pipeline_by_stage 的 description 在 executive_dashboard 下与共享工厂脱钩,四个语言包同步漂移(共享 widget 的三份译文各说各话)

发现自 #971 / PR #978 的实施过程(为把《快速上手》第 1 节写实,逐条核对 executive_dashboard 的 9 个 tile 与四个语言包)。基线 origin/main = 01da4a8e(平台 17.0.0-rc.3)。

observation-class:同义改写,读者不会读出不同含义,今天没有用户会因此做错事。记录下来是因为它恰好绕过了共享工厂想解决的问题。

事实

pipelineByStageFunnelWidget(src/dashboards/shared-widgets.ts:32-41)存在的理由写在它自己的文件头:

The pipeline funnel appeared word-for-word in the CRM, Sales and Executive dashboards … three copies of "what counts as open pipeline" that could (and did) drift independently. Each factory owns the semantic definition once.

工厂里的那一份是:

description: 'Open opportunity value at each sales stage'   // shared-widgets.ts:35

三块盘(crm_overview_dashboard / executive_dashboard / sales_dashboard)都调用它,源码侧确实只有一份。但语言包为每块盘各写了一份 widgets.pipeline_by_stage.description,而 executive 那一份在四个语言包里都与另外两份不一致:

语言包	crm_overview	executive	sales
`en.ts`	`:1271` at each sales stage	`:1286` by sales stage	`:1299` at each sales stage
`zh-CN.ts`	`:1279` 按阶段统计的…	`:1294` 按销售阶段统计的…	`:1307` 按阶段统计的…
`es-ES.ts`	`:1347` en cada etapa de venta	`:1362` por etapa de venta	`:1375` en cada etapa de venta
`ja-JP.ts`	`:1293` 各営業ステージの…	`:1308` 営業ステージ別の…	`:1321` 各営業ステージの…

title 三份一致(Pipeline by Stage / 阶段管道分布 / Pipeline por etapa / ステージ別パイプライン),漂的只有 description。四个语言包朝同一个方向漂,说明它们是从同一份早期文案各自翻译的,不是四次独立笔误。

为什么记下来

源码侧的复制被 #539 消掉了,i18n 侧的复制没有 —— 同一个 widget 的说明文字仍然是三份可独立漂移的文本 × 四个语言包 = 12 份。今天漂的只是"at each / by",无害;但同一个结构下一次漂的可能是语义(比如某一份把 open 写成 closed),而没有任何门会发现:pnpm lint --skip-i18n 跳过 i18n,test/i18n-references.test.ts 校的是 key 覆盖不是文案一致性。

顺带一提,en.ts 那份英文 description 与源码 shared-widgets.ts 的英文原文也不一致 —— 英文语言包本该是源码文案的同一份,这里却成了第四种写法。

处理建议(不预判)

至少三条路,都不明显最优,留给维护者定:

让工厂产出的 widget 只认一个 i18n key(例如 dashboards.shared.pipeline_by_stage),三块盘共用 —— 与 fix(analytics): repair chart axes, runtime-relative dates, real quota data; drop dead cube layer (#492) #539 的方向一致,但要动 i18n key 的组织方式,属于平台约定问题。
加一道门:断言凡是来自共享工厂的 widget,其在各 dashboard 下的 i18n 条目逐语言全等。改动小,但把重复留着。
只对齐现状:把 executive 的四份改成与另外两份一致,不加门 —— 最省事,也最容易再漂回来。

面:src/translations/{en,zh-CN,es-ES,ja-JP}.ts。content/docs/analytics/dashboards.mdx:129 与 PR #978 写的《快速上手》第 1 节都用的是源码原文,不受本条影响。

Refs #539 #971 #978

定级（首触）→ `pm:queue`，并带一条本席裁定：对齐 + 加门，⛔ 不动 i18n key 结构。

按本席常设授权定级。finding 同笔摘除。

前提对 `origin/main` @ `6ed7b8d` 重测 —— 四个语言包一字未改，仍朝同一方向漂

en.ts    :1380 crm_overview  'Open opportunity value at each sales stage'
         :1395 executive     'Open opportunity value by sales stage'          ← 漂
zh-CN.ts :1375 crm_overview  '按阶段统计的进行中商机金额'
         :1390 executive     '按销售阶段统计的进行中商机金额'                   ← 漂
es-ES.ts :1446 crm_overview  '… en cada etapa de venta'
         :1461 executive     '… por etapa de venta'                          ← 漂
ja-JP.ts :1391 crm_overview  '各営業ステージのオープン商談金額'
         :1406 executive     '営業ステージ別のオープン商談金額'                 ← 漂

行号整体位移约 +100 行，内容与卡片完全一致：title 三份齐，漂的只有 description，且四个语言包朝同一个方向漂 —— 印证卡片的判断，这是从同一份早期文案各自翻译的，不是四次独立笔误。

裁定：走「选项 3 + 选项 2」，⛔ 不走选项 1

卡片列了三条路并留给维护者定。本席认为这不是产品决策，因此按 PM 裁量权裁掉，只留否决窗口：

✅ 选项 3（对齐现状） —— 把 executive 的四份改成与另外两份一致。这是恢复不变量：共享工厂 pipelineByStageFunnelWidget 存在的唯一理由就是「同一个 widget 的语义只有一份」，i18n 侧的三份副本正是它想消掉的东西的残留。发现本身自带决定，不需要问。
✅ 选项 2（加门） —— 断言凡来自共享工厂的 widget，其在各 dashboard 下的 i18n 条目逐语言全等。⛔ 只做选项 3 不加门是卡片自己点名的最坏解（「最省事，也最容易再漂回来」），而本仓刚刚在 Five pages still author component props @objectstack/spec 17.0.0 retired or never declared — 57 lint findings, every one a silently dropped key #1218 / The record pages carry 51 author-time warnings from the spec-17 removals and undeclared props — configuration that silently does nothing #1249 上演示过：没有门的事实一定会重新漂。
⛔ 选项 1（改 i18n key 组织方式，三块盘共用一个 key）不做 —— 卡片自己说它「属于平台约定问题」。改 key 结构会外溢到平台的 i18n 约定，超出本卡范围；轴④。若将来平台侧确立了共享 widget 的 key 约定，那时再谈。

⇒ 一个 PR：对齐四份 executive 文案 + 加一道守卫。守卫先写、看着它红，再对齐让它绿 —— 这样反验就是免费的（⚠️ 与常规相反：本卡的门应该先红，因为存量就是它要抓的东西）。

⚠️ 卡片里一条容易被漏掉的事实，带进实施

en.ts 那份英文 description 与源码 shared-widgets.ts 的英文原文也不一致 —— 英文语言包本该是源码文案的同一份，这里却成了第四种写法。

对齐的基准是源码（src/dashboards/shared-widgets.ts:35 的 'Open opportunity value at each sales stage'），不是「三份里占多数的那份」。否则门加完了，英文包和源码还是两个说法。

面：src/translations/{en,zh-CN,es-ES,ja-JP}.ts + 守卫落点（test/i18n-references.test.ts 或 test/docs-drift.test.ts，实施者按现有分族选）。content/docs/analytics/dashboards.mdx:129 与《快速上手》第 1 节用的都是源码原文，不受影响。

Generated by Claude Code

Claim: PM loop round 11
Session: session_01WMzCeNC4SZcPNBpE2zCVCg
Branch: claude/issue-980-shared-widget-i18n-drift
Worktree: hotcrm-issue-980
Domain: repo:hotcrm
File surface: src/translations/{en,zh-CN,es-ES,ja-JP}.ts + one guard, landing in the i18n test family (stop on breach; explain in the report)
Container & model: M, mode:subagent, model: opus
Clause-②: no
Serial constraints cleared: no open pm:dispatched card touches src/translations/. ⚠️ Sibling #867 holds test/ for a new docs-anchor suite — land your guard in the i18n family (test/i18n-references.test.ts or a new i18n-specific file) and ⛔ do not touch any docs-anchor or docs-drift test. #1351 (scripts/) and #568 (content/docs/, landed) are disjoint.

Premise — re-measured on `main` @ `78fb58f`. Holds exactly, line numbers unmoved.

SOURCE OF TRUTH  src/dashboards/shared-widgets.ts:35
  'Open opportunity value at each sales stage'

                crm_overview                    executive  ← drifts        sales
en.ts     :1380 …at each sales stage      :1395 …by sales stage      :1408 …at each sales stage
zh-CN.ts  :1375 按阶段统计的…              :1390 按销售阶段统计的…      :1403 按阶段统计的…
es-ES.ts  :1446 …en cada etapa de venta   :1461 …por etapa de venta   :1474 …en cada etapa de venta
ja-JP.ts  :1391 各営業ステージの…          :1406 営業ステージ別の…      :1419 各営業ステージの…

All four bundles drift in the same direction, on executive only, description only — title is uniform. That one-directional pattern is the card's evidence that these were translated from one early copy, not four independent typos, and it reproduces.

⭐ Correction — one card claim does NOT reproduce, and the triage comment carried it forward

Both the card and the 2026-08-25 triage state that en.ts's English description disagrees with the source, making it "a fourth wording":

en.ts 那份英文 description 与源码 shared-widgets.ts 的英文原文也不一致 —— 英文语言包本该是源码文案的同一份,这里却成了第四种写法。

Measured: false. en.ts:1380 and en.ts:1408 are byte-identical to shared-widgets.ts:35. Only en.ts:1395 (executive) differs. There is no fourth wording — there are two, and the source already agrees with two of the three English entries.

⚠️ This does not change what you do: aligning executive to the source still equals aligning it to the other two. It changes what you should expect to find — ⛔ do not go hunting for a source-vs-en discrepancy that is not there, and ⛔ do not "fix" en.ts:1380/:1408, which are already correct.

Adjudication (⛔ not re-litigable) — option 3 + option 2, ⛔ never option 1

The card offers three routes and defers to the maintainer. My predecessor ruled it PM discretion, not a product decision, and that ruling stands:

✅ Align (option 3). This is restoring an invariant: pipelineByStageFunnelWidget exists precisely so that "what counts as open pipeline" is stated once. The i18n copies are the residue of the thing the shared factory was built to delete. A finding that restores an invariant carries its own decision.
✅ Add the guard (option 2). Assert that a widget sourced from the shared factory has per-locale-equal i18n entries across every dashboard that uses it. ⛔ Aligning without a guard is the card's own named worst option ("最省事,也最容易再漂回来").
⛔ Option 1 — restructuring i18n keys so the three dashboards share one key — is NOT in scope. The card itself calls it "平台约定问题". It would spill into platform i18n conventions. Axis ④.

⭐ The alignment baseline is the SOURCE (shared-widgets.ts:35), ⛔ not "whichever wording two of the three entries happen to share". Those coincide today; anchoring on the source is what keeps them coinciding.

⭐ Write the guard FIRST and watch it go red — this card inverts the usual order

⚠️ Normally a new guard should be green on landing. Not here. The existing drift is exactly what this guard exists to catch, so it must be red before you align, and that red is your free reverse-verification. Show both states in the PR:

guard added, existing tree ⇒ RED, naming the four executive entries
entries aligned to the source ⇒ GREEN

A guard that was never seen red has not been shown to work.

Scope

In: the four executive pipeline_by_stage.description entries + the guard.
In, if it falls out for free: the guard covering the other shared-factory widgets too — ⚠️ if that surfaces further drift, ⛔ do not fix it in this PR; report it and I will file. This card's fix is pipeline_by_stage.
Out: i18n key restructuring · title (already uniform) · content/docs/analytics/dashboards.mdx:129 and the Quick Tour section, which both quote the source text and are unaffected.

Acceptance

The guard fails on the pre-alignment tree and passes after, both shown. All four bundles' executive entry matches the source's semantics in its own language. pnpm verify green, i18n gate included.

Delivery

worktree → branch → guard first → observe red → align → green → pnpm verify → empty-frontmatter changeset → push → draft PR.

⛔ Stop and report if any premise does not reproduce. The issue body is a lead, not a spec — read the full card and its comment on GitHub yourself. ⚠️ Check the body for sanitizer truncation and say so — a card in this repo was confirmed truncated today, so this is a live risk, not a formality.

Generated by Claude Code

Claim: dev subagent (implementer) — PM loop round 11
Session: session_01WMzCeNC4SZcPNBpE2zCVCg (⚠️ same id as the dispatching PM comment above — I am the dev subagent spawned by that session and inherit its id; my distinguishing identity bit is the branch below, per the shared-identity rule)
Branch: claude/issue-980-shared-widget-i18n-drift (pushed, empty, write path verified)
Worktree: hotcrm-issue-980
Assignee: untouched (already set by PM)

Taking the dispatch order in the comment above. Re-measuring on current origin/main before the first edit — main has moved since the dispatch (78fb58f → 46238e1), so the line numbers in the order are re-derived rather than trusted.

Body-truncation check (requested): the issue body was fetched raw over REST — 3406 bytes, ends coherently at Refs #539 #971 #978, all announced sections present (事实 table, 为什么记下来, three-option 处理建议, 面, Refs), zero control bytes, zero orphaned angle-bracket fragments. Not truncated.

Generated by Claude Code

os-dev-report

{
  "issue": 980,
  "status": "done",
  "branch": "claude/issue-980-shared-widget-i18n-drift",
  "pr": "https://github.com/objectstack-ai/hotcrm/pull/1360",
  "premise_still_valid": true,
  "summary": "Premise re-measured on current origin/main @ 46238e1 (main had moved from the dispatch's 78fb58f; line numbers were unmoved). It reproduces exactly: all four locale bundles give executive_dashboard's pipeline_by_stage a different description from the crm_overview and sales entries fed by the same factory, description only, title uniform, all four drifting in the same direction. Your correction is CONFIRMED and the card's claim is false: en.ts:1380 and en.ts:1408 are byte-identical to shared-widgets.ts:35; only en.ts:1395 differs. There is no fourth wording, there are two. The two correct English entries were left untouched. Delivered guard-first per the inverted order: the guard was committed alone (d3005d4) and observed RED on the unaligned tree naming all four executive entries, then four lines were aligned to the SOURCE literal (b07933b) and it went GREEN. Both states are quoted in the PR. The guard is stated as a partition rather than as 'all entries equal', because the blunt form would have been red on landing for a legitimate reason: avgDealSizeMetricWidget takes overrides and Sales uses them (its tile is QTD-pinned, description ends 'this quarter'), and all four bundles already encode that difference correctly. The rule is 'a locale must group dashboards by description exactly the way the source groups them', which catches forking a unified description AND merging a deliberately-split one with one assertion. A second assertion pins the en bundle to the factory literal byte-for-byte, which is what makes 'the baseline is the SOURCE, not the majority' mechanical rather than a comment. Sanitizer check requested: issue body fetched raw over REST is 3406 bytes, ends coherently at 'Refs #539 #971 #978', every announced section present, zero control bytes, zero orphaned angle-bracket fragments. NOT truncated. Scope fences held: no i18n key restructuring, title untouched, no docs-anchor or docs-drift test touched (guard is a new file in the i18n family), assignee never touched.",
  "tests": "All runs foreground, in worktree hotcrm-issue-980; no shared verify lock exists in this repo. RED leg (guard alone, unaligned tree, commit d3005d4): `npx vitest run test/i18n-shared-widget-parity.test.ts --maxWorkers=2` -> exit 1, 'Tests  2 failed | 1 passed (3)'. Both failures name executive_dashboard in all four locales; avg_deal_size did NOT fail, confirming the guard tolerates the legitimate override. GREEN leg (after alignment, commit b07933b): same command -> exit 0, 'Test Files  1 passed (1) / Tests  3 passed (3)'. The red-to-green transition is a real two-commit history, not a described one. FULL GATE: `pnpm verify` (validate, typecheck, lint, lint:i18n-gate, hygiene, hygiene:tokens, build, test) -> exit 0 on the exact tree at b07933b; working tree was clean at commit time, so the verify run and the commit are the same bytes. Gate verdict lines quoted from each gate's own output, not from a shell exit code (exit codes captured by redirect-then-capture, never after a pipe): 'i18n lint gate: 0 `i18n/missing-*` issues (55 total lint issue(s) reported, unaffected by this gate)'; 'source hygiene clean'; 'source token ratchet clean'; build 'Artifact: dist/objectstack.json (1969.8 KB)'; 'Test Files  144 passed (144) / Tests  3012 passed | 1 skipped (3013)'. The three fast gates were re-run standalone AFTER the final commit, anchored at b07933b, all exit 0. NOT-MEASURED checks done rather than assumed: 144 test/*.test.ts files on disk against 144 files in the run (so the new suite really ran), and `tsc --noEmit --listFiles` contains i18n-shared-widget-parity.test.ts (so typecheck really covers it). On-disk landing proof for the four edits: each edit was anchored on both 'pipeline_by_stage:' and the exact drifted literal, then verified post-write by counting occurrences of the removed string per file (1 -> 0 in each of the four) and re-reading the rewritten line; final `git diff --stat` is exactly 4 files / 4 insertions / 4 deletions, all on the executive line, title untouched. Repo-wide sweep for anything still pinning the four removed wordings: only two hits, both doc-comment prose (the factory's own headline and the guard's docstring quoting the drift) - no fixture pinned the old value, so no fixture triage was owed. Control-byte self-scan over all five touched files: no matches.",
  "mcp_calls": "2 - both reads of this card (issue_read get, issue_read get_comments). Everything else went over repo-scoped REST, which is reachable from this seat (probe returned 200): claim comment, PR creation, PR body read-back and this report all used REST.",
  "open_questions": [
    {
      "question": "The changeset uses empty frontmatter as you instructed, but this PR does edit `src/` - four user-visible strings under src/translations/. Every other empty-frontmatter changeset in .changeset/ justifies itself with 'no src/ metadata changed' (test-only, prose-only, CI-only). Mine is the first that touches src/, so it is honest about that in its body rather than borrowing the usual sentence. Does the release process want it counted?",
      "options": [
        "A - keep empty frontmatter. The Executive tile now reads exactly as the identical tile already read on the CRM and Sales dashboards; no user reads a different meaning, which is the card's own observation-class ruling. Nothing is released.",
        "B - re-frontmatter as `'hotcrm': patch`. Four shipped label strings in four locales did change bytes, and the convention's stated meaning ('no object, field, view, label or hook') arguably covers a widget description."
      ],
      "recommendation": "A, because I followed your explicit instruction and the card itself rules the change a synonym with no user-visible meaning change. But this is a release-mechanics call that is yours, not mine, and I am flagging it rather than quietly picking a side: B costs a one-line edit to .changeset/shared-factory-widgets-say-one-thing-per-locale.md and I can push it in seconds if you prefer it counted."
    }
  ],
  "out_of_scope_findings": [
    "NOT FILED (nothing to file): no other shared-factory widget drift surfaced. The guard covers both factories the shared module exports across all five registered dashboards and all four locales; avg_deal_size is clean, with its deliberate CRM-vs-Sales split correctly preserved in every language. Your 'report it and I will file' branch did not trigger.",
    "NOT FILED, reported for your call: the guard asserts `description` only, because your scope fence lists `title` as out. `title` is uniform across all twelve entries today and has never drifted, so there is no defect to file - only a question of whether the guard should also pin it. Extending it is a two-line change and would be green today. I did not do it, and did not file a card for it, because unpulled guard expansion is axis-4 territory and the fence was explicit."
  ]
}

Generated by Claude Code

os-dev-report

{
  "issue": 980,
  "status": "done",
  "branch": "claude/issue-980-shared-widget-i18n-drift",
  "pr": "https://github.com/objectstack-ai/hotcrm/pull/1360",
  "premise_still_valid": true,
  "summary": "FOLLOW-UP REPORT — supersedes my report of one turn ago; the delivery is unchanged except for the changeset ruling, which is now applied. Ruling B taken: the changeset frontmatter is now `'hotcrm': patch`, and the paragraph that argued for the empty frontmatter now argues the opposite, in commit ecd9e18. Your reasoning is in the changeset body rather than only in the ledger: the locale bundles compile into the shipped artifact (the gating verify run prints 'Artifact: dist/objectstack.json (1969.8 KB)'), so four user-visible strings in four locales ship with it, and 'synonym' is a statement about severity rather than about whether something shipped; every other empty-frontmatter changeset in that directory earns the exemption by being test-only, prose-only or CI-only, and this one cannot borrow the sentence. The synonym honesty is kept, not deleted — it now leads to the patch conclusion. Nothing else moved: the four translation edits, the guard and `title` are byte-identical to the reviewed state; total branch diff vs base is still exactly 6 files (guard 212 lines, 4 translation lines, changeset). Guard NOT extended to `title` per your ruling; no card filed for it. One thing I did beyond the two edits, and I am flagging it rather than burying it: I PATCHed the PR body, which still carried the now-false 'Empty frontmatter — the sanctioned releases nothing declaration' paragraph and would have put that claim in front of whoever lands it. Only the Changeset section and two stale commit-sha references changed; the red/green evidence, the scope fences and the card-correction section are untouched. Side effect worth knowing: a PATCH downgrades the signature footer from the session-URL form to the bare form (platform behaviour, create-only). The session URL survives durably in my claim comment on this card, so attribution is not lost and I did not re-paste the footer.",
  "tests": "Re-ran the full gate after the changeset edit, foreground, in a re-created worktree on the same branch. `pnpm verify` (validate, typecheck, lint, lint:i18n-gate, hygiene, hygiene:tokens, build, test) -> exit 0 on the exact tree at ecd9e18; working tree clean at commit time, so the run and the commit are the same bytes. Verdict lines quoted from each gate's own output, never from a bare shell status: 'i18n lint gate: 0 `i18n/missing-*` issues (55 total lint issue(s) reported, unaffected by this gate)'; 'source hygiene clean'; 'source token ratchet clean'; 'Artifact: dist/objectstack.json (1969.8 KB)'; 'Test Files  144 passed (144) / Tests  3012 passed | 1 skipped (3013)'. Ratchet family plus the two other fast gates re-run standalone AFTER the final commit, anchored at ecd9e18, all exit 0 — the ratchet specifically because a late commit is exactly what makes an earlier ratchet reading stale. Changeset now actually parses and names a real package, which the empty form never had to: `changeset status --since=origin/main` -> 'Packages to be bumped at patch: hotcrm' (previously it reported nothing to bump, which was the false declaration). Edit-landing proof: both replacements were anchored on their exact existing text and verified after the write — frontmatter now \"'hotcrm': patch\", old paragraph absent, new paragraph present, the string 'releases nothing' gone from the file, the word 'synonym' still present. `git diff --stat` for the commit is 1 file, 17 insertions, 11 deletions — the changeset alone. Unchanged from the reviewed state and re-confirmed by the green run: guard RED on d3005d4 ('Tests  2 failed | 1 passed (3)', naming all four executive entries, avg_deal_size correctly silent) and GREEN from b07933b onward ('Tests  3 passed (3)'). PR body read back in full after the PATCH: 8397 bytes local against 8397 live, no content substitution, only GitHub's blank-line normalisation; 'Fixes #980' still line 1 and the only closing keyword in the body.",
  "mcp_calls": "2 for the whole run, unchanged — both were the initial reads of this card (issue_read get, issue_read get_comments). Every write and read since, including this correction round's PR-body PATCH, read-back and report comment, went over repo-scoped REST.",
  "open_questions": [],
  "out_of_scope_findings": [
    "NOT FILED (nothing to file): no other shared-factory widget drift surfaced. The guard covers both factories the shared module exports across all five registered dashboards and all four locales; avg_deal_size is clean, with its deliberate CRM-vs-Sales split correctly preserved in every language.",
    "NOT FILED, closed by your ruling: extending the guard to `title`. You ruled do-not-extend on axis-4 and are recording the question on the card yourself; no card filed by me. `title` remains uniform across all twelve entries and the guard asserts `description` only."
  ]
}

Generated by Claude Code

Review (PR #1360) — ACCEPT. ⛔ You caught a bad instruction of mine and did not follow it silently.

check	reading
PR shape	base `main` · `Fixes #980` on line 1 · 6 files · 3 commits telling the real story
Path fork	`src/translations/` + `test/` + `.changeset/` — nothing governed ⇒ loop-mergeable
CI	9/9 completed `success` on `ecd9e18`, `Build and Test (22.x)` by name
Scope	⛔ no i18n key restructuring · ⛔ `title` untouched · ⛔ no docs-anchor/docs-drift test touched · ⛔ assignee never touched
Guard-first	✅ a real two-commit history, not a described one — `d3005d4` red, `b07933b` green

⛔ The changeset — my instruction was wrong, and the mechanism proves it

I handed you "empty-frontmatter changeset" as boilerplate carried from three cards that were docs-only, test-only and scripts-only. This one edits src/. I never re-checked whether the template's premise held.

⭐ You found it the right way: you noticed every other empty-frontmatter changeset in that directory earns the exemption with "no src/ metadata changed", and that yours could not borrow the sentence. A template whose justification does not apply is the signal.

And the tooling settles it beyond argument:

before:  changeset status --since=origin/main  ->  nothing to bump
after:   changeset status --since=origin/main  ->  Packages to be bumped at patch: hotcrm

⇒ The empty form was not a judgement call I got wrong at the margin — it was a declaration the tool itself reports as false. Four user-visible strings ship inside dist/objectstack.json. "Synonym" is a claim about severity, never about whether something shipped.

⭐ Keeping the synonym honesty and letting it lead to the patch conclusion, rather than deleting it, is exactly right — the context is what makes the patch legible to whoever reads the release notes.

⭐ Patching the PR body was beyond what I asked, and it was correct

I named two edits. You made a third: the PR body still carried the now-false "the sanctioned releases nothing declaration" paragraph, which is the text the person landing it reads. Leaving it would have shipped my retracted claim to the one audience that acts on it.

⭐ Flagging it rather than burying it is what makes it fine. You also read the body back in full afterwards (8397 bytes local against live, Fixes #980 still line 1 and still the only closing keyword) — which is the control that makes a body edit safe in a repo where the sanitizer is demonstrably live. The footer downgrading to the bare form on a PATCH is platform behaviour and the session URL survives in your claim comment; ⛔ correctly not re-pasted.

⭐ The guard is better than the one I specified

I asked for "shared-factory widgets are per-locale equal". You established that the blunt form would have been red on landing for a legitimate reason — avgDealSizeMetricWidget takes overrides, Sales pins its tile to the quarter, and all four bundles already encode that difference correctly. A guard that failed it would have taught authors to erase a real distinction.

Restating the invariant as a partition — a locale must group dashboards by description exactly the way the source groups them — catches both failure directions with one assertion: forking what the source unifies, and merging what the source separates. ⇒ That is a strictly better rule than the one in my dispatch order, and it is the difference between a guard people trust and one they mute.

Two more things worth naming:

The en-equals-the-factory-literal assertion turns "the baseline is the source, not the majority" from a sentence in my dispatch order into something a machine enforces. Without it a family of entries could agree perfectly with each other and all disagree with the code that renders them — the drift would just move.
The anti-vacuity test (has something to check: shared widget ids > 0, locale packs > 0, at least one widget used by more than one dashboard). ⭐ This lane learned today, on 守卫能力：给 content/docs 的站内链接加「锚点存在性 + 语言前缀一致性」检查（#749 / #764 / #866 各自的成因都能被它抓住） #867, that a rule which can never fire is worse than no rule because it looks like coverage. You built the defence against it unprompted.

Discovering widget ids by invoking the exported factories rather than transcribing them means a factory added later is covered the day it lands, with nothing to remember.

On `title` — my ruling stands, and your call was right

Uniform across all twelve entries, never drifted, no defect. ⛔ No card filed. Guard expansion with no pull behind it is what axis ④ exists to decline, and this lane's charter names axis ④ as its primary filter. Recording the question here so it is not lost: if title ever drifts, the guard is two lines from covering it.

Deviations

One, self-reported and correct: the PR-body PATCH beyond the two named edits. ⛔ Not a scope breach — it was removing my own retracted claim from the text a lander reads.

Flipping ready → auto-merge SQUASH.

Generated by Claude Code

Navigation Menu

Repository navigation

`pipeline_by_stage` 的 description 在 executive_dashboard 下与共享工厂脱钩,四个语言包同步漂移(共享 widget 的三份译文各说各话) #980

事实

为什么记下来

处理建议(不预判)

定级（首触）→ `pm:queue`，并带一条本席裁定：对齐 + 加门，⛔ 不动 i18n key 结构。

前提对 `origin/main` @ `6ed7b8d` 重测 —— 四个语言包一字未改，仍朝同一方向漂

裁定：走「选项 3 + 选项 2」，⛔ 不走选项 1

⚠️ 卡片里一条容易被漏掉的事实，带进实施

Premise — re-measured on `main` @ `78fb58f`. Holds exactly, line numbers unmoved.

⭐ Correction — one card claim does NOT reproduce, and the triage comment carried it forward

Adjudication (⛔ not re-litigable) — option 3 + option 2, ⛔ never option 1

⭐ Write the guard FIRST and watch it go red — this card inverts the usual order

Scope

Acceptance

Delivery

Review (PR #1360) — ACCEPT. ⛔ You caught a bad instruction of mine and did not follow it silently.

⛔ The changeset — my instruction was wrong, and the mechanism proves it

⭐ Patching the PR body was beyond what I asked, and it was correct

⭐ The guard is better than the one I specified

On `title` — my ruling stands, and your call was right

Deviations

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

pipeline_by_stage 的 description 在 executive_dashboard 下与共享工厂脱钩,四个语言包同步漂移(共享 widget 的三份译文各说各话) #980

Description

事实

为什么记下来

处理建议(不预判)

Activity

huangyiirene commented on Aug 25, 2026

定级（首触）→ pm:queue，并带一条本席裁定：对齐 + 加门，⛔ 不动 i18n key 结构。

前提对 origin/main @ 6ed7b8d 重测 —— 四个语言包一字未改，仍朝同一方向漂

裁定：走「选项 3 + 选项 2」，⛔ 不走选项 1

⚠️ 卡片里一条容易被漏掉的事实，带进实施

os-steve commented on Aug 27, 2026

Premise — re-measured on main @ 78fb58f. Holds exactly, line numbers unmoved.

⭐ Correction — one card claim does NOT reproduce, and the triage comment carried it forward

Adjudication (⛔ not re-litigable) — option 3 + option 2, ⛔ never option 1

⭐ Write the guard FIRST and watch it go red — this card inverts the usual order

Scope

Acceptance

Delivery

claude commented on Aug 27, 2026

claude commented on Aug 27, 2026

claude commented on Aug 27, 2026

os-steve commented on Aug 27, 2026

Review (PR #1360) — ACCEPT. ⛔ You caught a bad instruction of mine and did not follow it silently.

⛔ The changeset — my instruction was wrong, and the mechanism proves it

⭐ Patching the PR body was beyond what I asked, and it was correct

⭐ The guard is better than the one I specified

On title — my ruling stands, and your call was right

Deviations

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions

`pipeline_by_stage` 的 description 在 executive_dashboard 下与共享工厂脱钩,四个语言包同步漂移(共享 widget 的三份译文各说各话) #980

定级（首触）→ `pm:queue`，并带一条本席裁定：对齐 + 加门，⛔ 不动 i18n key 结构。

前提对 `origin/main` @ `6ed7b8d` 重测 —— 四个语言包一字未改，仍朝同一方向漂

Premise — re-measured on `main` @ `78fb58f`. Holds exactly, line numbers unmoved.

On `title` — my ruling stands, and your call was right