refactor(ui-types): one source for the filter-builder field-type list, read by the schema, the interface and the renderer; close the three K4 review should-fixes #27369

Filing gate ③ (maintainer-directed). This is the follow-up to K4 (#27145, PR #27329, merged 2026-10-29 03:51Z) of the objectui.org docs fold. In the "Restore objectui.org docs deploy" project thread on 2026-10-29, Tesla asked 「k4 从长远考虑应该怎么处理」. The dispatching seat answered: merge K4 as it is, then file one card that gives the filter-builder field-type list a single source and also takes the K4 review's three should-fixes. Tesla confirmed with 「K4 先合并，然后再单独派卡是吗」.

维护者速读

filter-builder 能接受的 14 种字段类型，现在由三处手写列表共同决定：校验 schema、TS 接口、K4 测试里冻结的那份列表。另外，渲染器里还有一份按行为划分的分组表。objectui#6939、#7562、#8774 修的都是这几份列表对不上的问题。本卡把渲染器的分组表搬到 @object-ui/types，作为唯一来源。schema、接口和渲染器都从这里引用，测试只校验两点：每个类型都有渲染分支，除 status（spec 里没有这个类型，单独列出豁免）外，每个类型都是 spec 的 FieldType。K4 合约评审的三条建议改项也一并处理。

Wanted

1. One source. Move the renderer's behaviour buckets out of ui/packages/components/src/custom/filter-builder.tsx: numberLikeTypes (number, currency, percent, rating), dateLikeTypes (date, datetime, time), selectLikeTypes (select, status) and lookupLikeTypes (lookup, master_detail, user). Put them in a module of @object-ui/types, which @object-ui/components already depends on. Export from there the full accepted list, which is those buckets plus text and boolean, 14 members today.
   • FilterFieldSchema.type in ui/packages/types/src/zod/complex.zod.ts builds its z.enum from that list, and its .describe() stops citing "the published doc".
   • FilterField.type in ui/packages/types/src/complex.ts is typed from the same as const list, not restated as a union.
   • filter-builder.tsx imports the buckets and declares none of its own.
2. The pin moves. ui/packages/types/src/__tests__/filter-builder-mirror-6939.test.ts drops the frozen LAST_PUBLISHED_TYPES list, which K4 added because its doc authority had gone. In its place it asserts that:
   • every member of the list has a branch in the renderer, using the existing per-member render measurement in examples/schema-catalog/test/filter-builder-mirror-6939.test.tsx;
   • every member is a member of FieldType in packages/spec/src/data/field.zod.ts, except an explicitly named exemption list. On main c4c515ad exactly one member needs an exemption: status is not a spec FieldType. Exempt it with a one-line reason, and do not drop it, because the accepted set must not narrow.
     Rewrite the header's objectstack#27145 section to say where the authority now lives. The accepted set must not change: still 14 members, type still optional, text still the default.
3. K4 should-fix (1). The skill pointers in skills/objectui/guides/app-composition.md:62 and skills/objectui/rules/styling.md:318 are prose URLs, and no gate checks them. Either turn them into a form that check-skills-paths.mjs or check-doc-links.mjs already judges, or add skills to a gate's scan roots. State which in the PR.
4. K4 should-fix (2). At .github/workflows/ci.yml:454 the comment says "four roots outside ui/", but the list under it has five globs. Correct the comment.
5. K4 should-fix (3). ui/AGENTS.md:106 still tells contributors to update content/docs/guide/*.md, which no longer exists. Point it at content/docs/frontend/.

These should-fixes are recorded in triage's PASS records on #27329 (6485237613 on 6a7c7794, 6486708135 on df07a80e).

Who acts

A domain:devx card thread. The file surface is:

• ui/packages/types/src/**, the new module plus the zod and interface lines;
• ui/packages/components/src/custom/filter-builder.tsx;
• the two filter-builder-mirror-6939 tests;
• skills/objectui/{guides/app-composition.md,rules/styling.md};
• ui/scripts/check-doc-links.mjs or ui/scripts/check-skills-paths.mjs and their tests, only if item 3 needs them;
• .github/workflows/ci.yml, one comment line;
• ui/AGENTS.md;
• a changeset.

Item 4 makes the PR Tier H by path: it stays a draft until green, the triage seat writes the contract review, and the maintainer merges.

Business reason. A metadata author who puts a filter-builder in a page relies on two things agreeing: the types that validation accepts and the types the renderer draws. Today a third hand-kept list sits between them. Three past cards (objectui#6939, #7562, #8774) each fixed a case where these lists had drifted apart, and validation either refused a column the renderer could draw or accepted a spelling it drew as plain text. A single source removes that class of bug. The three should-fixes leave a stale pointer for contributors and two links that no gate checks.

Acceptance

• git grep -n "LAST_PUBLISHED_TYPES\|published doc's fourteen" prints nothing.
• Adding a fifteenth member to the shared list, without a renderer branch, makes the mirror test red. Removing one member from a bucket narrows the schema, the interface and the renderer together. Record both as one-time ablations in the PR.
• pnpm --dir ui exec vitest run packages/types examples/schema-catalog is green, and the Lint UI steps are green.
• The four components-complex-filter-builder/* catalog entries and the nested search-interface entry still validate.

Duplicate search. I ran a semantic issue search for "filter-builder FilterField type vocabulary single source list" on objectstack-ai/objectstack. Its one hit, #24640 (closed, about filterableFields), is a different surface. New.

Size: S–M. Clause-②: no. Metadata-protocol: none.