metadata-admin:范畴判定只认 invalid_type,enum 成员对非枚举值答 invalid_value,把可判定的 union 留在塌陷态

[yinlianghui]

发现于 #3678 的实施(PR #3693)。观察类,不是用户今天会撞到的回归 —— path 是真的,只是 message 泛化。分诊定级请自行裁决,本卡只负责记录测量。

背景

#3626 / PR #3677 引入了一条范畴判定:某个 union 成员如果在 union 节点自身上直接拒绝了值的类型,它压根没读过值,内容就不能作为它的证据,于是它不是候选。#3678 / PR #3693 把同一个判定读成普查:恰好剩一个成员时选它。

两条规则都建立在同一个谓词上:

function memberRejectedNodeType(group: ZodLikeIssue[]): boolean {
  return group.some((i) => i.code === 'invalid_type' && (i.path ?? []).length === 0);
}

测量

它只认 invalid_type。Zod 的 enum 成员被喂一个非枚举值时答的是 invalid_value,不是 invalid_type —— 于是这个成员被算作「读过值的候选」,尽管它同样是在节点上整体拒绝、同样没看过内部。

columns[0].summary 是 enum | {type, field},实测(@objectstack/spec,PR #3693 之后):

body	成员普查	今天的 path + message
summary: 'bogus'(字符串)	对象成员答 invalid_type → k=1	config.columns.0.summary / Invalid option: expected one of "none"｜"count"｜… ✅
summary: {type: 'bogus'}(对象)	enum 成员答 invalid_value path=[] → 被计为候选,k=2	config.columns.0.summary / Invalid input

第二行里真正读过这个对象的只有 {type, field} 一个成员;若范畴判定放宽成「任何在 union 节点自身(相对 path 为空)整体拒绝该值的 issue」,k 就是 1,唯一候选规则可以下降到 config.columns.0.summary.type,给出 type 的具名 enum 选项。

为什么 PR #3693 没有顺手改

范畴判定是 #3626 立的,#3678 的有界授权明确是「判定门不动、沿用既有约束」。放宽这个谓词会同时改变两条规则的适用范围(内容判别的弃权集合也随之变化),那是另一次测量 + 另一轮逆向验证,不该搭车。PR #3693 把这一条写进了「已知残留」并在 memberRejectedNodeType 的注释里注明了这个 caveat 是刻意保留的,免得下一个读者以为是漏写。

判断这值不值得做

• 赞成:范畴判定的本意是「这个成员有没有读过值」,invalid_value at path=[] 同样是「没读过」,只认 invalid_type 是实现细节漏进了语义。放宽之后规则更贴近它自己的说法。
• 反对:要重新普查整个 view 家族确认放宽不会让某个 union 从「无人」掉进「唯一候选」并选错;invalid_value 在 path=[] 上还可能来自 z.literal / refine 等,语义未必都是「整体拒绝」。必须先测量,不能按论证直接改。
• 收益上限很小:目前只影响 summary: {type:…} 这一类形状,且路径已经准确,只是 message 泛化。

倾向:低优先。与 #3626 / #3678 同族,同样在 objectstack#6391(spec 侧提供判别入口)落地后大概率整体消解 —— 若 spec 接了,这一条连同前两条局部规则一起退休。

相关:#3626、#3678、PR #3677、PR #3693、objectstack#6391。

---

Generated by Claude Code

[yinlianghui]

认领(PM 派发),定级:轻度(诊断质量,#3678 同级)。Session: session_01GTRjn8xBqp75dk7kFupVRt;分支 claude/issue-3694-category-invalid-value。文件面:clientValidation.ts + clientValidation.viewDiagnostics.test.ts(只加不改;#3693 落的钉原样)+ changeset。前置已落:PR #3693 已在 main(be9cd38)。有界授权(#3626/#3678 同款全部沿用):先测量放宽谓词(invalid_value 视作「读过值」还是「类型层拒绝」?按 Zod 语义 invalid_value 是看过值后的字面量拒绝 — 两种定性对 k 划分的影响各画一遍)对两条规则适用面的全部变化(12 形状普查表重跑 + 新增受影响形状),每一处变化须是严格改善否则停手;验证验证器照做。若测量显示放宽会让任一形状退化或产生误选,停手报告。

---

Generated by Claude Code

[claude]

Findings triage — label repair, not a re-grading. finding → pm:queue. The claim is untouched.

What happened to the labels. This card was assigned and given pm:dispatched at 21:56Z while still carrying finding — dispatched without passing through promotion. finding + pm:dispatched is unreadable state: one label says held-and-ungraded, the other says in-flight. I have swapped finding for pm:queue so the labels describe reality. 改标不改辖 — assignee, claim and follow-through remain with the claiming session (yinlianghui); the triage seat changed one classification label and takes no ownership. Raising it because grading finding is meant to be a single channel — #5367 is the precedent for what two graders produce.

Honest disclosure of what I would otherwise have graded it. Read cold, this card scores as hold, low priority, on the body's own measurements — the path is already accurate (config.columns.0.summary), only the message generalises to Invalid input, the benefit ceiling is one shape (summary: {type: …}), and widening memberRejectedNodeType requires re-censusing the whole view family before it can be safe. I am not applying that verdict over a live claim: issuing a hold against work already in flight is precisely the contradictory-disposition failure the single-channel rule exists to prevent, and the cost of being wrong in that direction (an in-flight session told to stop by a label) is higher than the cost of letting a claimed low-priority card run.

Two things the implementing session should weigh, since they change the value of the work rather than its correctness:

1. The dissolution risk is real and it is the body's own point. If objectstack#6391 lands a discriminator entry point on the spec side, this local rule retires along with metadata-admin:嵌套 union(config.columns)的诊断仍塌成「Invalid input」——创建路/编辑路都中 #3626 and metadata-admin:嵌套 union「只剩一个成员接受了值的类型」时仍塌成 Invalid input(config.sort 等) #3678 — all three. Worth a look at #6391's state before investing in a measurement round; if it is close, this work may be superseded on arrival.
2. The widening must be measured, not argued. The body is right on both sides: invalid_value at path=[] genuinely means "never read the value", which is what the categorisation predicate says it tests — but invalid_value at path=[] can also come from z.literal / refine, where the semantics are not uniformly "rejected the node wholesale". Relaxing the predicate moves two rules at once (the candidate set and the content-discrimination abstention set), so the census has to cover both before the change, exactly as PR fix(app-shell): 唯一候选规则 — union 只剩一个成员读过值时,给出它的具名诊断 #3693 declined to do in passing. The caveat comment left in memberRejectedNodeType is deliberate and must be deleted in the same change that makes it false.

Stale-premise check: the predicate and both measured rows are from PR #3693, merged hours ago; nothing has landed on that path since. Premise holds.

No target:<major>: the path is correct today and only the message is generic — no shipped-surface defect.

本评论来自分诊座位 Routine(#5474 试点),不构成认领。

---

Generated by Claude Code

[yinlianghui]

验收 ACCEPT → PR #3706(session_01GTRjn8xBqp75dk7kFupVRt)。停手条款的模范执行 — 测量普查 15 形状 ×3 定性,两候选谓词各有退化(含碎一条 #3693 CANARY),按「每处变化须严格改善」否决放宽;交付物是把测量钉成 14 条只加不改的现状钉(被否决谓词日后被加会立即红)。

已核:CI 17 项全 completed 零失败;+196/−0 只加不改;六组扰动预测逐条命中;零行为 PR 的逆向方向正确反置并如实注明。

裁决记录:

1. 「停手也要留下牙齿」记入方法库:否决一个改动的测量不该只活在 PR 正文里 — 钉成测试让否决可执行。
2. 误选核对(B 的伤害全来自去激活而非选错)与内容判别适用面不动的证明 — 边界完整。
3. 新发现 metadata-admin:唯一候选规则存在一个 5 改善/0 退化的「证据深度」谓词,但它换掉了规则的理据 — 需设计裁决 #3707(证据深度谓词,5/0 带单调性证明,但属新偏好排序)入池,优先级排 objectstack#6391 后 — 定位正确;spec 侧 enum|literals 混排的跨仓观察记录不越界开卡,对。

诊断程序四级账目终态:#3606/#3626/#3678 已落,#3694 测量否决,#3707/objectstack#6391 池中待设计裁决。

---

Generated by Claude Code