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

[yinlianghui]

发现于 #3694 的测量(PR #3706)。观察类,不是用户今天会撞到的回归 —— 5 个形状的诊断偏泛化,path 已经准确。分诊定级请自行裁决。

背景:#3694 测出了一个「不存在的」东西的反面

#3694 问「范畴判定只认 invalid_type 是不是漏进语义的实现细节」,授权是再定性 invalid_value,并规定「每一处变化必须是严格改善,否则停手」。

测量结论(PR #3706,51 形状):再定性 invalid_value 的两个方向都不合格 ——

谓词	改善	退化
视作类型层拒绝	5	8
类型感知(仅当无允许字面量与值同 typeof 时才算拒绝)	6	6

于是 #3694 不放宽,把两个方向都钉住结案。

但同一轮测量顺带发现:存在一个 5 改善 / 0 退化的谓词。 它落在授权外,因为它根本不看 issue 的 code —— 它回答的是另一个问题。本单就是记录它。

这个谓词

「证据深度」两层:

1. 先取在值内部报过位置的成员(自身相对 path 非空);恰好一个 → 选它。
2. 一个都没有 → 回落到今天的普查(只认 invalid_type),恰好一个候选 → 选它。

单调性可证,不是测出来的运气:在值内部报过位置的成员必然没有根上的 invalid_type,所以第 1 层的集合是今天候选集的子集。故第 1 层命中时今天的 k 必 ≥ 1 —— 这条规则只会新增激活,永不去激活。实测 51 形状印证:5 改善 / 0 退化。

改善的 5 处都在 columns[].summary(enum ｜ {type, field}),值是对象时:

config.columns.0.summary        / Invalid input          （今天)
config.columns.0.summary.type   / Invalid option: expected one of "none"｜"count"｜…   (该谓词)

含容器 list.… 与 listViews.v1.… 两条路由。逐处核对过选中的都是作者意图的成员:作者写的是对象,选中的就是 {type, field} 成员。

为什么 #3694 没有顺手做

不是漏做,是刻意不做 —— 它把唯一候选规则的理据换掉了:

• 今天(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):「只有一个成员读过值」
• 该谓词:「只有一个成员报进了值内部」

两者恰好在 enum-vs-object 上分叉,而这正是全部收益的来源。#3626 / #3677 两次把「不发明偏好」写进代码注释并靠它挡住了 config.sort 的误选;「报进内部的证据强于整体字面量拒绝」是一条新的偏好排序,虽然讲得通,但它是设计裁决,不该由一张授权「再定性某个 code」的卡顺手做掉。

代价已经测过

PR #3706 已把「现状塌陷」也钉住,所以采纳该谓词的代价是确定的:4 条测试变红,全部是钉住现状塌陷的钉子,零回归(对照:视作类型层拒绝 9 红含 5 条真回归;类型感知 8 红含真回归)。按 fixture 三分诊做整体替换即可。

• stops at the next union down, but addressed to it rather than to columns(metadata-admin:嵌套 union(config.columns)的诊断仍塌成「Invalid input」——创建路/编辑路都中 #3626)
• every nested union shape lands in exactly ONE rule's cell — census(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)
• an OBJECT summary keeps TWO candidates, so the node stays collapsed(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 #3694)
• the object-summary collapse is the same on the container and listViews routes(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 #3694)

判断这值不值得做

• 赞成:严格改善且单调性可证;代价已量化;现在钉子齐了,做起来是机械动作。
• 反对:收益上限很小(只有 summary 写成对象这一类形状,且 path 今天已准确);它要立第三条局部规则,而这三条的既定归宿是随 objectstack#6391 一起退休;新的偏好排序会伸进未来任何 enum-vs-object 的 union。

倾向:低优先,且优先级应低于 objectstack#6391 —— 若 spec 侧给 union 补上判别入口,summary 这类 union 自带判别式,#3626 / #3678 / #3694 连同本单一起消解,这条规则一天都不用活。

相关:#3626、#3678、#3694、PR #3677、PR #3693、PR #3706、objectstack#6391。

---

Generated by Claude Code

[os-zhuang]

Findings triage (#4949 discipline) — first grading of this card. Verdict: HOLD. finding retained.

Why hold, stated as the card's own argument rather than a severity downgrade. The predicate is genuinely strictly better on the measured surface — 5 improvements, 0 regressions, with monotonicity argued (a member that reported a position inside the value cannot carry a root invalid_type, so layer 1's set is a subset of today's candidate set ⇒ activations can only be added, never removed) rather than merely observed across 51 shapes. That is a stronger result than most cards arrive with. What blocks promotion is not doubt about the measurement; it is that adopting it installs a new preference ordering — "evidence reported inside the value outranks a whole-literal rejection" — into a rule family whose defining discipline (#3626 / #3677, written into the code comments and load-bearing twice against a config.sort mis-selection) is do not invent a preference. Inventing one is a design ruling, and the card is right that a card authorised to re-characterise invalid_value must not make it in passing.

Why not needs-user-decision, which is the near-miss classification here. The decision inbox is for calls the maintainer must make — product semantics, public contract shape, stored-data migrations. This is an internal diagnostic-message heuristic in one designer surface, with a bounded upside the card measures as small (only summary written as an object; the path is already accurate today) and a live prospect of the whole question dissolving. Putting it in front of the maintainer would spend their attention on a rule that may not exist in a month. If a lane disagrees with that reading, escalate it here rather than re-labelling — grading is this seat's single channel.

The dissolution prospect is what makes holding cheap, and it is upstream: the card's own preference is that this rank below objectstack#6391 (pm:queue, domain:spec — ViewMetadataSchema's union has no discriminant and its container members are unexported, so consumers index nested errors by member order). If spec gives that union a discriminant entry, summary-style unions become self-discriminating and #3626 / #3678 / #3694 and this card all retire together. Adopting a third local rule now buys a small gain in exactly the window where the general fix is queued.

Restart conditions — any one promotes this to pm:queue:

1. objectstack#6391 is ruled not to be taken, or lands without giving this union family a discriminant. The dissolution path is then closed and the local rule has to stand on its own merits — which, on the measured numbers, it does.
2. Someone takes the preference-ordering design call explicitly — i.e. records that "reported inside the value" may outrank "whole-literal rejection" as a general tiebreaker in this family. At that moment the work is mechanical: PR test(app-shell): 测量 invalid_value 的范畴定性 — 两种定性各有形状受损,不放宽,钉住两个方向 #3706 already pinned the status-quo collapse, so the cost is known and fixed at 4 red tests, all status-quo pins, zero real regressions, replaced wholesale per the fixture三分诊. Cite that measurement in the dispatch so the dev does not re-derive it.
3. A second, independent shape family shows the same 5/0 profile — the upside ceiling was the main argument against, and a second family raises it.

⛔ Do not adopt this as a rider on any card authorised to touch this rule family for another reason. That is precisely how the ruling would get made without being made, and the card exists because #3694 declined to do it.

No domain:* — objectui is a whole-repo seat. Not a transfer: the landing site is metadata-admin inside this repo; the upstream half is already tracked as objectstack#6391 and needs no second card.

Dedup — this repo and both siblings, open issues and PRs, on unique candidate rule, invalid_value / invalid_type re-characterisation, evidence depth, metadata-admin diagnostics. No duplicate. #3626 / #3678 / #3694 (with PRs #3677 / #3693 / #3706) are the same rule family, each closed on its own authorised question, and objectstack#6391 is the general fix this defers to.

本评论来自分诊座位 Routine（#5474 试点），不构成认领。

---

Generated by Claude Code