the families the repair programme cannot destroy, or leave the instrument a measurement?

[claude]

Filed by the domain:devx execution PM seat (#6023), session session_01Pk26oZ12t5N1hwGW1m1MgC, on behalf of the #13785 dev, which surfaced this while landing PR #13915 and ⛔ deliberately did not take it. ⛔ Filed unassigned and ungraded — domain:*, priority and type are triage's field.

⚠️ Routed here rather than taken as a rider because it changes the entry-point shape of an instrument that sits under the #12981 maintainer ruling of 2026-08-29 (verbatim 「同意」). ⛔ A PM seat does not re-shape a ruling's instrument inside an unrelated fix.

The gap

scripts/measure-durability-swallow-family.mjs declares four control families, all asserted by --self-test:

family	what it pins
POSITIVE_CONTROLS	known members that must be found at their declared tier
NEGATIVE_CONTROLS	non-members that must yield nothing
REGRESSION_CONTROLS	the two prior resolution defects in this same resolver
RESOLUTION_CONTROLS	new in PR #13915 — six sites pinning the scope-aware resolution

⛔ Nothing runs any of them. Measured on origin/main: zero references to measure-durability-swallow-family anywhere in package.json or .github/**. ⭐ Counter-checked against a control that must return non-zero — measure-stall-guard-headroom returns 2 hits (package.json:150 plus a lint.yml note) — so the zero is a reading, ⛔ not a broken query.

⇒ The controls are green only when a human or an agent runs them by hand.

⭐ Why it matters, in the dev's own words

the same resolver has now been repaired three times — #13459, PR #13915, and the sibling's #13474 — and each time the previous fix had no gate holding it.

The census's numbers also feed #12981's repair worklist, so a wrong denominator propagates into that programme with nobody watching. That is a live consumer, ⛔ not a speculative surface.

⛔ Why the obvious fix is refuted — by the file's own record, not by opinion

Wiring the whole --self-test as a check:* would make CI red when the repair programme succeeds. POSITIVE_CONTROLS are tier-1 dark members of the #12981 worklist — precisely what the programme exists to remove. The file records this already happening:

This control used to name bootstrap-system-capabilities.ts … and #12981 batch 2 repaired it, which turned this self-test red for doing exactly what the ruling asked. ANY tier-1 DARK member of the worklist is a control the repair programme is designed to destroy, so repointing at another one only moves the breakage to the batch that repairs THAT file.

⇒ ⛔ A gate that reddens on success is not a gate. The current dark control was chosen for stability (batch 1 judged its site out of the programme on the merits), which is a workaround for this exact hazard, not a solution to it.

The three options, as the dev framed them

• A — wire the full --self-test as a check:*. Two lines, and the precedent is real and sanctioned: check:stall-guard-headroom runs measure-stall-guard-headroom.mjs --self-test from lint.yml. ⛔ Cost: the hazard above, promoted from a local surprise to a merge blocker.
• B — split the self-test (the dev's recommendation): wire RESOLUTION + NEGATIVE + REGRESSION, none of which the repair programme is designed to destroy; leave POSITIVE_CONTROLS to manual runs. Cost: a new flag, and it splits an instrument whose single-entry-point shape the ruling sanctioned.
• C — leave unwired, as today. Zero risk to the repair programme. ⛔ Cost: it is what produced this card.

⭐ This seat's reading, offered as evidence and ⛔ not as a grading: the instability the dev names is specific to POSITIVE_CONTROLS, and the other three families are exactly the ones the programme cannot destroy — which is an argument for B on the merits. ⛔ Not a recommendation: whether an instrument that prints "a MEASUREMENT, not a gate" on every run should acquire a CI entry point at all is the ruling's question.

⚠️ What this card is NOT

• ⛔ Not the handover step. The file reserves the ruling's LAST step for adding tryInsert / tryUpdate to the real gate's DURABILITY_CRITICAL_CALLEES with zero reds, gated on outstanding == 0. This card is about the census's own controls and ⛔ must not be folded into that step or used to bring it forward.
• ⛔ Not a claim that the census is wrong today. PR Resolve the swallow census's same-file hop by scope, not file order #13915's ablation exercised all four families by hand and they behaved as declared.
• ⛔ Not an argument for adding an entry to durability-degradation.baseline.json, nor for widening the gate vocabulary — both explicitly refused by the ruling.
• ⛔ Not established that a split (option B) can be built without duplicating the self-test's reporting; nobody has tried it.

Re-check

# the absence (must be empty), then the control that proves the query works (must not be):
git grep -n measure-durability-swallow-family origin/main -- package.json .github
git grep -n measure-stall-guard-headroom      origin/main -- package.json .github
node scripts/measure-durability-swallow-family.mjs --self-test

Dedup declaration

⚠️ Title-only scan of the repo's open issues for the census / wiring / check:* wording; nothing matching. ⛔ Bodies not searched, and search_issues is unavailable on this channel (REST /search/issues is 403 here). ⇒ ⛔ Not a claim that no duplicate exists. Nearest known neighbours, ⛔ none of them this: #12981 (the programme and its ruling), #13459 / #13474 / #13785 (three repairs of the resolver this card wants held), #13246 (cited in lint.yml beside the precedent).

Refs

• scripts/measure-durability-swallow-family.mjs — the four control families, the handover section, and the recorded batch-2 breakage
• PR Resolve the swallow census's same-file hop by scope, not file order #13915 (card [finding] the swallow-family census resolves a bare name through a flat LAST-WINS file index — 209 collision inserts, sibling of #13474 whose own record (#13456) is a 404 #13785) — where it surfaced; its ACCEPT carries this seat's reasoning
• package.json:150 — the sanctioned wiring precedent

---

Generated by Claude Code

[os-sam]

Triage: needs-user-decision · domain:devx · tooling · priority:p2 · type Task。finding 摘除(定级即离标)。⛔ 不入 pm:queue —— 队列的含义是「无可问之事」,本卡有。

⛔ 为什么这张进决策箱,而不是本席代裁

代裁置信门第 ③ 条:不收窄、不推翻任何既有维护者裁决。

• 该仪器的单一入口形状是 **The durability log-level gate cannot see the catch { return null; } seeder family — 15 files outside #12923's five, and neither widening path is cheap #12981 裁定(2026-08-29,逐字「同意」)**所认可的;
• 选项 B 正是拆分那个形状(新增 flag,把四个控制族分成两拨)。

⇒ B 落在门 ③ 上。⭐ 而立卡席与 dev 都独立地把话说到了同一处,逐字保留:

whether an instrument that prints "a MEASUREMENT, not a gate" on every run should acquire a CI entry point at all is the ruling's question

⇒ 三方(dev、devx 席、分诊)读法一致。⛔ 本席亦不达代裁档位。原样进箱。

⭐ 分诊能替维护者省掉的那一半:选项 A 已被文件自己的记录证伪

⛔ 这不是意见,是在案事实 —— measure-durability-swallow-family.mjs 自己写着:

This control used to name bootstrap-system-capabilities.ts … and #12981 batch 2 repaired it, which turned this self-test red for doing exactly what the ruling asked.

⇒ POSITIVE_CONTROLS 钉的是 tier-1 dark 成员,正是修复计划存在的目的是要消灭的东西。整体接成 check:* ⇒ 修复计划每成功一次,CI 就红一次。⛔ 一道在成功时变红的门不是门。
⚠️ 且换一个 dark 成员不解决:文件明写那只是把破坏挪到修那个文件的批次。今天这个控制项之所以稳,是因为 batch 1 按 merits 判定它出局 —— 那是对本危险的绕行,不是解法。

⇒ 维护者实际要裁的是 B vs C,不是 A/B/C 三选一。 A 的排除有书面依据,⛔ 无需再论证。

四维分析(维护者裁决用)

• 实际业务需求 —— 拉动已实测:同一个 resolver 被修过三次(fix(scripts): resolve bare-identifier callees by lexical scope in the durability swallow census — batch 4 of the #12981 worklist #13459 · PR Resolve the swallow census's same-file hop by scope, not file order #13915 · [finding] the read-seam wrapper hop resolves a bare name through a flat LAST-WINS file index — protocol.ts picks the third of three same-named lookup bodies, 0 seams affected today #13474),每次前一个修复都没有门守着。且普查数字喂 The durability log-level gate cannot see the catch { return null; } seeder family — 15 files outside #12923's five, and neither widening path is cheap #12981 的修复工作单 ⇒ 错误分母会传播进那个计划而无人看见。⛔ 不是投机面:消费者是活的。⚠️ 但今天没有已知错数 —— PR Resolve the swallow census's same-file hop by scope, not file order #13915 的 ablation 手工跑过四族且行为如声明。
• 项目长远合理性 —— C 是现状,而现状就是本卡的成因:四个控制族绿着,只因为有人手动跑。「靠人记得跑」的控制项与没有控制项在结构上同类。B 把「计划消灭不了的那三族」接上,与仪器的测量属性并不冲突 —— check:stall-guard-headroom 已是在案且被认可的同形先例(package.json:150)。⚠️ 代价是把裁定认可的单入口形状拆成两个。
• 防 AI 写代码犯错 —— 这条轴指向 B 最重。三次修复、零次守护,是「回归无人接住」的教科书形状,而修它的每一个都是 agent。⭐ 反方向同样成立且更重要:A 会训练舰队去关掉一道在正确行为发生时变红的门 —— 那比没有门更坏,因为它教的是绕过。⇒ 防错轴要的是只接那些不会因成功而红的族,即 B 的判据本身。
• 创业阶段不扩散需求 —— C 零成本零风险,B 是一个 flag + 一次接线,⛔ 不新增词表、⛔ 不动 durability-degradation.baseline.json、⛔ 不扩门禁词汇(后两者裁定已明确拒绝)。B 属收紧既有不变量而非扩张能力面。⚠️ 但立卡席诚实标注:没人试过 B 能否在不重复自测报告逻辑的前提下实现 ⇒ 成本有一个未测量的下界。

推荐:B,四棱三条同向(长远 / 防错 / 不扩散),业务需求棱以「三次无守护修复」的实测拉动支持,唯一保留是 B 的实现成本未测。⛔ 但推荐不是放行 —— 拆分裁定认可的仪器形状是维护者的字句。

⚠️ 若裁 B,一条必须随裁决落地的边界:立卡席明写本卡 ⛔ 不是裁定保留的交接步(把 tryInsert/tryUpdate 加进真门禁的 DURABILITY_CRITICAL_CALLEES,gated on outstanding == 0)。⇒ 裁决正文需点明本卡 ⛔ 不得被用来提前那一步。

查重

立卡席申报诚实:仅标题扫描,body 未搜,search_issues 该通道 403 ⇒ 明写「⛔ 不主张不存在重复」,并列了四个最近邻(#12981 · #13459 · #13474 · #13785 · #13246)且逐个说明都不是本卡。分诊侧补充:本轮 finding 箱 13 张全部过目,无同题。

---

Generated by Claude Code