spec: the option-1 population is measured — 6 rows, and it cannot reach the ≥7 duration keys that name no unit anywhere. What is the goal?

[claude]

维护者速读

一年前定下的规矩是:一个表示时长的数字,单位要写在键名里,比如 timeoutMs 而不是光叫 timeout——这样 AI 或人照着写配置时,不会把 60000 毫秒当成 60000 秒。

2026-09-07 你批过一次(batch #65):当时提议「让检查器也去读源码注释(JSDoc)里的单位」,没有采纳,理由白纸黑字写着——「这一类到底有多少个,没数过」。

现在数过了。结论有点出人意料:

• 当初被否掉的那个改动,总共只够到 6 个键。小到不值得为它改规则。
• 但普查顺带挖出另一批至少 7 个:它们表示时长或时刻,任何地方都没写单位——注释里没有,描述里也没有。读不读 JSDoc 对它们毫无帮助,因为它们压根没有 JSDoc 单位可读。
• 还有一批 4 个是「时刻」(时间戳),它们该用仓库里已经存在、已经有 6 处在用的那个专用类型,却写成了普通数字。这一批不需要动任何规则,只是没跟上既有约定。

所以真正要你定的不是「要不要读 JSDoc」,而是:这件事的目标到底是哪一个。

• A — 就按当初那个提议做:让 JSDoc 算一个单位来源。够到 6 个键。
• B — 只做「时刻」那一批:让它们改用已有的专用类型。不改任何规则,4 个键(外加 6 个同形的)。
• C — 改的是「哪些键进入检查范围」本身,让「哪儿都没写单位」的时长也能被点名。改动最大,但只有它够得到那至少 7 个。
• D — 什么都不做,batch Add comprehensive test coverage for ObjectStack spec protocols - 100% coverage achieved #65 维持原判。现在你知道被否掉的那个东西是 6 个键。

我的推荐是 B,理由和回退在下面。请回一个字母:A / B / C / D。

---

Background

check:duration-unit-keys refuses a duration-shaped numeric key that does not carry its unit in the key name. The #15939 epic closed on 2026-09-13 having remediated 21 such rows and landed the gate widening (#17635 → 484484041a). This card is the follow-on that epic deliberately held back.

Governing text — quoted, ⛔ not paraphrased

Decision batch #65, 2026-09-07, maintainer 「同意」 (#15939 comment 5564447683):

the gate does not start reading JSDoc as a unit channel (option 1 not adopted; its population is unmeasured)

Keys with no unit in either channel stay "listed, not judged" (the #14519 shape) — unchanged.

And the one prior classification of a member of that shape, packages/spec/src/migrations/entries/semantic/18.kernel-health-check-and-hot-reload-durations-unit-in-key.ts:63-65:

The sibling shutdownTimeout on HotReloadConfig is deliberately NOT renamed with them: its JSDoc reads "Graceful shutdown timeout" and names no unit anywhere, so it is the #14519 unit-nowhere shape the #14478 gate leaves outside its verdict, not part of this row set.

⚠️ That sentence classifies one key out of one PR's row set. It does ⛔ not settle the #14519 question, and this card is not overturning it.

Protocol declaration

No protocol change is proposed by this card. Options A and C would change gate behaviour and therefore need their own implementation cards; option B changes no gate at all. ⛔ Nothing here is adopted by filing.

The measured population — the precondition batch #65 named

Census #18078, report-only, at origin/main @ fb29f62cee. Re-check: git fetch origin main && cd packages/spec && tsx scripts/check-duration-unit-keys.ts --list.

class	count	reachable by option 1?
A unit nowhere, inside the census	23 mechanical → 6 genuine durations	no — no JSDoc unit to read
B duration, unit in JSDoc only, outside the census	2	yes
C instant as bare z.number() instead of EpochMs	4	yes
D duration/instant, no unit in any channel and outside the census	≥ 7	no
non-member scoping (the maxMemory shape — a size, not a duration)	6 exact / 58 wider	n/a

⇒ Option 1 as declined reaches 6 rows (B + C) and cannot reach D at all.

Two measured facts that bear on the choice:

1. A's 23 is not a population. Only 6 are genuine durations; the other 17 are dimensionless — counts, backoff multipliers, one instant — matching only because the predicate is a 25-token name list. metrics.zod.ts:224 ageBuckets (JSDoc "Number of age buckets", default(5)) is in it because age is a token. The rule's remedy is "move the unit into the key name", which is meaningless where there is no unit.
2. Adopting A brings a known false positive live. plugin-registry.zod.ts:110 downloadsLastMonth has JSDoc "Downloads in the last 30 days"; the gate's PROSE_COUNTED_RE (:281) matches 30 days. Today the row never enters the census so the misread is inert. On the day JSDoc becomes an admission channel, the gate starts refusing a download count.

Options × real cost

option	what happens	cost a user or author can feel
A JSDoc becomes a census-admission channel	2 durations + 4 instants become judgeable	a gate change plus 6 renames; and downloadsLastMonth starts being refused as though it were a duration
B instants only → EpochMs	4 rows (+6 more of the same shape) adopt an existing convention	⛔ no gate change, no rule change, no migration prescription; it is drift from a live convention with 6 declared users
C widen census admission itself	a unit-nowhere duration becomes listable regardless of its name	the largest change, and the only one that reaches the ≥7 in class D; it changes what the census is
D do nothing	batch #65 stands, now with a number	the ≥7 stay invisible; a reader still cannot tell 60000 ms from 60000 s on those keys

业务含义直译

• A = 「把注释也当成说明书的一部分」。听起来合理,但实测只影响 6 个键,而且会让一个下载计数被当成时长退回——为很小的收益引入一个新的误判面。
• B = 「本来就有的规矩,把漏网的几个补上」。不新增任何义务,只是让既有约定一致。
• C = 「把体检的名单本身扩大」。这是唯一真正堵住那个洞的做法,也是唯一需要重新想清楚「什么算时长」的做法。
• D = 「维持原判」。现在这个判断有数字支撑了,不再是因为不知道有多少而搁置。

四维分析(从业务立场)

① 项目长远合理性。 B 缩小特例:它让一个已经存在、已经有 6 处在用的约定被一致地遵守,不新增任何契约表面。A 扩大特例——它让「注释」成为第三个有语义的说明书通道,而这个通道的内容不进入发布产物,读者在参考页上看不到它,于是产生「代码里写了、文档里没有」的长期分裂。C 是三者里唯一减少增生的:它把判断依据从一张 25 个词根的名字表(一个必然长期漂移的清单)换成「有没有单位」这个语义问题。

② 实际业务拉动。 今天真正会撞上的人:拿着参考页照抄配置的作者与 AI。B 的 4+6 行是时间戳,写错的后果是把毫秒当秒传进去,静默地差 1000 倍。A 的 2 行时长同理但只有 2 行。D 类那 ≥7 行是拉动最大的一批——恰恰是三个选项里两个够不到的那批。⛔ 没有任何一类有已发生的事故记录,所以这四项都不是救火。

③ 防 AI 犯错。 这一轴决定推荐。⚠️ 出错时谁看见什么:一个 AI 照着参考页写 timeout: 60000,如果单位只写在源码注释里,参考页上什么都没有,它没有任何线索,写错了也不会被拒绝——静默的 1000 倍偏差,直到运行时行为不对才被发现。这正是 ③ 说的「响亮拒绝优于静默容忍」被违反的形态。B 把这批键换成一个闭合的类型(EpochMs),等于把自由结构换成枚举,是这一轴上最干净的一步。A 只在 2 个时长上做同样的事,但同时引入一个新的静默错误方向:downloadsLastMonth 会被响亮拒绝,而它是对的——假阳性教作者不信任这道门,比漏判更伤。

④ 创业阶段不扩散。 B 不新增任何已声明的键与永久义务,它只是让既有的用起来。A 新增一条规则分支加 6 个重命名的迁移义务,换 6 行。C 的义务最大,但它换来的是唯一一个不会随名字表漂移而失效的判据——如果只做一件长期的事,C 是那件;如果现在只做一件便宜的事,B 是那件。

推荐 + 回退 + 置信缺口

推荐 B,并把 A 与 C 分开处理:B 不改规则、有现成约定、覆盖静默偏差最严重的一批;A 的 2 行时长可以随后单独做,但要连带处理 downloadsLastMonth 这个会被激活的假阳性;C 应当是它自己的卡,因为它改的是普查的定义而不是读散文的方式。

回退:B 若在实施中发现某个时间戳其实不是 wall-clock(比如是相对偏移),那一行退出本批、单独记录,⛔ 不为了凑数改类型。

⚠️ 置信缺口——本分析看不见的:(1) D 类的 ≥7 是下界不是计数——普查明说 "at least 7",要数准需要先答「什么算时长」,而那正是本卡的下游问题;(2) 没有任何事故或用户报告数据,四类的排序完全基于「读者能否在发布面上看到单位」的推理,而不是基于谁真的写错过;(3) 本分析只看 packages/spec,#15642(扫描子树的范围问题)仍独立开着,别的 workspace 包里同形的键没有数过。

裁后执行段

• A ⇒ 一张 domain:spec 实现卡:普查预测入队面 packages/spec/scripts/check-duration-unit-keys.ts + 6 处重命名的 ADR-0087 转换与墓碑;同卡必须处理 downloadsLastMonth 假阳性,⛔ 不得留给下一个人发现。
• B ⇒ 一张 domain:spec 实现卡:4 行(外加 6 行同形)改用 EpochMs;这是已发布键的类型收窄,按 clause ② 判定后决定是否需要席内契约复核。
• C ⇒ 先一张 needs-user-decision 子卡答「什么算时长」(A 类 23 vs 6 的边界),再谈实现。
• D ⇒ 本卡关 not_planned,在 spec: tenant.zod.ts idleTimeout / sessionTimeout publish a describe with no unit, while the JSDoc one line above says seconds #14519 上留一条评论把这次的数字记进去,免得第三次因为「没数过」而重开。

下游问题 —— ⛔ 本卡不问,答完上面那个字母才谈

1. 什么算「一个时长」? A 类机械 23 / 语义 6 的边界。17 行的差全是非时间单位的间隔与年龄(snapshotInterval = 每 N 个事件,reflectionInterval = 每 N 次交互,ageBuckets = 桶数)加 5 个退避倍数。B 类的边界同样悬在这个答案上。
2. 要不要扩到所有带单位的数值,而不只是时长? 精确同形 6 行(maxMemory 那种),宽泛同形约 57–58 行(字节、像素、百分比)。

相关单与 PR

#15939(已关,母单与 Ruling A)· #18078(已关,本卡的普查)· #14519(unit-nowhere 形状)· #15676(ruling B,时刻形状)· #18075(不同的洞:经 keyUnits 进普查却逃过判决的 2 行)· #15642(扫描子树,独立)· #17635 / #484484041a(闸门加宽)· batch #65 台账在 #12708。

---

os-decision-facets

• ① 项目长远合理性:B 让一个已有 6 处在用的约定被一致遵守,⛔ 不新增契约表面;A 让源码注释成为第三个说明书通道(而它不进发布产物),扩大分裂;C 是唯一把判据从一张会漂移的名字表换成语义问题的选项。
• ② 实际业务拉动:今天撞上的是照着参考页写配置的作者与 AI;拉动最大的是 D 类那 ≥7 行,而 A 与 B 都够不到它;⛔ 四类均无已发生事故记录,都不是救火。
• ③ 防 AI 犯错:单位只写在源码注释里时,参考页上什么线索都没有,写错静默地差 1000 倍(毫秒/秒);B 把这批换成闭合类型(EpochMs)是这一轴最干净的一步;⚠️ A 会让一个下载计数被响亮误拒(downloadsLastMonth),假阳性比漏判更伤信任。
• ④ 创业阶段不扩散:B 零新增永久义务;A 用一条新规则分支加 6 个迁移义务换 6 行;C 义务最大但换来唯一不随名字表漂移的判据。

推荐:B(只做时刻那一批,不改规则)。选项:A 采纳 option 1 / B 只做时刻 / C 扩大普查本身 / D 维持 batch #65。

置信缺口:D 类的 ≥7 是下界不是计数(要数准需先答「什么算时长」);无任何事故或用户报告数据,排序基于「发布面上看不看得见单位」的推理;只覆盖 packages/spec,别的 workspace 包同形键未数(#15642 独立开着)。

_{Filed by the epic PM for #15939, session_015c5G6TmpMKgnusmTpD7Ntt, 2026-09-13T23:15Z. Population from census #18078 at fb29f62cee; every figure quoted here was re-verified by this seat against the tree, not relayed.}

---

Generated by Claude Code

[os-elon-musk]

Ruling: batch #133 item 4 · letter C · maintainer 「决裁批 #133 同意」 2026-09-14T03:33Z (to 1D·2a·3A·4C·5C)

Ruling recorded — C: the goal is making unit-nowhere durations judgeable — census admission moves from the 25-token name list to 「does this key declare a unit」. B (the 10 instants → EpochMs) executes now as tranche 1 under this ruling; the definition of 「a duration」 is tranche 2's precondition and is its own decision card. A (JSDoc as an admission channel) and D (leave batch #65 as is) are not taken (director seat, decision batch #133 item 4, 2026-09-14)

Derived first from the long-term axis: two years out every duration/instant numeric in @objectstack/spec is either unit-in-name or a closed type, and the gate judges declarations rather than a name list that must drift — only C is that shape; A makes source comments a third contract channel that never reaches a reader of the published reference; D leaves ≥7 keys where 60000 cannot be told from 60 s. ②③④ do not overturn ①: ③ agrees (the D-class rows are the silent ×1000 case and only C reaches them); ④ only orders the work — the cheap, already-conventional B half first, the definition second.

What is ruled

1. Tranche 1 — spec: route the 10 bare z.number() wall-clock instants to EpochMs (census #18078 class C ×4 + unit-nowhere instants ×6) — tranche 1 of ruling C on #18080 #18114 (domain:spec, pm:queue, p2): the 10 bare z.number() wall-clock instants adopt EpochMs (census class C ×4 + the six unit-nowhere instants, tokenExpiry included; lamport clocks excluded). No key rename. Whether any producer writes non-integers is measured in the PR and decides banner/disposition, ⛔ not the letter.
2. Tranche 2 precondition — [Decision] what counts as a duration for the unit-in-key census — replace the 25-token name list with a declared shape? (class A 23→6 and class D ≥7 of census #18078; tranche 2 precondition of ruling C on #18080) #18115 (needs-user-decision, domain:spec, p2): 「what counts as a duration」 — declared shape (closed Duration* types + unit-in-name, name list retired to a hint) vs curated ledger vs prose heuristics; presented in the next decision batch with facets and speed-read.
3. Tranche 2 implementation (the 6 + 2 + ≥2 genuine durations, and the census-admission change in check-duration-unit-keys.ts) is filed only after [Decision] what counts as a duration for the unit-in-key census — replace the 25-token name list with a declared shape? (class A 23→6 and class D ≥7 of census #18078; tranche 2 precondition of ruling C on #18080) #18115 is ruled; ⛔ nobody widens the gate before then.
4. A is refused: it reaches 6 rows, none of them the D class, and activates a measured false positive (downloadsLastMonth). D is refused: batch Add comprehensive test coverage for ObjectStack spec protocols - 100% coverage achieved #65's stated precondition (「its population is unmeasured」) is now met, and the measurement points past option 1 rather than at it.
5. spec: tenant.zod.ts idleTimeout / sessionTimeout publish a describe with no unit, while the JSDoc one line above says seconds #14519 (the unit-nowhere shape) receives a pointer to this ruling so the 「listed, not judged」 resting place is read as superseded once [Decision] what counts as a duration for the unit-in-key census — replace the 25-token name list with a declared shape? (class A 23→6 and class D ≥7 of census #18078; tranche 2 precondition of ruling C on #18080) #18115 lands; finding: check:duration-unit-keys scans packages/spec/src/** only — the same offender shape exists in workspace packages the gate never reads #15642 (scanned subtree) stays independent.

State

needs-user-decision removed; closed completed — the question this card asked (「what is the goal?」) is answered and the work lives on #18114 / #18115. domain:spec / tooling kept.

---

Generated by Claude Code