fix(spec): translateObject serves validations[].message in the reader's language - #27202

📓 Docs Drift Check

3 anchor(s) derived from 1 changed package(s); no hand-written page names any of them, so this run has nothing to list — not a clean bill of health. This check sees only pages that NAME a derived anchor: one that documents this change in prose, or enumerates it in an authoring dialect, names none and stays invisible to it on every run.

# while this PR is open — GitHub drops the merge commit once it closes
git fetch origin 9c798fdb390a53e27033d0b9f186480df4c539d7 && git checkout 9c798fdb390a53e27033d0b9f186480df4c539d7
# afterwards, rebuild it from the two parents, which stay fetchable
git fetch origin 9dacccf96127755bd196e1844ce5c4cd61717d57 c8e5fee888af2f7cad5ed0b42cc4813784e0027e && git checkout -B drift-repro 9dacccf96127755bd196e1844ce5c4cd61717d57 && git merge --no-ff c8e5fee888af2f7cad5ed0b42cc4813784e0027e

node scripts/docs-audit/affected-docs.mjs --json 9dacccf96127755bd196e1844ce5c4cd61717d57

Contract review

Served-tier: CONTRACT_REVIEW_TIER
Head-sha: c8e5fee888af2f7cad5ed0b42cc4813784e0027e
Local-runs: none
Scope: packages/spec/src/system/i18n-resolver.ts + packages/spec/src/system/i18n-resolver.object-validations.test.ts, packages/objectql/src/engine-validation-locale-negotiation.test.ts, packages/rest/src/meta-object-validation-message-i18n.test.ts, ui/packages/plugin-kanban/src/ObjectKanban.tsx, .changeset/27135-object-validation-message-i18n.md; the code those hunks make claims about, read on origin/main 9dacccf9: packages/objectql/src/validation/rule-validator.ts (authoredRuleMessage, checkConditional), packages/objectql/src/validation/record-validator.ts (ValidationMessageContext, resolveFieldLabel), packages/objectql/src/engine-validation-messages.ts (validationMessageContext, negotiatedMessageLocale), packages/services/service-i18n/src/file-i18n-adapter.ts (t, resolveFromLocale, getTranslations), packages/core/src/fallbacks/memory-i18n.ts (t), packages/rest/src/meta-item-read-gate.ts (translateMetaList, translateMetaDocument, translationBundleOf, metaTranslateOptions, metaRequestLocale), packages/rest/src/meta-item-read-routes.ts (the locale-keyed ETag), packages/rest/src/meta-collection-routes.ts (Vary), packages/rest/src/form-routes.ts L300-325, packages/cli/src/utils/i18n-extract.ts (pushValidationMessages), packages/spec/src/system/translation.zod.ts (_validations), ui/packages/react/src/hooks/useSettledSchema.ts (the kanban's objectDef source), skills/objectstack-i18n/SKILL.md L167, content/docs/ui/translations.mdx L388-396, docs/adr/0029 D9.2 and D9.2a; PR #27133 (body, file list, merge ec66ac0d); card #27135 body and comments 6463784251, 6465192030, 6466141109, 6466175269; PR #27202 body, file list, both commits, compare/ac138e64e...c8e5fee88, net diff against main, the one PR comment (docs-drift bot), and the head's check-runs.

Card, PR, diff and cited code read 2026-10-28T11:04Z; the head was c8e5fee88 at every read and had not moved.

① Derived judgments

1. One key — holds. translateObjectValidations (i18n-resolver.ts L3996-4038 at the head) spells the address with objectValidationMessageKey(objectName, rule.name) and walks it as a dotted path through each locale's data; rule-validator.ts L3769 authoredRuleMessage asks II18nService.t for the output of the same function. No second spelling and no second resolver: the locale ladder is the shared localeChain + pickData (resolveBundleLocale) every overlay in this file uses, and the write path negotiates the request tag with the same resolveBundleLocale over getLocales() (negotiatedMessageLocale) before an exact t lookup, so a bare zh reaches zh-CN on both. The spec test pins that the retired top-level validationMessages and a bare objects.OBJECT.RULE.message stay unread, and builds every bundle entry at the path objectValidationMessageKey spells, parsed by TranslationBundleSchema.

2. /meta equals the refusal — holds, with three pre-existing edges shared with field labels. The chains compared: the resolver walks locale → fallbackChain = [getFallbackLocale()] → authored, skipping the chain when locale is the defaultLocale (#15711); FileI18nAdapter.t walks locale exact → fallbackLocale → the key (miss) → the validator keeps the authored text. They agree on: a rule with no name (both authored: validator L3767, resolver L4015); a bundle with no entry (both authored); an inactive or warning-only rule (the overlay translates its message regardless of active/severity; no refusal is emitted for such a rule, so there is no second sentence to disagree with, and the kanban asks the gate before naming a rule); an extended object (both key on the merged object's name, and pushValidationMessages writes the same address for an objectExtensions rule, as i18n-standard.mdx documents); a tenant/org authored translation layer (translationBundleOf reads getTranslations, the merged static+authored view; t probes the same two maps authored-first — one precedence); a conditional rule (checkConditional hands the branch to evaluateRule, so the refusal carries the branch's own name and message; the overlay translates the branches and leaves the wrapper authored, pinned with a wrapper entry present in the bundle).
Where they can differ, none introduced here and each shared with field labels through the same lookupObjectFieldAttr / resolveFieldLabel pair: (a) a request for the default locale whose own bundle lacks the entry while the fallback locale's has it (the dev declared this); (b) an empty-string entry in the requested locale — t returns '' and takes no fallback leg, so the validator keeps the authored text, while localeChain walks on and may serve the fallback locale's string; (c) createMemoryI18n().t reads the DEFAULT locale's bundle before fallbackLocale when the requested locale has no bundle at all. Recorded, not failed: the card's Expected is "the same way it overlays field labels", and that is what this is. The depth bound MAX_VALIDATION_DEPTH = 10 equals the extractor's; the validator has none, so a branch nested deeper than ten is translated by the refusal and not by /meta — no extracted key exists for it either; negligible, noted.

3. Nothing else moves — holds. label/pluralLabel/description, fields, actions and listViews are computed before the new call and are untouched; the test pins fields and actions equal with and without validations, and no validations key added to a document without one. /meta's shape is unchanged: the same array, each rule spread with only message replaced, and the authored array by reference when nothing matched. Caching: the item ETag is simpleHash(locale + JSON.stringify(item)) (meta-item-read-routes.ts L649) and both transports send Vary: Accept-Language (meta-collection-routes.ts L219, L634) — unchanged by this PR, so no locale's translated message can be served from another locale's cached /meta. One door the PR body's reader measurement does not name: form-routes.ts L317 hands the public form's object schema to translateMetadataDocument('object'), so the anonymous form's served validations[].message is now translated too — the same direction as the refusal of the same submit, and the fix's intent; recorded. formula-option-labels.ts L150 and service-analytics L1168 pass { name, fields } only — unaffected.

4. The kanban comment — true at this head; patch honest. objectDef is useSettledSchema(schemaKey, dataSource), the adapter's /api/v1/meta/object/:name read, so ObjectKanban.tsx L1838-1840 now states what /meta serves: the bundle's _validations.RULE.message for the request locale, the refusal's string (edges in ①2 aside). compare/ac138e64e...c8e5fee88 is one commit: the three comment lines and three changeset lines, nothing else. @object-ui/plugin-kanban is published (18.2.0, not in .changeset/config.json ignore), so the patch line is owed and is comment-only, as declared.

5. Declarations — hold. Clause-②: no: no Zod schema, no export, no ObjectLike member moved; no api-surface/ or authorable-surface/ shard is in the file list and check:generated had nothing to write. Metadata-protocol: none: no authorable key, value or meaning changes — the bundle key's meaning (the sentence a rejected write returns) is unchanged, it gains a second reader. patch: kind none on a behaviour fix. Skills-docs: none: skills/objectstack-i18n/SKILL.md L167 ("message only — the sentence a rejected write returns") stays true; no skills/** or .claude/skills/** text teaches that /meta serves the message untranslated (grep _validations, objectValidationMessageKey, untranslated); the row now understates the key's reader set — a docs follow-up, not a false sentence. content/docs/ui/translations.mdx L388-396 likewise stays true. Fixes #27135 is complete: the card's Expected (overlay from the one key; authored on a miss) and its "is a server-side test owed?" question are both answered.

6. Tests — pin the claims; nothing weakened. Spec (12): the overlay in zh-CN; zh → zh-CN; the declared chain then authored; the translateMetadataDocument('object') door; no input mutation; nameless rule, missing entry, empty entry, no bundle; no other spelling read; fields/actions unchanged; no key added. ObjectQL (+1): translateObject(..., { locale, fallbackChain: ['en'] }) as /meta calls it, against the real engine's refusal (ql.update → authoredRuleMessage) for zh-CN, zh, en, de, in one table. REST (3): GET /meta/:type/:name and GET /meta/:type under zh-CN and en through RestServer with i18nServiceProvider as the 14th constructor argument (position verified against rest-server.ts). Not pinned: edges (a)-(c) above and the authored-translation layer — all pre-existing paths. The diff's 3 deletions are comment lines (kanban 2, docblock 1): no existing assertion changed. The ablation (anchor line deleted → 5/12 spec, 1/7 objectql, 2/3 rest red; restored green) is the dev's reading on ac138e64e, read, not re-run.

7. Gates — green on the head. Check-runs on c8e5fee888af… were read four times: at 2026-10-28T10:44Z (44 runs, 23 in_progress), 2026-10-28T10:55Z (65 runs, 23 in_progress), then every two minutes until 2026-10-28T11:03Z: 67 runs, 57 success, 10 skipped (opt-in, matrix-stub and schedule-gate rows), 0 failure, 0 in_progress. The eight required contexts are each success: Lint & Repo Gates, TypeScript Type Check, Test Core (8/8 shards), Build Core, Governed Surface Queue Guard, Lint UI, Type Check UI, Build & E2E UI; Check Changeset, PR-body guards, Spec property liveness, Dogfood Regression Gate and Test UI (16/16) too. The dev's 85/85 derived gates with exit 0 on c8e5fee88 and the suites and ablation on ac138e64e are read as reported, not re-run (Local-runs: none). Head unchanged at every read.

② Semver level

@objectstack/spec patch: a served-document behaviour fix with Clause-②: no and Metadata-protocol: none — kind none takes patch. @object-ui/plugin-kanban patch: a source comment, and a published package whose src/** the diff moves. One changeset, both lines, both declarations present. Correct.

③ Boundary flags

F1 — the dev's open question (ADR-0029 D9.2a): answered — A, a declared deviation, right for this card. D9.2a binds the three scalars and says in its own first paragraph that for fields, validations and indexes the fold "still does" run as before: validations is named as outside the clause, so A contradicts no ADR. What A deviates from is this translator's own convention (translateObject docblock: field description "follows the rule above, as every string this translator added since does" — description, inlineTitle and the embedded listViews are each judged by valueOverridesPackagedBase). The deviation is right because the comparison is a property of the PAIR of readers, not of one: authoredRuleMessage has no packaged base and applies the catalog unconditionally, so B would serve the tenant's rewrite on /meta and the packaged translation in the refusal — the card's split re-opened for exactly the rewritten rules — and field label, the card's named precedent, is already flat catalog ?? document here. The deviation is declared where it lives (the helper's docblock "No packaged-base comparison") and in the acceptance notes. What it leaves, stated plainly: under a non-default locale, a tenant who rewrote a rule's message in a D9 overlay (same rule name) now sees the packaged translation on /meta as well as in the refusal, where before this PR /meta showed the rewrite. The #8284 sentence "an explicit override beats a packaged default" holds on neither surface for rule messages; the refusal half pre-exists. The complete answer is D9.2a on BOTH readers — the validator comparing against the packaged owner's rule (the registry already holds getPackagedObjectBase), then this overlay following the convention — a validator-side card outside this claim's surface. The lane seat files it as a card (the #8284 class, on a surface the 2026-08-13 ruling did not reach); the acceptance note meets Prime Directive 10's floor, so this is owed, not a condition of the verdict. Nit: the docblock declares the deviation without the ADR's id; one ADR-0029 D9.2a token there would make it greppable (Prime Directive 13).

F2 — the default-locale edge (dev-declared): answered in ①2(a). The #15711 contract, pre-existing and shared with every overlay; not this PR's to move.

F3 — a ui/ path in a domain:spec PR: within the amended claim. Comment 6466175269 widened the surface to the one comment and the changeset line; the file list matches it exactly. No governed path (skills/**, docs/adr/**, .claude/**, .github/workflows/**) is in the file list; Governed Surface Queue Guard success.

F4 — branch claude/project-thread-h0y6db, not claude/issue-27135-*: declared in the claim (platform-assigned thread branch, written as measured); the Claim: comment decides ownership and the PR-body guards passed. Not a flag.

F5 — not re-measured here: the dev's reading that Studio's rule editors run on the draft or the /layers baseline, so a translated message cannot be saved back. I did not re-measure it; the class is the ADR-0029 D9.2a "honest edge" already recorded for every translated string the served body carries, unchanged in class by this PR.

Implemented-by: claude/project-thread-h0y6db
Reviewed-by: session_01WGufzgbJhFoXYsLSe5WSPP

VERDICT: PASS

维护者速读 (席位终稿)

• 改了什么：translateObject 现在也翻译校验规则的提示语 validations[].message。它读翻译包里 objects.<对象>._validations.<规则>.message，和服务器拒绝写入时用的是同一个键（objectValidationMessageKey）。所以 /meta 和服务器的拒绝提示，在同一语言下显示同一句话。看板关闭列的提示（fix(plugin-kanban): a lane that refuses the dragged card says why #27133）随之显示译文，那段过时的代码注释也一并改正。
• 为什么改：中文、日文等部署里，同一条规则的说明，服务器提示是译文，看板上却是原文，用户会读到两种语言的两句话。
• 风险与回滚：没有 schema 变化，只是读取时多一层翻译覆盖。没有译文的规则保持原文。fields、actions 的翻译不变，缓存按语言区分也不变。回滚 = revert 本 PR。
• 先看：packages/spec/src/system/i18n-resolver.ts 里新增的 translateObjectValidations。
• 两分钟验证：pnpm --filter @objectstack/rest exec vitest run meta-object-validation-message-i18n 预期全绿（同一规则、同一语言下，/meta 与拒绝提示一致）。
• 需要你做的：无，席位落地。

席位意见：接受。契约复核 PASS（6467333940），确认只有一个键、/meta 与拒绝提示在各种情况下一致、其它翻译不受影响。开发提出的问题（校验提示不做“已打包原文”比较）复核判定为有声明的偏离，而且是对的：只改一边会重新造成两句话不一致。租户改写过的规则提示会被打包译文盖住，这是本 PR 之前就有的问题，另立卡 #27221，两边一起修。

席位核对 · ACCEPT · domain:spec seat 3 · os-support-ai · session session_01WGufzgbJhFoXYsLSe5WSPP · 2026-10-28T11:07Z

Head: c8e5fee888af2f7cad5ed0b42cc4813784e0027e · card #27135 · claim 6465192030, amendment (the kanban comment) · triage 6463784251 (one key)

1. 形态: a draft PR to main, Fixes #27135. 6 paths, +535 / −3:

   • i18n-resolver.ts and its test;
   • an objectql and a rest test;
   • the kanban comment;
   • one changeset.

   All are inside the amended claim surface. check-governed-merges --pr 27202 reports NOT governed (538 lines), so the ordinary queue landing applies. The assignee is os-support-ai.

2. 前提实测: the premise holds on main. translateObject overlays fields and actions but not validations[].message, while the rule validator translates the refusal through objectValidationMessageKey. The review read the head and confirmed four things:

   • translateObjectValidations reads that key verbatim, through the shared locale ladder.
   • /meta equals the refusal for a nameless rule, a missing entry, inactive / warning-only rules, an extended object, the tenant authored layer and conditional branches. Three edges can differ ([Decision] the default locale's text — is the authored label the default-locale source (a default-locale request stops there), or must every supported locale ship a bundle? os i18n check says the former, localeChain does the latter; the resolver's literal ['en'] default rides on the answer #15711 default-locale skip, an empty-string entry, the memory-i18n default leg); all are pre-existing and shared with field labels.
   • The fields / actions / listViews overlays are untouched.
   • The ETag stays locale-keyed, with Vary: Accept-Language.

3. 门禁 check-runs (on this head): 67 runs, 57 success, 10 skipped, none red or pending, read at 2026-10-28T11:03Z. All eight required contexts are green. Dev-side readings:

   • resolver suites and the ablation on ac138e64e;
   • dispatch gates on c8e5fee88: 85 derived, 85 exit 0, 0 NOT-MEASURED.

   c8e5fee88 adds only the kanban comment and its changeset line.

4. 接受的偏差: (a) Validation messages use a flat catalog ?? document, with no ADR-0029 D9.2a packaged-base comparison. This is declared in the helper docblock and in Acceptance notes, and the review judged it right for this card: D9.2a names validations as outside its clause, and comparing on one reader alone would re-open the split. (b) The branch name is the platform-assigned thread branch.

5. 范围外去向: a tenant-rewritten rule message loses to the packaged translation under a non-default locale. That is pre-existing on the refusal and now consistent on /meta; it is filed as i18n: a tenant-rewritten validation rule message loses to the packaged translation under a non-default locale, on both the refusal and /meta — apply ADR-0029 D9.2a to both readers #27221 (D9.2a on both readers). The public form door (form-routes.ts ~L317) now serves the translated message too, in the same direction as its refusal; that is named in the body. The review's nit (cite ADR-0029 D9.2a in the helper docblock) is not owed for landing.

6. 落地条款: the seat checked the prose faces against the diff on this head: the PR body (with the second door and i18n: a tenant-rewritten validation rule message loses to the packaged translation under a non-default locale, on both the refusal and /meta — apply ADR-0029 D9.2a to both readers #27221 added in this round), the changeset, and the kanban comment. The seat readies the PR and arms auto-merge in this act, then follows it to MERGED.