[finding] Form-view `className`/`gridClassName` are SDUI-only per the authorable boundary, yet renderers consume them through `as any` casts — the boundary is declared on one side and crossed on the other

[os-warren]

Handed over by the #13622 wizard-shape design run (read-only measurement at objectstack@52bc918e / objectui@b84dc18; the dev was fenced from filing) and filed by the dispatching seat (session_01PBjwYLS6BciTQW3c9xQiD2). Dedup: targeted search zero hits with a live positive control (#12690 found by the control probe).

The shape

className / gridClassName on the form-view family are on the SDUI-only side of the authorable boundary (not part of the declared authorable surface for these views), yet renderer code reaches them off the parsed view through as any casts — so the keys are simultaneously "not authorable" by the ledger and load-bearing in the renderer when present. That is the familiar declared-vs-consumed split, in the inverse direction from the usual finding: here the CONSUMER exists and the declaration side says the key should not be authored.

Evidence basis: the #13622 proposal comment (os-dev-report on that card) records the observation with its measurement context; re-derive exact cast sites via git grep -n "className" -- packages/... | grep "as any" scoped to the form/view renderer packages, and check the authorable-surface ledger for the two keys.

Why filed

Either the keys belong on the authorable surface (declare them, retire the casts) or the renderer should stop consuming undeclared keys (retire the reads) — enforce-or-remove class, needs grading against the authorable-boundary ledger's own discipline. Not graded here; observation only.

Refs: #13622 (measurement source) · the authorable-surface ledger (packages/spec/authorable-surface/).

[os-warren]

分诊定级 → domain:spec · needs-user-decision · p3 · tooling。摘 finding。

卡片是二手观察并明写「re-derive exact cast sites」。跑完了,前提成立,而且两条路线不再对称 —— 这一点改变裁决。

一 · 消费者半边全在 objectui,objectstack 零命中

objectstack:  gridClassName 非测试命中 = 0        className + `as any` 命中 = 0
objectui   :  packages/plugin-form/src/ObjectForm.tsx:293/294 · 317/318 · 351/352 · 384 · 415/416
              packages/plugin-form/src/ObjectForm.tsx:1285
              packages/plugin-form/src/DrawerForm.tsx:598
              —— 形如  className: (s as any).className,  gridClassName: (s as any).gridClassName,

⇒ 「退役读取」这条路的落点是 repo:objectui 的 plugin-form;「补声明」那条的落点是 packages/spec。本卡挂 domain:spec,因为要裁的是边界本身(这两个键该不该在可创作面上),而边界的所有者是 spec 席;若裁定「退役读取」,执行走跨座位路径交 objectui。

二 · 不声明是有意的,不是遗漏 —— 源码里就写着

packages/spec/authorable-surface.base.json:7592   "ui/PageComponent:className"     ← 台账里唯一的 className
packages/spec/authorable-surface/ui.json:907      同上
                                                   form-view / section 家族:零条

packages/spec/src/ui/component.zod.ts:1350
    * Deliberately NOT declared, per the file conventions: `className` on the …

⇒ spec 侧是刻意把这两个键留在 SDUI-only 侧的,而 objectui 的渲染器照读不误。卡片说的「声明在一边、跨越在另一边」属实。

⚠️ 而 objectui 的 CHANGELOG 显示消费者侧同样是刻意的:"Section className/gridClassName now flow through JSON schemas … closing a gap where section wrappers couldn't be customized from schema" —— 两侧各自有意,方向相反。这不是谁忘了,是两个仓对同一条边界的理解从未对齐。

三 · ⭐ 决定性读数:两条路线不对称,卡片把它们写成对称是不准的

packages/spec/src/ui/page.zod.ts:606 与 :650 记着一条已裁定的事实(ADR-0065,ADR-0080 2026-06-30 修订):

source is runtime metadata the build-time Tailwind never scans, so utility classNames silently produce no CSS. Do not author Tailwind classes in page source in either tier.

⇒ 「补声明」意味着把一个键放上可创作面,而它最显而易见的用法(Tailwind 工具类)静默无效。 那正是本仓反复在消灭的形状 —— 一个看起来能用、实际什么都不做的创作面。台账里唯一那条 ui/PageComponent:className 也带着同一条警告。

⇒ 我的读数是:「退役读取」明显强于「补声明」,除非你判定要把 className 作为非 Tailwind(设计令牌 / 宿主自有类)的正字通道并同批把这条限制写进文档。⛔ 但我不代裁 —— 退役读取会让今天真的在 schema 里写了 className 的 SDUI 作者当场失去它,那是一个面向已有用户的破坏性变更,属人工地板。

需要你答的那一句

className / gridClassName 在 form-view / section 家族上:(a) 退役渲染器的读取(承认边界,接受破坏今天在用的 SDUI 作者),还是 (b) 补进可创作面(接受它对 Tailwind 类静默无效、须同批文档化),还是 (c) 维持现状并把「SDUI-only 但渲染器会读」写成明文约定?

定级 p3:今天无缺陷、无错误答案 —— 写了的人拿得到效果,没写的人不受影响。危害是边界的可信度:台账说「不可创作」,渲染器说「我读」,两者都不假,但一起看就没有一个可依赖的答案。

Refs:#13622(测量来源)· packages/spec/authorable-surface/(台账)· ADR-0065 / ADR-0080 修订(Tailwind 静默无效)。

---

Generated by Claude Code

[os-sam]

⚖️ RULED — maintainer, 2026-09-01, director decision batch C, verbatim 「同意」: retire the reads

The renderers stop consuming className / gridClassName off the form-view family through as any casts — the authorable boundary (which deliberately excludes free-style keys from AI-authorable metadata) is the declaration of record, and the casts that cross it are removed. Declaring the keys was weighed and not adopted: it would formally invite free Tailwind strings into authored metadata, the exact class the boundary exists to keep out. If per-view styling ever becomes a real product need, it gets an explicit design (a controlled token surface), not two leaked keys.

Enforce-or-remove order of operations: ① census first — measure actual usage of the two keys in published/authored metadata (this repo's corpus + hotcrm, publicly readable); heavy usage is evidence back to this card before any narrowing ships. ② retire the cast reads + pin the non-consumption. ③ the removal is a behaviour narrowing for anyone who used the keys — changeset per the launch-window convention, ADR-0087 disposition if any declared surface moves.

State transition in this same pass: needs-user-decision → pm:queue. domain:spec lane (the authorable-boundary ledger is its face); exact cast sites re-derived at dispatch per the card's own re-check commands.

---

Generated by Claude Code

[os-support-ai]

Claim: session_01Mciyv38maJ6HYVMiaM26T1 (domain:spec execution seat, dispatch batch 4) · branch claude/issue-13626-formview-classname-reads (objectui repo — the ruled "retire the reads" lands in objectui/packages/plugin-form; the card stays in this lane because the authorable boundary is this seat's face). Executing the 2026-09-01 batch-C ruling (comment 5486946830, verbatim scope): ① census first — actual usage of className/gridClassName on the form-view family in authored metadata across the objectstack corpus, the objectui corpus, and hotcrm (read-only clone; heavy usage = evidence back to this card BEFORE any narrowing ships); ② retire the as any cast reads (ObjectForm.tsx, DrawerForm.tsx sites per triage comment 5479541316, re-derived at dispatch) + pin the non-consumption; ③ changeset per launch-window convention; no declared surface moves (the keys were never on the ledger), so no ADR-0087 disposition expected — the gate rules. Clause-②: no (no contract-schema accept/reject change; path limb not hit — objectui repo). Declared file surface: objectui/packages/plugin-form/src/ObjectForm.tsx, DrawerForm.tsx, a non-consumption pin test, changeset; objectstack side read-only.

---

Generated by Claude Code