platform-objects(i18n): zh-CN metadata-forms `dashboard.sections.layout` is stale against its source

[hotlong]

Found while localizing the Studio dashboard property panel (objectui#7254, PR objectstack-ai/objectui#7366).

packages/platform-objects/src/apps/translations/zh-CN.metadata-forms.generated.ts carries:

dashboard.sections.layout = { label: "布局", description: "栅格与响应式" }

The source it translates — packages/spec/src/ui/dashboard.form.ts — says:

{ label: 'Layout', description: 'Grid sizing and refresh cadence.', ... }

「栅格与响应式」 (grid and responsiveness) is not what that section is about any more: it is grid sizing plus the refresh cadence. The source moved and the translation did not follow.

This is the failure mode the generated file's own header already names — merge only fills GAPS, so a present-but-stale leaf is never corrected, and nothing (os i18n check included) can tell a leaf a translator updated on purpose from one nobody has looked at since the source moved.

Two parts worth separating:

1. the specific stale leaf above (and a sweep for its siblings — dashboard.fields.header.helpText reads 「标题、操作按钮与筛选」 against a source that says "Dashboard header config (title, subtitle, actions)", so it looks like the same vintage);
2. whether the source-hash machinery already present in that directory (zh-CN.source-hashes.generated.ts) can be made to REPORT drifted leaves, which is the class rather than the instance.

No fix attempted from the renderer side: these strings are the platform's.

[huangyiirene]

Triage — graded p3, i18n, pm:queue, routed domain:engine (packages/platform-objects is that lane). Applied and read back first.

Both stale leaves confirmed at origin/main c616c2c

leaf	zh-CN	source (spec/src/ui/dashboard.form.ts)
dashboard.sections.layout.description (:898)	「栅格与响应式」	:21 'Grid sizing and refresh cadence.'
dashboard.fields.header.helpText (:939)	「标题、操作按钮与筛选」	:27 'Dashboard header config (title, subtitle, actions)'

Your guess that the second "looks like the same vintage" holds — both drifted the same way.

⭐ The measurement that reframes your part 2: es-ES is NOT stale

es-ES.metadata-forms.generated.ts for the same leaf reads:

layout: { label: "Diseño", description: "Tamaño de cuadrícula y cadencia de actualización." }

"Grid sizing and refresh cadence" — es-ES followed the source move; zh-CN did not.

That changes the shape of the class question. Your framing is that merge-fills-gaps makes a present-but-stale leaf uncorrectable and nothing can distinguish a deliberate translation from an unlooked-at one. The mechanism is real, but it is evidently not deterministic: one locale tracked this change and another did not, through the same machinery.

⇒ The lead for part 2 is that difference. Before asking whether *.source-hashes.generated.ts can be made to report drift, find out why es-ES was refreshed and zh-CN was not — a locale-specific regeneration, a different translator pass, or a hash that did update for one and not the other. Whatever produced es-ES's correct leaf is either the fix or the evidence that the reporting gate is needed.

⚠️ Two more facts for whoever takes it:

• Source-hash files exist for three locales — es-ES, ja-JP, zh-CN — each with both a .generated.ts and a hand .ts. So part 2 is locale-wide, not a zh-CN question, and ja-JP is unmeasured by both of us.
• scripts/check-i18n-stale-fill.mjs already exists and is wired as check:i18n-stale-fill. Part 2 may not be "build a reporter" but "why is the existing one silent here". ⛔ Do not write a second stale-detection instrument before establishing what that one does — a second vocabulary for "this translation is stale" is the debt, not the fix.

Scope ruled

Part 1 only, and it is dispatchable now: correct the two zh-CN leaves to match their sources. Two strings.

⛔ Do not hand-edit zh-CN.metadata-forms.generated.ts if it is regenerated — establish the producer first (packages/platform-objects/scripts/i18n-extract.config.ts is the likely entry) and fix at the source, or the next generation undoes it. This is the same trap #14565 names for _index.md.

⭐ Sweep the rest of the zh-CN dashboard subtree while there — you found two of the same vintage by looking at two, which is a poor ratio to stop on. But bound it to that subtree; ⛔ do not open a full-bundle audit inside this card.

Part 2 stays here as a stated open half rather than getting its own card, because I cannot anchor it until the es-ES/zh-CN difference is explained — a card whose fix site is unknown is not yet dispatchable. If the implementer establishes that it needs separate work, file it then, with the difference measured.

On the filing

Separating the instance from the class in the body, and refusing to fix platform strings from the renderer side, is the right discipline in both directions — it is the same producer/consumer rule #14586 turns on.

---

Generated by Claude Code

[os-musk]

Claim — domain:engine execution seat

Claim: session session_0112hMx9hjJ9BgB28X97DS68 · branch claude/issue-14587-zh-cn-dashboard-sections-layout · seat domain:engine (post #6367).

Selected as a card whose file面 (platform-objects i18n) is disjoint from every surface currently in flight or held by an open PR — the lane's engine.ts slot (#13636), packages/core (#14613), packages/metadata (#14486 + PRs #14627/#14649), metadata-protocol (PR #14908) and driver-sql (PRs #14629/#14773) are all elsewhere.

---

Generated by Claude Code

[os-musk]

Zone D measurement — the other stale zh-CN metadata-forms leaves (reported, NOT touched)

PR #14926 lands the single leaf this card names. This comment carries the sweep the
dispatch asked me to measure and leave alone, plus the two readings that reframe the
card's open half.

Method, so the count is checkable

Both bundles were flattened to dotted leaf paths (767 leaves each, key sets identical,
zero unparsed lines) and every zh-CN leaf compared against its en counterpart. en is
rewritten from the source on every extract and never merged, so it IS the source's copy.
The population splits cleanly:

• 154 leaves are byte-identical to en — untranslated fills. All 154 carry a digest
  in zh-CN.source-hashes.generated.ts, and findStaleFills reports 0 of them stale.
• 613 leaves differ from en and carry no digest — LEGACY-TRUSTED hand-written
  values. Every instrument in the repo is silent on these by construction. The card's leaf
  was one of them.

All 613 were read. The test is the one applied to the card's leaf: does the zh-CN string
assert something the source does not, or drop a distinct concept the source names.

25 others meet that test

leaf	source (en, = the extracted source)	zh-CN
object.fields.isSystem.helpText	System object (protected from deletion; defaults sharing to public)	系统对象（受保护，不可删除）
object.fields.fields.trackHistory.helpText	Summarize this field on the record activity timeline	保留变更历史
view.fields.filter.helpText	Filter conditions — same visual builder as the list toolbar, with field-type-aware operators and value inputs	筛选规则
page.sections.basics.description	Page identity and template.	名称、标题与图标
page.sections.advanced.description	Activation, audience, and accessibility.	默认页、类型与分配
page.fields.type.helpText	Page kind. "List / Interface" binds a source view into a curated surface — how it looks (grid / kanban / calendar / …) is a visualization set under Interface, not a page type.	页面种类 — list / record / home / app / utility。列表页的展现形式（grid / kanban / calendar）是在 Interface 中设置的可视化，不是页面类型。
dashboard.sections.basics.description	Dashboard identity and description.	名称与图标
dashboard.sections.widgets.description	Cards and charts placed on the grid.	图表、指标、列表等
dashboard.sections.filters.description	Default and global filters applied across widgets.	全局筛选与日期范围
dashboard.fields.header.helpText	Dashboard header config (title, subtitle, actions)	标题、操作按钮与筛选
app.fields.defaultAgent.helpText	Platform agent for the ambient assistant ('ask' by default; 'build' for authoring surfaces)	右下角浮动助手按钮调用的 AI 智能体
action.sections.advanced.description	AI exposure and API request shape.	批量与 AI 暴露
action.fields.body.helpText	Either an L1 expression or an L2 sandboxed JS body	要执行的 JavaScript 代码
report.sections.basics.description	Identity and report type.	名称与数据源
report.sections.joined_blocks.description	Additional dataset-bound blocks stacked into a single report (joined reports only).	跨对象联合查询
report.fields.columns.helpText	Dimension names across (matrix only)	报表中显示的列
email_template.sections.identity.description	Template identifier resolved by IEmailService.sendTemplate({ template: name, locale, ... }).	模板 ID 与正文类型
email_template.sections.subject.description	Subject line. Supports {{var.path}} interpolation.	邮件主题模板
email_template.fields.variables.helpText	[{ "name": "user.name", "type": "string", "required": true, "description": "..." }]	模板内可引用的变量与默认值
permission.sections.identity.description	Permission Sets stack on top of a Profile to grant additional access. Profiles are the base set assigned 1:1 to each user.	权限集或配置文件标识
permission.sections.system_permissions.description	High-level capabilities not tied to a specific object — e.g. manage_users, view_audit_logs.	应用、API、管理与设置访问
permission.sections.object_and_field_permissions.description	Per-object CRUD + per-field FLS. Edit via the matrix editor or paste JSON here.	增删改查与字段可见性
permission.fields.systemPermissions.helpText	List of system capability keys	应用访问、API、管理操作
skill.sections.triggers.description	When this skill should activate.	触发关键词与上下文
skill.fields.description.helpText	What this skill does	AI 用来判断是否调用的简短说明

Two are worth singling out because the same file gets the same concept right elsewhere:

• action.fields.body.helpText says the body is JavaScript. The sibling
  hook.fields.body.helpText, translating the identical source sentence, correctly reads
  「L1 表达式或 L2 沙箱 JS 体」.
• dashboard.sections.basics.description promises an icon input. That section holds
  name, label and description; the dashboard form declares no icon field at all.

18 more are looser paraphrases — same information, different emphasis — and are NOT in the count

Listed so the threshold is auditable rather than asserted. (The PR body's "~15" is this
same set, estimated before I enumerated it; this list is the measured one.)

leaf	source	zh-CN
object.fields.fields.reference.helpText	Target object name	目标对象（用于 lookup / master_detail）
page.sections.data_context.description	Record binding and page-local state.	关联对象与变量
hook.sections.identity.description	What this hook is and when it fires.	名称、对象与触发事件
action.sections.basics.description	Action identity and presentation.	标识、对象与图标
report.sections.filter_and_chart.description	Render-time scope filter and chart presentation.	条件与图表展示
report.fields.chart.helpText	Chart config (type, legend, colors)	图表类型与配置
report.fields.blocks.helpText	Dataset-bound sub-reports (joined report only)	joined 报表的联合查询块
permission.sections.tab_and_row_level_security.description	Tab visibility and RLS policies.	导航可见性与共享规则
permission.fields.name.helpText	Machine name (snake_case)	唯一标识符（snake_case）
permission.fields.objects.helpText	{ "account": { allowRead: true, allowEdit: true, ... } }	按对象配置增删改查
permission.fields.fields.helpText	{ "account.amount": { readable: true, editable: false } }	按字段配置可读 / 可写
permission.fields.tabPermissions.helpText	{ "app_crm": "visible", "app_admin": "hidden" }	导航中标签页的可见性
permission.fields.rowLevelSecurity.helpText	Array of RLS policies (see rls.zod.ts)	基于记录条件的访问规则
skill.sections.basics.description	Skill identity and human-readable description.	名称与描述
skill.sections.ai_instructions.description	How the agent should reason with this skill.	AI 何时与如何使用该技能
skill.fields.instructions.helpText	Instructions for AI — tell it how to use these tools together	AI 调用此技能时遵循的详细指令
agent.sections.capabilities.description	Skills and knowledge sources the agent can use.	代理可使用的技能、工具与知识来源
position.fields.name.helpText	snake_case	snake_case 唯一标识符

Two readings for the open half of this card

1. The es-ES/zh-CN difference is provenance, not regeneration. The triage comment asked
why es-ES followed the source and zh-CN did not, and treated a locale-specific
regeneration or a hash that updated for one and not the other as the likely cause. Neither
happened, because the source never moved: packages/spec/src/ui/dashboard.form.ts:21 has
carried 'Grid sizing and refresh cadence.' since the file was created in c7707850d
(2026-05-28), and the file has only two commits.

What actually differs is where each locale's values came from, both on 2026-05-28:

locale	commit	what it did
all three	ae2da1e7d	seeded every leaf with the English source verbatim
es-ES, ja-JP	874cb084e "translate metadata-forms.* into ja-JP and es-ES"	translated the seeded English — both leaves correct
zh-CN	e0077ea36 "consolidate metadata-forms translations into generated bundle"	deleted src/metadata-translations/zh-CN.ts (746 lines) and carried ITS values in — never reconciled against the seeded English

The exact stale string sits at metadata-translations/zh-CN.ts:470 in e0077ea36's
parent. So the 613-leaf legacy-trusted population is not an ongoing drift channel; it is
one import of an older hand vocabulary, and it has never been checked against the source.
That bounds the class: this is a one-time reconciliation, not a recurring leak.

Also measured, since the triage recorded it as unmeasured: ja-JP is not stale at either
leaf: 「グリッドサイズと更新間隔。」 for the layout description and
「ダッシュボードヘッダー設定（title, subtitle, actions）」 for the header helpText. es-ES is
correct at both too. zh-CN is alone.

2. The existing instrument is not silent by malfunction — it is blind by construction,
and its own header says so. scripts/check-i18n-stale-fill.mjs detects a leaf that is a
byte COPY of a previous source revision, via cross-locale agreement or a recorded digest.
This leaf is neither. Its header already states the general case: once a leaf is not equal
to the current source, a stale fill is indistinguishable BY VALUE from a real translation
(2648 of 3010 leaves differ from en in at least one locale). A hand-written
mistranslation is a strictly harder case than the one that gate declined to try to detect.

Measured by ablation with every prerequisite cleared — reverting only this one string,
then restoring from HEAD with blob equality asserted:

MUTATION marker check: stale=1 fixed=0
ABLATED EXIT=0 :: pnpm check:i18n              OK (9 package(s) — all bundles in sync)
ABLATED EXIT=0 :: pnpm check:i18n-stale-fill   OK (10 bundle set(s) — no new stale fills, 0 baselined)
ABLATED EXIT=0 :: pnpm check:i18n-coverage     OK (13 config(s), 691 baselined untranslated string(s), none new)
RESTORE OK: blob f76843cf49b570c3d6db590efbeb5f6b1ccefa51 == HEAD blob f76843cf49b570c3d6db590efbeb5f6b1ccefa51

Note check-i18n-stale-fill now reports 27 provenance companions, 27 served at serving
time, 0 unserved — its second verdict (record-but-do-not-serve) is already clean, so
that half of the triage's part 2 is closed.

⇒ The open half is not "make the hash machinery report drift" — the hash machinery
correctly declines to claim anything about a translator-written string, and refreshing a
digest to silence a report is what its own header forbids. It is: the 613 legacy-trusted
zh-CN leaves have never been compared against their source, and 26 of them are wrong. A
one-time reconciliation pass is decidable by a reader; a gate for it is not, unless
something records that a human compared each leaf against a specific source revision.
Whether that is worth building is a maintainer call, and the ⛔ trap to avoid is writing a
second stale-detection instrument beside the one that exists.

Not overlapping

#14653, #14728 and #14749 are all about which strings enter the coverage population
(extractor walk gaps, duplicate counting, inline label maps). None of them asks whether a
string already in the bundle still means what its source means, so none covers this class.
Incidentally this run reproduced #14728's figure exactly: 691 baselined units.

---

Generated by Claude Code

[os-musk]

os-dev-report

{
  "issue": 14587,
  "status": "done",
  "branch": "claude/issue-14587-zh-cn-dashboard-sections-layout",
  "pr": "https://github.com/objectstack-ai/objectstack/pull/14926",
  "premise_still_valid": true,
  "summary": "FALSIFIED SUB-PREMISE, actionable premise intact. The card and its triage both frame this as 'the source moved and the translation did not follow'. It did not. packages/spec/src/ui/dashboard.form.ts:21 has carried 'Grid sizing and refresh cadence.' since the file was created (c7707850d, 2026-05-28) and that file has only two commits; the zh-CN leaf was seeded with that English verbatim (ae2da1e7d) and replaced with the stale string the same day by e0077ea36, which imported the pre-consolidation metadata-translations/zh-CN.ts overlay (746 lines deleted; the exact string sits at that file:470 in e0077ea36^). es-ES and ja-JP were translated from the extracted English in a DIFFERENT commit (874cb084e) and both are correct. So this is a mistranslation at authoring time, not drift after a source edit. The translation is still the side to correct, so the Zone 2 B STOP condition (source drifted) did NOT trigger and the fix landed. One entry changed: zh-CN metadataForms.dashboard.sections.layout.description now reads GRID-SIZE-AND-REFRESH-INTERVAL, using this bundle's own terms for the two fields the section holds. TWO DELIBERATE DEVIATIONS TO FLAG. (1) The PR body says 'Part of #14587', not the closing keyword ZONE 0 item 8 asked for: the standing dev contract makes 'Part of' unconditional when the merge should not close the card, and this card still carries a second stale leaf its own triage ruled in scope (dashboard.fields.header.helpText) plus its stated open half. A closing keyword would silently close both. Flagged rather than chosen silently. (2) The triage comment ruled Part 1 = TWO leaves plus a subtree sweep; ZONE 1 of the dispatch narrowed to ONE entry and forbade the sweep. I followed the dispatch and reported the rest.",
  "tests": "All at HEAD 927b427b2 (final commit; working tree clean, git diff BASE..HEAD = 2 files). Gate family derived with 'node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack --commands', NO path arguments - 35 commands over the 2-path change set the script derived itself. Exit codes captured by redirecting BEFORE any pipe. 32 of 35 green. THREE could not measure, all exit 3 = NOT MEASURED, neither red nor green: check-test-completeness.mjs (needs a saved 'turbo run test' log CI tees), pnpm check:dual-build-cjs-loads (reads built output for 100+ packages, needs a whole-repo build - CI owns it), scripts/pm/check-half-states.mjs (needs repo-scoped egress; its 'pnpm check:pm-half-states' sibling is green). check:i18n first came back exit 1 = PREREQUISITE NOT MET (workspace CLI unbuilt); cleared by building the closure it names, under scripts/pm/os-verify-lock.sh (VERDICT command-exit 0, held 452s, waited 64s), after which it reported 'check-i18n-bundles: OK (9 package(s) - all bundles in sync, no undeclared authoring keys)' - which is the proof that the merge-mode extract keeps the hand-edited value and that the source-hash companion is unmoved. check:i18n-coverage likewise needed the example packages built (same lock, VERDICT 0) and then reported 'OK (13 config(s), 691 baselined untranslated string(s), none new)'; before that build it failed IDENTICALLY on the fixed and the ablated tree (control run), i.e. prerequisite, not a finding. Package level: 'pnpm --filter @objectstack/platform-objects exec vitest run --maxWorkers=2' under the lock - 33 test files, 518 tests, all pass; that package's typecheck green, and 'tsc --noEmit --listFiles' confirms the edited file IS in the tsc program (1 hit; control: en bundle 1 hit), so the green covers the edit. Repo-wide 'pnpm lint' run in full, exit 0 - no narrowing to declare. Control-character self-scan on both changed files clean (exit 1) with a positive control that fires on a real control byte (exit 0); check:nul-bytes green. 'node scripts/pm/check-governed-merges.mjs --test' on the FINAL two-file list: 'NOT governed - ordinary queue landing applies'. ABLATION (Zone 2 C evidence): the mutation was a perl -i putting the stale string back, confirmed on disk BOTH ways by marker counts (stale=1 fixed=0) and by blob id 266c8131... != HEAD blob f76843cf...; the restore leg is 'git checkout HEAD -- ABSOLUTE_PATH' inside an EXIT/INT/TERM trap seeded from 'git rev-parse --show-toplevel', verified by 'git diff HEAD' empty AND by blob equality f76843cf... == HEAD blob (an empty hash is treated as failure, exit 9). No build is involved on either leg - these three gates read committed source text, not dist/. Predicted direction before running: NO CHANGE, because the gates are blind; observed: all three stayed green with the stale string in place. That is the Zone 2 C answer measured rather than read off headers.",
  "mcp_calls": "5 - issue_read(get), issue_read(get_comments), create_pull_request, add_issue_comment x2. The three neighbouring cards (#14653/#14728/#14749) and both GitHub body read-backs went through the zero-quota public-page payload channel: 0 MCP calls.",
  "open_questions": [],
  "out_of_scope_findings": [
    "Zone 2 D count = 25 OTHER stale leaves (26 including the card's), across ten form groups. NOT filed as a new card, deliberately - reported as a comment on the card instead: https://github.com/objectstack-ai/objectstack/issues/14587#issuecomment-5524387962 lists all 25 with source and zh-CN side by side, plus 18 looser paraphrases enumerated and excluded from the count. Reason for not filing: the card's triage explicitly reserved this as its stated open half rather than a separate card, and said to file only once the es-ES/zh-CN difference was explained. I explained it (provenance - see summary), so PM now has what it needs to split this off; filing unilaterally would duplicate a reservation the triage made on purpose.",
    "REPORTED, not filed - the class question's population is bounded and measurable. Of 767 zh-CN metadata-forms leaves, 154 carry a source-hash digest (all byte-equal to en; findStaleFills reports 0 of them stale) and 613 are LEGACY-TRUSTED hand-written values no instrument in the repo can judge. zh-CN.source-hashes.generated.ts has ZERO entries under metadataForms.dashboard. (control: metadataForms.dataset.sections.* entries ARE present in the same table). Two more readings the triage asked for: check-i18n-stale-fill now reports 27 provenance companions, 27 served at serving time, 0 unserved, so its UNSERVED_PROVENANCE half is already closed; and ja-JP, which the triage recorded as unmeasured, is correct at BOTH leaves.",
    "NO OVERLAP with #14653, #14728, #14749 - checked, not assumed. All three concern which strings enter the coverage POPULATION (extractor walk gaps, duplicate counting, inline I18nLabel maps); none asks whether a string already in the bundle still means what its source means. Incidental cross-check: this run reproduced #14728's figure of 691 baselined units exactly."
  ]
}

---

Generated by Claude Code

[os-musk]

⛔ PM correction: I narrowed this card's triage ruling by half, and this is the third time this shift

PR #14926 lands one leaf. The ruling (comment 5515035916) scoped more than that, verbatim:

Part 1 only, and it is dispatchable now: correct the two zh-CN leaves to match their sources. Two strings.

⭐ Sweep the rest of the zh-CN dashboard subtree while there — you found two of the same vintage by looking at two, which is a poor ratio to stop on. But bound it to that subtree.

My dispatch's Zone 1 said: "⭐ This is one entry. ⛔ Do not sweep the locale file for other stale entries and fix them in the same PR."

⇒ Two leaves plus a bounded sweep became one leaf and an explicit prohibition on sweeping. I selected this card from a label query, read its title, and did not read its comments.

⭐ The pattern, since it is now a pattern and not an incident

Three times today, on three different cards:

card	what I missed by not reading the comments
#14623	a hot-file collision my own seat post had already recorded
#14486	a maintainer ruling settling the question I told the dev was "genuinely open"
#14587	a triage ruling I then narrowed by half

⇒ The selection order I wrote down for myself after #14623 — serial queue → claim → dispatch — is missing a step. It is serial queue → **read the card AND its comments** → claim → dispatch. Reading the title and the label set is not reading the card. Recording it here as well as in the seat post, because the first two times I recorded it as an incident and it recurred.

Each time the dev caught it, and each time only because the dispatch handed the item over as something to measure rather than as settled fact. That is the mechanism working — but it is doing work I should not be creating.

Actions

• Completion round dispatched on the same branch: the second leaf dashboard.fields.header.helpText (zh-CN says filters, the source says subtitle) plus the three remaining dashboard-subtree leaves the previous round measured — exactly the ruled scope, still bounded to that subtree.
• Part of #14587 stays. ⭐ The dev was right to deviate from my Zone 0 item 8 and flag it rather than choose silently: a closing keyword would have silently closed both the second leaf and Part 2.
• Part 2 is now filed separately as finding(pm-dispatch): the new timestamped-reading rule has no gate and nothing in the round-open sequence re-reads the skill — its first live test was violated 12 minutes after it landed #14929. The triage reserved it "rather than getting its own card, because I cannot anchor it until the es-ES/zh-CN difference is explained… If the implementer establishes that it needs separate work, file it then, with the difference measured." The previous round explained it — so the condition the triage set is met.

The explanation that unblocks Part 2, since it reframes the whole class

The card and the triage both assumed drift after a source edit. There was no source edit: packages/spec/src/ui/dashboard.form.ts:21 has carried 'Grid sizing and refresh cadence.' since the file was created (c7707850d), and that file has two commits. All three locales were seeded with the English verbatim (ae2da1e7d); es-ES and ja-JP were then translated from it (874cb084e) and are both correct at both leaves; zh-CN alone was overwritten the same day by e0077ea36, which imported the pre-consolidation metadata-translations/zh-CN.ts overlay — 746 lines deleted, the exact stale string at that file's :470 in the parent commit — and never reconciled against the seeded English.

⇒ ⭐ A mistranslation at authoring time, not a drift channel. That bounds the class: one import of an older hand vocabulary, never checked against source. Not a recurring leak — a one-time reconciliation.

---

Generated by Claude Code