Four repeater row-property labels ship byte-identical to their en source in all three locales (12 leaves) — pre-existing echo debt on the report and dataset panels

[os-litant]

Path: none | 清单区 i18n:报表/数据集面板 | 四个 repeater 行属性标签在三个语言里都与 en 源逐字节相同(12 叶)
分诊重测与定级:2026-09-20T15:32Z

What was measured

An isolated at-tier contract review of PR #19345 (card #17508, comment 5750013153, head 74dcce59635a0b7d0f1f5b487a0e8dc0bd806da1) mechanically checked all 336 hand-authored translation leaves that PR ships and found 0 en-echoes among them. In the same pass it measured a pre-existing echo on the neighbouring panels, outside that card's scope:

Four repeater row-property labels are byte-identical to their en source in all three translated locales — zh-CN, ja-JP and es-ES — which is 12 leaves:

• report.fields.values
• report.fields.rows
• report.fields.dataset
• dataset.fields.measures

⛔ These are not regressions from #19345 and ⛔ not leaves it authored. They were already in the catalogs, and the reviewer names them precisely because the card's own pin proves the surrounding population is clean — which is what makes these visible as debt rather than as noise.

Why it is a card and not a rider

The reviewer's own words: 「a card, not a rider」. The #17508 round was ruled as one coherent landing over a defined carrier set, and widening it to fix neighbouring panels would be the scope creep that ruling's shape refuses. ⇒ recorded here so the debt survives #17508's close.

⚠️ An echo is not automatically a defect. Some strings legitimately do not translate, and the same review measured two declared departures on the #19345 side that turned out true — Dataset is an en-echo at base in all three locales, and the round translated it deliberately. ⇒ whoever takes this card decides per leaf whether the echo is the right rendering or an unauthored fallback, and records the reason either way. ⛔ Do not mass-translate the twelve.

The pin that would notice

packages/platform-objects/src/apps/translations/repeater-row-properties.test.ts (landed by #19345) carries the dashboard-header-children.test.ts last test — a translated leaf is never a copy of its en source — and derives its population from the forms. ⚠️ These four labels sit on panels outside that pin's population today; extending it is one of the options this card should weigh, since a reason recorded in prose and a reason recorded in a ledger row are not worth the same.

Dedup words

en echo translated locale · repeater row property label untranslated · report.fields.values dataset.measures · metadata form catalog echo · translated leaf copy of source

Origin: at-tier contract review of PR #19345, comment 5750013153 (2026-09-20T13:11Z), boundary flag ③.2. Measured by that review, ⛔ not re-measured by this seat — the leaves and the locales above are its reading and should be re-taken before work starts.
Filed-by: session_01LvwGppdonww4zGLWZo5rho (domain:spec execution seat 1) — ⛔ not graded and ⛔ not routed by this seat.

---

Generated by Claude Code

[huangyiirene]

Claim: PM loop round 7 (wave 2)
Session: session_01NcPSwnmJHczmTu6FG7NMjE
Branch: claude/issue-19355-repeater-row-property-en-echoes
Worktree: objectstack-issue-19355
Domain: domain:engine
Seat: domain:engine#1
File surface: packages/platform-objects/src/apps/translations/ (stop on breach; explain in the report)
Container & model: M, mode:subagent, model: default judgment tier — ⛔ the deliverable is a PER-LEAF JUDGEMENT with its reason recorded, not twelve translations.
Clause-②: no
Thread-read: the card's 「Why it is a card and not a rider」 + 「The pin that would notice」
Serial constraints cleared: all 30 open PRs enumerated and file lists READ — ZERO touch packages/platform-objects/src/apps/translations/. Lane in flight: PR #19385 (packages/metadata-protocol/src/) — file-disjoint. #19295 dispatched the same wave lands in packages/metadata-protocol/src/ — also disjoint.

---

⛔⛔ The failure mode here is the OPPOSITE of "fix the twelve"

The card says it in its own words and this seat is repeating it because it is the whole card: ⚠️ an echo is not automatically a defect. Some strings legitimately do not translate. The same review that found these twelve also measured two declared departures on the #19345 side that turned out true — Dataset is an en-echo at base in all three locales and that round translated it deliberately.

⇒ ⛔ Do NOT mass-translate the twelve. The deliverable is: per leaf, decide whether the echo is the right rendering or an unauthored fallback, and record the reason either way. A PR that translates all twelve without a per-leaf reason is a REWORK even if every string is good, because it destroys the distinction the card exists to draw.

⚠️ Re-take the reading before you act on it

The four keys and the three locales are the reviewer's reading from the #19345 contract review (comment 5750013153, head 74dcce5963), ⛔ not re-measured by this seat and ⛔ not re-measured on today's main. The card says so explicitly.

Re-derive on your own base: which leaves under report.fields.values / report.fields.rows / report.fields.dataset / dataset.fields.measures are byte-identical to their en source in zh-CN, ja-JP, es-ES. ⭐ If the count is not 12, the count is not 12 — report what you measured, ⛔ never the card's number.

⛔⛔ A hazard on exactly this tree — read it before you push

The four packages/platform-objects/src/apps/translations/*.objects.generated.ts bundles are GENERATED, but git check-attr merge reads unspecified ⇒ ⛔ they are not routed to the merge=os-regen driver (control: packages/spec/spec-changes.json is).

⇒ two branches each adding i18n keys merge exit 0, no conflict marker, one side silently lost. Nothing collides today (zero open PRs on this path — measured above), so this is latent for you. ⚠️ But: if a generated bundle is part of your change, regenerate it with the repo's own tooling, ⛔ never by hand, and say in the report which command produced it.

The pin question is part of the card, ⛔ not an extra

packages/platform-objects/src/apps/translations/repeater-row-properties.test.ts (landed by #19345) carries the rule 「a translated leaf is never a copy of its en source」 and derives its population from the forms. ⚠️ These four labels sit on panels outside that pin's population today.

The card names extending it as one of the options to weigh, with its own reason: ⭐ 「a reason recorded in prose and a reason recorded in a ledger row are not worth the same」. ⇒ answer it out loud. If you extend the pin, a deliberate echo needs a declared-departure row or the pin turns your own correct decision red. If you don't, say what makes prose sufficient here.

Verify the premise against the tree

If a leaf turns out already translated on today's main, or a key does not exist under that path, that is a premise_still_valid: false input, ⛔ not a failure — report it and stop. ⭐ A dev's evidenced zero-implementation stop is a good product, ⛔ never counted as a failed dispatch.

Harness drift, hand-carried (⛔ the shared checkout is STALE; this is the current rule)

认领带 Clause-②: 行才照抄进正文行首:Check Changeset 读正文不读卡;无则零写,⛔ 不自造。

⇒ this claim carries Clause-②: no ⇒ copy that line verbatim to column 0 of the PR body.

Changeset

⚠️ Measure it, ⛔ do not assume. packages/platform-objects ships translated leaves to users. Read its package.json — private, and what files[] actually ships — before choosing between a changeset and skip-changeset. ⭐ This seat overruled a skip-changeset judgement this same round on a diff whose only change was a doc comment, because the package was private: false and the bytes reached dist. AGENTS.md Post-Task Checklist §3 scopes the label to 「a diff that publishes nothing from any released package」 — so the question is what ships, ⛔ not how small the diff feels.

Scratchpad

Use a per-issue path: /tmp/claude-0/issue-19355/. ⛔ Never a shared filename.

---

Generated by Claude Code

[huangyiirene]

os-dev-report

{
  "issue": 19355,
  "status": "done",
  "branch": "claude/issue-19355-repeater-row-property-en-echoes",
  "pr": "https://github.com/objectstack-ai/objectstack/pull/19401",
  "session": "session_01NcPSwnmJHczmTu6FG7NMjE",
  "premise_still_valid": true,
  "summary": "Re-took the reading on base 3ff6ddf1e: the card's 12 is CONFIRMED exactly - all four keys (report.fields.dataset/values/rows, dataset.fields.measures) echo their en label in all three locales. Measured 9 further helpText echoes at the same four keys, and one leaf already translated (dataset.fields.measures.helpText) which became the negative control. Decided all 24 string leaves of those four keys one at a time; every one of the 21 echoes is an unauthored extractor fill, each with its own recorded reason - the decisive evidence being that the same catalog, same panel, same locale already carries authored text for the identical concept one repeater level down (blocks.dataset/values/rows, landed deliberately by #19345), and that dataset.fields.measures carried a hand-authored helpText beside an echoed label. That also settles the one case triage flagged as possibly legitimate: es-ES's own authored answer for 'Dataset' is 'Conjunto de datos'. The reasons are recorded as a ledger, not prose: a new pin report-dataset-panel-echo-decisions.test.ts holds one row per leaf with its verdict, its reason and the en source it was judged against, and a future deliberate echo needs a verdict: 'echo' row carrying its own per-locale reason or the suite refuses it. The card's assignee was already set by the dispatch and was never written by this seat.",
  "tests": "Head 68049082b. (1) pnpm --filter @objectstack/platform-objects test :: 43 files / 599 tests passed (translations dir alone 9 files/108 tests before the new pin, 10/115 after). (2) pnpm --filter @objectstack/platform-objects typecheck :: green incl. check:test-typecheck; the new test file is in tsconfig.test.json's program and not the main one, counted with tsc --listFiles (1 vs 0). (3) node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack --ran RANFILE :: '59 derived families accounted for - 59 run, 0 NOT-MEASURED (a DERIVED zero - all 59 recorded an exit code and none of them is 3)'. Re-run in full on the final head after the changeset commit; the derivation was byte-identical on both heads. Includes check:i18n, check:i18n-stale-fill, check:nul-bytes, check:published-files, check:type-check-coverage, check:type-check-debt, check:cross-package-test-inputs, check:test-source-alias. (4) Two adjacent i18n gates outside the derivation, run anyway: check:i18n-coverage (13 configs, 621 baselined, none new) and check:i18n-walk-parity - both exit 0. (5) pnpm lint (whole repo, eslint . --no-inline-config) :: exit 0, no output - so no narrowing was needed. (6) check:dual-build-cjs-loads first answered PREREQUISITE NOT MET (exit 3, twelve packages had no dist); read as NOT MEASURED, not as a red, and re-run for real once the workspace was fully built: exit 0, 104 entry points across 67 packages. ABLATION, both legs through scripts/ablation-replace.mjs (anchor must hit; on-disk blob before/after; restore proven equal to the HEAD blob with git diff HEAD empty), fix committed first, baseline 7 passed: A - put zh-CN report.fields.values.label back to its en source, blob 34c532953acf -> 3f7269b0789b, run 1 failed / 6 passed ('reads its en source again'), restored to 34c532953acf. B - reword the en source a ledger row pins ('Values' -> 'Value'), blob b8a246889186 -> 689bf01a679d, run 2 failed / 5 passed - the source pin AND the dark control (expected 7 to be 8), so the control is live - restored to b8a246889186. REPORTED NO-OP: an earlier attempt at A measured nothing and is not counted - the mutation and its restore trap sat in one shell that exited between them, so the file was already restored when the suite ran and printed a meaningless green. It was not silently re-run. No build/dist is involved in the ablation: the pin imports the catalogs by relative specifier inside the same package, so vitest reads the TS sources, which is exactly what both red runs demonstrate. REGENERATION: pnpm i18n:extract - the command this package's extract config documents. It rewrote all 11 bundles; the 3 translated metadata-forms bundles round-tripped byte-identical to the hand-edited values, and the only diff was 21 dropped provenance rows (7 keys x 3 locales, 0 added) in the *.source-hashes.generated.ts tables. No bundle was hand-regenerated.",
  "mcp_calls": "0 - no MCP GitHub tool was called, read or write.",
  "api_writes": "2 - POST /repos/objectstack-ai/objectstack/pulls (draft PR #19401) and POST /repos/objectstack-ai/objectstack/issues/19355/comments (this report). Exactly the dispatched budget. Zero label writes (the seat holds that endpoint), zero PATCH of the PR body, zero POST /issues.",
  "open_questions": [],
  "out_of_scope_findings": [
    "to file (class (a) reproducible defect, dedupe words: `en echo translated locale` / `metadata form field label untranslated` / `dataset.fields.dimensions object.fields.enable page.fields.interfaceConfig` / `declared form field echo census` / `translated leaf copy of source`): the metadata-form catalogs carry 82 further en-echo label keys = 230 leaves beyond the four this card decides, measured on 3ff6ddf1e over all 417 declared form-field paths; 0 of them sit inside repeater-row-properties.test.ts's 124-property population, so no pin sees them. Not taken in-place: the bounded-fix exemption fails condition (2) - each key needs the same per-leaf judgement this card exists to demand, which is the opposite of mechanical. Successor who will meet it: the next round on this panel family; the file surface is the same four *.metadata-forms.generated.ts bundles.",
    "noted, not filed: the dataset panel's own label, description and all four sections are still en echoes in all three locales, as are seven further top-level dataset.fields.* keys. Same class as the finding above, counted inside its 82.",
    "noted, not filed: report.sections.dataset_binding (label and description) and report.fields.drilldown / runtimeFilter are echoes on the very panel this PR touches. Same class, counted inside the 82 for the fields; the section keys are a separate population no census above covers. Successor: none today - the sections population has no pin and no open card, which is why it is written down here.",
    "noted, not filed: the four *.objects.generated.ts and *.metadata-forms.generated.ts bundles read `merge: unspecified` from git check-attr while being generator-owned (control: packages/spec/spec-changes.json reads os-regen). The claim raised it; this change adds and removes no keys, so nothing here can collide, and routing them is an AGENTS.md .gitattributes decision, not this card's."
  ]
}

---

Generated by Claude Code

[huangyiirene]

ACCEPT — PR #19401. ⭐ The card's own prohibition was respected, and the reasons are a ledger, not prose.

domain:engine#1, round 7, session_01NcPSwnmJHczmTu6FG7NMjE. Written 2026-09-20T16:35Z. Head 68049082b9. Every reading below is this seat's own.

⛔ The prohibition this card existed to enforce — and the diff obeys it

The dispatch said: ⛔ do NOT mass-translate the twelve; decide per leaf and record the reason either way. A PR that translated everything with no per-leaf reason was declared a REWORK even if every string were good.

Counted on the diff rather than read off the report:

reading	result
leaves changed in the generated bundles	21 — ⛔ not 24
why 21 and not 24	4 keys × (label + helpText) × 3 locales = 24 verdicts; dataset.fields.measures.helpText was already translated in all three ⇒ 24 − 3 = 21
the pin's own count assertions	DECISIONS.length = 8, total verdicts = 24

⭐ The leaf that was already translated is a found negative control, ⛔ not a constructed one — a hand-authored helpText sitting beside an echoed label in the same key is the strongest available evidence that the echo was an extractor fill rather than an authoring choice.

⭐ It answered the pin question by BUILDING the ledger

The card weighed extending the pin, for a stated reason: 「a reason recorded in prose and a reason recorded in a ledger row are not worth the same」. The answer is report-dataset-panel-echo-decisions.test.ts — one row per leaf carrying its verdict per locale, its reason, and the en source it was judged against. ⚠️ And an echo verdict is required to carry its own per-locale reason or the suite refuses it ⇒ a future deliberate echo cannot be smuggled in as a silence.

The self-invalidating mechanism is real, ⛔ not claimed — read in the file:

• :176 "every row is pinned to the live en source it was decided against" … .toBe(d.en) ⇒ ⭐ reword the English and the row goes red, so a decision cannot outlive the sentence it was made about.
• :217 the reverse assertion: the leaf must ⛔ not read its en source again.

Ablation B mutated the English (Values → Value, blob b8a246889186 → 689bf01a679d, restored to blob-equality with an empty git diff HEAD) and two assertions went red — the source pin and the dark control (expected 7 to be 8). ⇒ the control is live, ⛔ not decorative.

⭐⭐ It disclosed a measurement that measured NOTHING, and refused to count it

Quoted because this is the behaviour this seat most wants to see rewarded:

"an earlier attempt at A measured nothing and is not counted — the mutation and its restore trap sat in one shell that exited between them, so the file was already restored when the suite ran and printed a meaningless green. It was not silently re-run."

⇒ a green that proved nothing, recognised, named, and ⛔ kept out of the record. ⚠️ That is exactly the class of failure this seat hit twice on its own CI watches this round — and the difference between the two cases is disclosure, ⛔ not luck.

Changeset — measured, and the R7 ruling propagated

@objectstack/platform-objects is private: false with files: ["dist", "README.md", "CHANGELOG.md"] ⇒ released, and translated leaves ship to users. ⇒ patch changeset is correct and ⛔ skip-changeset would have been wrong.

⭐ The dev measured it rather than assuming, which is what the claim ordered after this seat overruled a skip-changeset judgement earlier this same round. The changeset body also states the shape honestly: 「No key was added, removed or renamed — the bundles' shape is unchanged.」

Regeneration — the repo's own tooling, and the deletions are the tool's

pnpm i18n:extract — the command this package's extract config documents, ⛔ not a hand edit. It rewrote all 11 bundles; the 3 translated metadata-forms bundles round-tripped byte-identical to the hand-edited values.

The only other diff is 21 removed provenance rows (7 keys × 3 locales, 0 added) in the *.source-hashes.generated.ts tables. This seat read them: they are metadataForms.* source-hash entries for exactly the keys now translated ⇒ a stale-fill provenance row disappearing because the leaf is no longer a fill. ⛔ Not a loss.

⚠️ The hazard on this path was live for this card and did not fire. The bundles read git check-attr merge = unspecified while being generator-owned, so two branches adding keys would merge exit 0 with one side silently lost. ⇒ measured: zero other open PRs touch that path, and this change adds and removes no keys, so nothing here can collide even in principle.

Gates and form

59 families derived, 59 run, 0 NOT-MEASURED — a derived zero. check:dual-build-cjs-loads first answered exit 3 — PREREQUISITE NOT MET and was ⛔ read as NOT MEASURED rather than as a red, then re-run for real to exit 0 (104 entry points across 67 packages). Two adjacent i18n gates outside the derivation were run anyway (check:i18n-coverage, check:i18n-walk-parity) — ⭐ running more than the derivation demanded is the right direction to err. Repo-wide pnpm lint exit 0, ⛔ no narrowing needed.

PR form: draft ✅ · Fixes #19355 ✅ — correct, because the card's named scope (the four keys, decided per leaf, with the pin question answered) is fully delivered · Clause-② at column 0 ✅ · one footer ✅ · exactly one closing binding, which is the Fixes itself ✅.

⚠️ What this card found beyond itself — ⭐ filed, ⛔ NOT taken in place

82 further en-echo label keys = 230 leaves, measured over all 417 declared form-field paths on 3ff6ddf1e, and ⛔ 0 of them sit inside repeater-row-properties.test.ts's 124-property population ⇒ no pin sees any of them.

⭐ The dev declined the in-place fix and named which condition failed: the bounded-fix exemption's condition (2), mechanical — because each key needs the same per-leaf judgement this card exists to demand, which is the opposite of mechanical. That is the correct refusal, and refusing for a named reason is worth more than refusing.

Filed separately by this seat. Also recorded from the same sweep, ⛔ not filed: the dataset panel's own label/description and four sections are echoes in all three locales (counted inside the 82), and report.sections.* is a population no census covers and no pin sees — ⚠️ 承接者:无, which is precisely why it is written down.

Landing

packages/platform-objects + .changeset/ + a test ⇒ ⛔ not a governed surface. Ordinary queue landing applies once every check is green or an expected skip.

---

Generated by Claude Code

[huangyiirene]

LANDED — PR #19401, verified by two readings. Card closed by Fixes; pm:dispatched removed by hand in the same act.

domain:engine#1, round 7. Written 2026-09-20T17:30Z.

reading	result
squash on origin/main	6af29010b9a367c1079f42a8b2a99034a6128a10 — parent count 1, and --is-ancestor origin/main ⇒ YES
the decision ledger	report-dataset-panel-echo-decisions.test.ts is on the tree

⭐ The content reading is a DELTA with its control moving the other way

⛔ Not a grep for a string — this seat had already measured the whole population before the landing, so the landing has a prediction to fail:

	before	after	Δ
label keys echoing their en source in all three locales	83	79	−4 ⇐ exactly the four decided keys
leaves	249	237	−12
⭐ POSITIVE CONTROL — labels genuinely translated (zh-CN)	455	459	+4

⇒ the echo count fell by four and the translated count rose by four, in the same measurement, on the same population. ⚠️ A parser matching too broadly cannot produce a two-directional agreement like that — which is precisely what this seat's earlier raw-string control on #19399 could not do, and said so on that card.

What this card delivered, restated because it is the unusual part

⛔ Not twelve translations. Twenty-one leaves decided one at a time, each with its recorded reason, and one leaf (dataset.fields.measures.helpText, already translated in all three) left as a found negative control. The reasons live in a ledger rather than prose — one row per leaf with its verdict, its reason and the en source it was judged against, pinned equal to the live source so rewording the English turns the row red, and an echo verdict must carry its own per-locale reason or the suite refuses it.

⭐ And the dev disclosed an ablation run that measured nothing — its mutation and restore sat in one shell that exited between them, so the suite ran against an already-restored file and printed a meaningless green. It named it and refused to count it. ⚠️ That is the same failure class this seat hit twice on its own CI watches this round; the difference is disclosure.

Out of scope, filed

The sweep behind this card found 82 further en-echo label keys on its base. This seat re-measured independently (83 keys / 249 leaves on its own base and population, with the positive control above) and filed #19403, stating in the card that the two numbers differ by definition, ⛔ not by disagreement — different base, and a wider population — and that whoever takes it must say which population it is fixing.

⛔ The dev correctly declined to take them in place and named the failing condition: mechanical. Each key needs the same per-leaf judgement this card exists to demand.

---

Generated by Claude Code