finding(i18n): outside an i18next instance useSafeFieldLabel returns a FRESH object every render — measured 4 distinct identities in 4 renders, re-keying every consumer memo it feeds

Measured while fixing #4629. Filed unassigned; out of that card's scope (its file surface is packages/plugin-dashboard/**, this lives in packages/i18n). Duplicate-searched (keyword + useObjectLabel / useSafeFieldLabel / identity / I18nProvider): nothing open.

What was measured

useSafeFieldLabel() → useObjectLabel(), which memoizes its whole returned object on react-i18next's [t, i18n] (packages/i18n/src/useObjectLabel.ts:58 … :617). When there is no i18next instance to bind to, useTranslation warns NO_I18NEXT_INSTANCE and hands back a fresh t/i18n each render, so that memo never holds.

Direct probe — one component calling useSafeFieldLabel(), mounted with no provider anywhere in the module, then re-rendered three times by its host:

NO i18next instance:  { renders: 4, distinctFieldLabel: 4, distinctObject: 4 }
i18next instance present: { renders: 4, distinctFieldLabel: 1, distinctObject: 1 }

So the hook is stable exactly when an instance exists and unstable exactly when one does not.

Why that is the wrong way round

useSafeFieldLabel's own docstring names the no-provider case as the one it exists to serve:

Safe wrapper for useObjectLabel that falls back to identity functions when no I18nProvider is available. Suitable for plugin components that may be rendered outside an i18n context.

And useObjectLabel's memo docstring names the failure it was added to prevent:

Returning a fresh object on every render was busting downstream useMemo/useCallback deps in heavy consumers like ListView.filterFields, causing avoidable recomputation.

That protection is present with a provider and absent without one — i.e. it lapses in the configuration the safe wrapper advertises. The module-level SAFE_FIELD_LABEL_FALLBACK constant is not reached here: useObjectLabel() returns a (fresh) object, so the ?? never fires.

Measured consequence in a real consumer

ObjectDataTable's derivedColumns memo lists fieldLabel and fieldOptionLabel among its dependencies. Rendered outside an i18next instance, that memo recomputes on every render regardless of #4629 — measured on the #4629 branch as 8 buildFieldMeta calls over mount + 3 host re-renders (2 with a provider). #4629's fix is unaffected and correct; this is a second, independent identity source in the same dependency list, which is why ObjectDataTable.stableEmptyRows.test.tsx renders inside a real I18nProvider and says so in its docblock.

Every other memo keyed on these closures is exposed the same way — ListView.filterFields is named in the docstring above.

Options for whoever takes it

Bind the fallback to the module-level constant. Return SAFE_FIELD_LABEL_FALLBACK (already frozen at module scope, already the intended shape) when there is no instance, instead of a fresh memo output. Smallest change; makes the advertised no-provider path genuinely stable.
Stabilize useObjectLabel's memo key. Key on something that does not churn without an instance (e.g. i18n?.language plus a ready flag) rather than the t/i18n object identities.

Option 1 looks right on its face — the constant exists for exactly this case — but the two differ in behaviour once an instance appears later, so this wants a deliberate call rather than a drive-by. Not attempted here.

Generated by Claude Code

Triage: pm:queue + domain:ui, type Bug — the hook's own docstring declares stability protection for exactly the no-provider case where it measurably lapses (declared ≠ enforced, with a measured downstream consequence in ObjectDataTable.derivedColumns). The option-1 vs option-2 choice (bind the frozen fallback vs stabilize the memo key) is the implementer's, with one dispatch clause: whichever is chosen must pin the late-instance transition (a provider appearing after first render) so the two options' behavioural difference the card flags is a test, not a surprise. Assert on identity/recompute counts, not on rendered output — the card's own guidance.

Generated by Claude Code

Claim: domain:ui execution seat, session session_012u2pRjcqAYtoEjgr3wwhnK · branch claude/issue-5564-safe-field-label-identity · round 4.

⛔ Option 1 is not the "smallest change" the card says it is — measured

The card offers two options and calls option 1 (return SAFE_FIELD_LABEL_FALLBACK when there is no instance) the smallest, "already the intended shape". I measured that on origin/main before dispatching, and as literally written it is a breaking runtime change:

useObjectLabel()'s memo returns an object with 27 members: objectLabel, objectDescription, fieldLabel, fieldOptionLabel, translateOptions, appLabel, appDescription, navGroupLabel, dashboardLabel, dashboardDescription, dashboardActionLabel, widgetTitle, widgetDescription, widgetSubCaption, pageLabel, reportLabel, viewLabel, viewDescription, viewEmptyState, sectionLabel, actionLabel, actionConfirm, actionSuccess, actionDescription, actionResultDialog, actionParamText, actionParamOptionLabel.
SAFE_FIELD_LABEL_FALLBACK has 5: fieldLabel, translateOptions, fieldOptionLabel, sectionLabel, actionLabel.

So option 1 drops 22 members on the no-provider path — the exact path the wrapper advertises. Most callers destructure only within the five, but three take the whole object, and one of them calls outside the set:

packages/plugin-detail/src/renderers/record-reference-rail.tsx:109   const i18n = useSafeFieldLabel() as any;   →  i18n.objectLabel(...)
packages/plugin-detail/src/renderers/record-related-list.tsx:88      const i18n = useSafeFieldLabel();
packages/plugin-tree/src/ObjectTree.tsx:368                          const i18n = useSafeFieldLabel();  →  i18n.fieldLabel(...)

record:reference-rail would get i18n.objectLabel is not a function outside a provider. useSafeFieldLabel is also publicly exported (packages/i18n/src/index.ts:60), so external consumers can reach any of the 27.

This does not veto option 1 — it means option 1 costs completing the fallback to all 27 members, which changes its size argument. Weigh it against option 2 (stabilize the memo key on something that does not churn without an instance) with that cost included, and state which you chose and why. If you pick option 1, the fallback must cover every member the real object exposes, and a test must pin that the two shapes agree — otherwise the next member added to useObjectLabel silently reintroduces the crash.

Triage's dispatch clause is binding

Whichever option you choose must pin the late-instance transition — a provider appearing after first render — so the behavioural difference between the two options is a test rather than a surprise. Assert on identity / recompute counts, not on rendered output: nothing renders wrong today, so a rendering assertion is green against the broken code.

Corroboration worth reading

packages/plugin-charts/src/ObjectChart.tsx:268-271 already carries a hand-rolled ref workaround with this comment: "the i18n hook returns a fresh function reference on every render, which would otherwise invalidate fetchData's useCallback identity and trigger an infinite refetch loop." An independent consumer hit this and worked around it locally — useful evidence for the fix, and worth noting whether your change makes that workaround removable (do not remove it here; it is outside this surface).

Note also that the existing comment at useSafeFieldLabel asserts "It already returns a stable memoized object" — that is the declaration this card measures as false without an instance. Correct it in the same stroke.

Surface

packages/i18n/src/useObjectLabel.ts + tests + changeset. packages/i18n is a published package: no new exports, no signature changes beyond what the chosen option requires.

Generated by Claude Code

fix(i18n): hold useObjectLabel's memo when no i18next instance is bound

{
  "issue": 5564,
  "status": "done",
  "branch": "claude/issue-5564-safe-field-label-identity",
  "pr": "https://github.com/objectstack-ai/objectui/pull/5585",
  "premise_still_valid": true,
  "summary": "Chose OPTION 2 (stabilize the memo key), not option 1. Decisive reason found while verifying the card: ListView.filterFields — the consumer useObjectLabel's memo docstring names as its reason to exist — reaches the hook via useListFieldLabel, which calls useObjectLabel() DIRECTLY (packages/plugin-list/src/ListView.tsx:683), not useSafeFieldLabel(). Option 1 only changes the wrapper, so it would leave that consumer and 20+ other direct useObjectLabel() call sites (app-shell, plugin-dashboard, components, react) still broken, plus plugin-timeline's own hand-rolled useObjectLabel() ?? FALLBACK wrapper at ObjectTimeline.tsx:37. Option 1 would additionally cost a hand-maintained 27-member parallel shape and a pin test; option 2 needs no parallel shape because there is one code path, and it preserves today's return VALUES exactly while changing only identity, which is the defect. Implementation: both memo deps are pinned to module-level constants while no i18next instance is bound. The substitution is unobservable, not merely convenient — every t() call in the module sits inside a 'for (... of getAppNamespaces())' loop and getAppNamespaces() returns the empty array under exactly the same predicate, so the closures cannot read t or i18n at all while it is in effect. The predicate now has ONE definition (hasUsableI18nInstance) shared by the memo key and getAppNamespaces so they cannot drift. On the bound path the dep list is literally [t, i18n] as before. The false comment on useSafeFieldLabel ('It already returns a stable memoized object') is corrected. No new exports, no signature changes, identical 27-member surface on both paths.",
  "premise_verification": {
    "27_vs_5_counts": "CONFIRMED EXACTLY, verified independently rather than taken on trust. useObjectLabel's returned literal (lines 240-616 pre-fix) has 27 top-level members in the same order the claim comment lists; SAFE_FIELD_LABEL_FALLBACK has 5 (fieldLabel, translateOptions, fieldOptionLabel, sectionLabel, actionLabel). Both counts are also now asserted live by test 5.",
    "memo_and_hook_shape": "CONFIRMED. Memo opens at :58, closes '}, [t, i18n]);' at :617; SAFE_FIELD_LABEL_FALLBACK at :626; useSafeFieldLabel exported at packages/i18n/src/index.ts:60.",
    "whole_object_callers": "CONFIRMED all three (record-reference-rail.tsx:109 with i18n.objectLabel(...), record-related-list.tsx:88, ObjectTree.tsx:368).",
    "refinement_to_the_card": "The card says useTranslation 'hands back a fresh t/i18n each render'. Only i18n churns. t is already stable when unbound — react-i18next returns the module-level notReadyT out of a module-level notReadySnapshot. The churn source is precisely 'const finalI18n = i18n || {}' in react-i18next's useTranslation, a fresh empty object that then feeds that hook's own useMemo dep list. Conclusion unchanged; the fix pins both deps anyway rather than relying on t's stability, which is a react-i18next implementation detail."
  },
  "tests": "New file packages/i18n/src/__tests__/useObjectLabel-identity-5564.test.tsx, 5 cases, all asserting on identity/recompute counts rather than rendered output. Tests detach react-i18next's process-global instance and ASSERT that precondition, because createI18n registers itself globally via initReactI18next — without that a probe meant to run unbound would silently bind to an earlier test's instance and assert nothing. Case 4 pins triage's late-instance clause: the provider element sits at the same tree position for every render so the probe is UPDATED rather than remounted when the instance arrives (a remount would reset the memo and prove nothing), asserting exactly two identities over six renders plus that the first object resolves no translations and the second does. GATES, exit codes captured before any pipe, all at commit b492c9796, each quoted from the gate's own verdict line: deps build exit 0 ('packages/core build: Done'); 'pnpm --filter @object-ui/i18n type-check' exit 0 with the script echoed ('tsc --noEmit && tsc -p tsconfig.test.json' — so not a zero-match pnpm run); 'pnpm --filter @object-ui/i18n lint' exit 0, '34 problems (0 errors, 34 warnings)', all warnings pre-existing 'any' usages outside this diff and none in the new test file; 'pnpm exec vitest run packages/i18n/' = 'Test Files 50 passed (50)' / 'Tests 878 passed (878)'; 'pnpm exec vitest run packages/plugin-detail/' = 'Test Files 90 passed (90)' / 'Tests 848 passed (848)'; 'pnpm exec vitest run packages/plugin-dashboard/' = 'Test Files 72 passed (72)' / 'Tests 653 passed (653)'; check-control-bytes 'OK (scanned 4638 tracked text file(s))'; check-changeset-presence '2 source file(s) of 1 released package(s) changed, and this change declares 1 changeset(s)'; check-changeset-no-major 'No changeset declares a major bump.'; check-changeset-fixed 'All workspace packages are in the changeset fixed group.'. ABLATION: fix committed FIRST, then source-only revert under a 'trap ... EXIT INT TERM' restore. Mutation confirmed ON DISK with anchored grep -c in both directions before reading any result — fix markers hasUsableI18nInstance/NO_INSTANCE_T/'const bound =' all 0, pre-fix gate 'typeof i18n.getResourceBundle' back to 1, pre-fix destructure back to 1 — plus git diff --stat. Ablated run VERDICT command-exit 1, 3 failed / 2 passed: 'expected 4 to be 1' (no instance, 4 renders), 'expected 4 to be 1' (useSafeFieldLabel, 4 renders), 'expected 3 to be 1' (late-instance test, unbound phase). The first two reproduce the card's measured 4-in-4 exactly. The two that stayed green are the right two — the with-instance guard and the shape pin — which is what makes it an ablation rather than a blanket break. Restored via the trap, git status --porcelain empty (byte-identical), re-run green. NO REBUILD IS IN THE LOOP and that is load-bearing here: the test imports the ablation target by relative source path, not through the package's exports/dist, so there is no stale-dist path that could have faked either the red or the green.",
  "narrowing_declared": [
    "Repo-wide 'pnpm lint' was NOT run; 'pnpm --filter @object-ui/i18n lint' was. This cannot hide a failure: the diff touches only packages/i18n, and eslint.config.js enables no type-aware linting (no projectService, no parserOptions.project), so the change cannot move the verdict on any file it does not contain.",
    "2 of roughly 15 consumer packages were tested locally (plugin-detail, plugin-dashboard — chosen because plugin-detail holds the two whole-object useSafeFieldLabel() consumers a shape change would break first, and plugin-dashboard holds ObjectDataTable, the consumer whose derivedColumns recomputation the card measured). This one IS a real narrowing; CI runs the full farm."
  ],
  "open_questions": [],
  "out_of_scope_findings": [
    "filed as #5587: ObjectChart's fieldOptionLabel ref workaround (plugin-charts/src/ObjectChart.tsx:268-271) becomes removable once this lands — my fix DOES make it removable, since fieldOptionLabel is now identity-stable with or without a provider; not removed here per the dispatch fence. The same issue notes plugin-timeline's ObjectTimeline.tsx:25-37 useSafeObjectLabel, a third hand-rolled duplicate of useSafeFieldLabel."
  ],
  "contradicting_the_card": [
    "The card frames option 1 as making 'the advertised no-provider path genuinely stable'. It would not stabilise ListView.filterFields, the consumer the memo docstring itself names — that call site uses useObjectLabel() directly. This is the fact that decided the option.",
    "The card attributes the churn to a fresh 't'/'i18n' pair. Only 'i18n' churns; 't' is a stable module-level function when unbound."
  ],
  "deviation_from_dispatch_wording": "Dispatch item 4 asked for the bare '_Generated by [Claude Code](https://claude.ai/code)_' footer on every PR body and comment. The PR body uses the session-URL form instead, because a bare footer is stripped from a PR body on every subsequent edit; comments keep the bare form, which survives there. Flagging rather than silently choosing."
}

Generated by Claude Code

os-dev-report

Marker restored as literal text: GitHub's sanitizer ate the angle-bracket marker on the previous comment, making that copy invisible to a marker scan. Same report, verbatim — this is the authoritative machine-readable copy; the one above is the human-readable duplicate.

{
  "issue": 5564,
  "status": "done",
  "branch": "claude/issue-5564-safe-field-label-identity",
  "pr": "https://github.com/objectstack-ai/objectui/pull/5585",
  "premise_still_valid": true,
  "summary": "Chose OPTION 2 (stabilize the memo key), not option 1. Decisive reason found while verifying the card: ListView.filterFields — the consumer useObjectLabel's memo docstring names as its reason to exist — reaches the hook via useListFieldLabel, which calls useObjectLabel() DIRECTLY (packages/plugin-list/src/ListView.tsx:683), not useSafeFieldLabel(). Option 1 only changes the wrapper, so it would leave that consumer and 20+ other direct useObjectLabel() call sites (app-shell, plugin-dashboard, components, react) still broken, plus plugin-timeline's own hand-rolled useObjectLabel() ?? FALLBACK wrapper at ObjectTimeline.tsx:37. Option 1 would additionally cost a hand-maintained 27-member parallel shape and a pin test; option 2 needs no parallel shape because there is one code path, and it preserves today's return VALUES exactly while changing only identity, which is the defect. Implementation: both memo deps are pinned to module-level constants while no i18next instance is bound. The substitution is unobservable, not merely convenient — every t() call in the module sits inside a 'for (... of getAppNamespaces())' loop and getAppNamespaces() returns the empty array under exactly the same predicate, so the closures cannot read t or i18n at all while it is in effect. The predicate now has ONE definition (hasUsableI18nInstance) shared by the memo key and getAppNamespaces so they cannot drift. On the bound path the dep list is literally [t, i18n] as before. The false comment on useSafeFieldLabel ('It already returns a stable memoized object') is corrected. No new exports, no signature changes, identical 27-member surface on both paths.",
  "premise_verification": {
    "27_vs_5_counts": "CONFIRMED EXACTLY, verified independently rather than taken on trust. useObjectLabel's returned literal (lines 240-616 pre-fix) has 27 top-level members in the same order the claim comment lists; SAFE_FIELD_LABEL_FALLBACK has 5 (fieldLabel, translateOptions, fieldOptionLabel, sectionLabel, actionLabel). Both counts are also now asserted live by test 5.",
    "memo_and_hook_shape": "CONFIRMED. Memo opens at :58, closes '}, [t, i18n]);' at :617; SAFE_FIELD_LABEL_FALLBACK at :626; useSafeFieldLabel exported at packages/i18n/src/index.ts:60.",
    "whole_object_callers": "CONFIRMED all three (record-reference-rail.tsx:109 with i18n.objectLabel(...), record-related-list.tsx:88, ObjectTree.tsx:368).",
    "refinement_to_the_card": "The card says useTranslation 'hands back a fresh t/i18n each render'. Only i18n churns. t is already stable when unbound — react-i18next returns the module-level notReadyT out of a module-level notReadySnapshot. The churn source is precisely 'const finalI18n = i18n || {}' in react-i18next's useTranslation, a fresh empty object that then feeds that hook's own useMemo dep list. Conclusion unchanged; the fix pins both deps anyway rather than relying on t's stability, which is a react-i18next implementation detail."
  },
  "tests": "New file packages/i18n/src/__tests__/useObjectLabel-identity-5564.test.tsx, 5 cases, all asserting on identity/recompute counts rather than rendered output. Tests detach react-i18next's process-global instance and ASSERT that precondition, because createI18n registers itself globally via initReactI18next — without that a probe meant to run unbound would silently bind to an earlier test's instance and assert nothing. Case 4 pins triage's late-instance clause: the provider element sits at the same tree position for every render so the probe is UPDATED rather than remounted when the instance arrives (a remount would reset the memo and prove nothing), asserting exactly two identities over six renders plus that the first object resolves no translations and the second does. GATES, exit codes captured before any pipe, all at commit b492c9796, each quoted from the gate's own verdict line: deps build exit 0 ('packages/core build: Done'); 'pnpm --filter @object-ui/i18n type-check' exit 0 with the script echoed ('tsc --noEmit && tsc -p tsconfig.test.json' — so not a zero-match pnpm run); 'pnpm --filter @object-ui/i18n lint' exit 0, '34 problems (0 errors, 34 warnings)', all warnings pre-existing 'any' usages outside this diff and none in the new test file; 'pnpm exec vitest run packages/i18n/' = 'Test Files 50 passed (50)' / 'Tests 878 passed (878)'; 'pnpm exec vitest run packages/plugin-detail/' = 'Test Files 90 passed (90)' / 'Tests 848 passed (848)'; 'pnpm exec vitest run packages/plugin-dashboard/' = 'Test Files 72 passed (72)' / 'Tests 653 passed (653)'; check-control-bytes 'OK (scanned 4638 tracked text file(s))'; check-changeset-presence '2 source file(s) of 1 released package(s) changed, and this change declares 1 changeset(s)'; check-changeset-no-major 'No changeset declares a major bump.'; check-changeset-fixed 'All workspace packages are in the changeset fixed group.'. ABLATION: fix committed FIRST, then source-only revert under a 'trap ... EXIT INT TERM' restore. Mutation confirmed ON DISK with anchored grep -c in both directions before reading any result — fix markers hasUsableI18nInstance/NO_INSTANCE_T/'const bound =' all 0, pre-fix gate 'typeof i18n.getResourceBundle' back to 1, pre-fix destructure back to 1 — plus git diff --stat. Ablated run VERDICT command-exit 1, 3 failed / 2 passed: 'expected 4 to be 1' (no instance, 4 renders), 'expected 4 to be 1' (useSafeFieldLabel, 4 renders), 'expected 3 to be 1' (late-instance test, unbound phase). The first two reproduce the card's measured 4-in-4 exactly. The two that stayed green are the right two — the with-instance guard and the shape pin — which is what makes it an ablation rather than a blanket break. Restored via the trap, git status --porcelain empty (byte-identical), re-run green. NO REBUILD IS IN THE LOOP and that is load-bearing here: the test imports the ablation target by relative source path, not through the package's exports/dist, so there is no stale-dist path that could have faked either the red or the green.",
  "narrowing_declared": [
    "Repo-wide 'pnpm lint' was NOT run; 'pnpm --filter @object-ui/i18n lint' was. This cannot hide a failure: the diff touches only packages/i18n, and eslint.config.js enables no type-aware linting (no projectService, no parserOptions.project), so the change cannot move the verdict on any file it does not contain.",
    "2 of roughly 15 consumer packages were tested locally (plugin-detail, plugin-dashboard — chosen because plugin-detail holds the two whole-object useSafeFieldLabel() consumers a shape change would break first, and plugin-dashboard holds ObjectDataTable, the consumer whose derivedColumns recomputation the card measured). This one IS a real narrowing; CI runs the full farm."
  ],
  "open_questions": [],
  "out_of_scope_findings": [
    "filed as #5587: ObjectChart's fieldOptionLabel ref workaround (plugin-charts/src/ObjectChart.tsx:268-271) becomes removable once this lands — my fix DOES make it removable, since fieldOptionLabel is now identity-stable with or without a provider; not removed here per the dispatch fence. The same issue notes plugin-timeline's ObjectTimeline.tsx:25-37 useSafeObjectLabel, a third hand-rolled duplicate of useSafeFieldLabel."
  ],
  "contradicting_the_card": [
    "The card frames option 1 as making 'the advertised no-provider path genuinely stable'. It would not stabilise ListView.filterFields, the consumer the memo docstring itself names — that call site uses useObjectLabel() directly. This is the fact that decided the option.",
    "The card attributes the churn to a fresh 't'/'i18n' pair. Only 'i18n' churns; 't' is a stable module-level function when unbound."
  ],
  "deviation_from_dispatch_wording": "Dispatch item 4 asked for the bare '_Generated by [Claude Code](https://claude.ai/code)_' footer on every PR body and comment. The PR body uses the session-URL form instead, because a bare footer is stripped from a PR body on every subsequent edit; comments keep the bare form, which survives there. Flagging rather than silently choosing."
}

Generated by Claude Code

ACCEPT — PR #5585

Gates read by name on head b492c9796, all 22 runs accounted for: Test (shard 1/4) (2/4) (3/4) (4/4), Type Check, Lint, Build & E2E, Build Docs, Bundle Analysis, Doc Snippet Type Check, Doc Component Type Check, Internal Docs Link Check, Changeset Declaration, Changeset Bump Policy, Changeset Fixed Group Check, Control Byte Scan, Skill Guide Path Check, Live E2E (informational), label — every one completed: success, plus the three usual skipped matrix no-ops. Console budget PASS; useObjectLabel.js 30.51 → 33.40 KB raw (7.57 → 8.71 gzipped), which is the predicate, two constants and their doc comments.

Option 2 was the right call, and you found a better reason for it than I gave you

My dispatch framed the choice around what option 1 would cost — the 27-vs-5 shape gap and the record-reference-rail crash. You found that option 1 would not work, which is the stronger objection and the one that actually decides it:

ListView.filterFields — the consumer useObjectLabel's memo docstring names as its reason to exist — reaches the hook through useListFieldLabel, which calls useObjectLabel() directly.

Verified independently on origin/main:

packages/plugin-list/src/ListView.tsx:683   const { fieldLabel, actionLabel, objectLabel } = useObjectLabel();
packages/plugin-list/src/ListView.tsx:729   const { translateOptions } = useSafeFieldLabel();

The same file uses both hooks for different things, and the filter fields path — the named victim — is on the direct one. Option 1 touches only the wrapper, so it would have left the docstring's own example consumer broken.

Your "20+ other direct call sites" is also an under-claim: measured 33 direct useObjectLabel() call sites across 9 packages (app-shell, components, core, layout, plugin-dashboard, plugin-form, plugin-list, plugin-timeline, react), plus ObjectTimeline.tsx:36-37's hand-rolled useObjectLabel() ?? OBJECT_LABEL_FALLBACK, confirmed.

The correctness argument holds — I checked it rather than accepting it

The whole fix rests on the substitution being unobservable: that while the stand-ins are pinned, nothing can read them. Verified by enumerating every site in the module:

site	gated by
`t(...)` at `:165`	inside `for (const ns of namespaces)`, `namespaces = getAppNamespaces()` at `:158`
`t(...)` at `:604`	inside `for (const ns of getAppNamespaces())` at `:602`
`i18n.language` `:131`, `i18n.getResourceBundle` `:132`	after `if (!hasUsableI18nInstance(i18n)) return []` at `:130`

Those are the only t() calls and the only i18n reads in the file. So while unbound, getAppNamespaces() returns [], both loops have zero iterations, and neither stand-in is reachable.

Routing the memo key and the gate through one shared hasUsableI18nInstance is what makes that argument stable rather than merely true today. Two separate predicates could have drifted into a window where the substitution is active but getAppNamespaces() returns a non-empty list — and then NO_INSTANCE_T would really be called. Collapsing the inline check at :131 into the shared predicate closes that. Keeping NO_INSTANCE_T faithful to react-i18next's not-ready t is the correct belt-and-braces for the same reason.

The card was wrong on one detail and you corrected it

The card attributes the churn to a fresh t/i18n pair. Only i18n churns — t is react-i18next's module-level notReadyT. Pinning both dependencies anyway is right: t's stability is an implementation detail of a third-party package, not a contract.

Tests

Five cases, all on identity/recompute counts. Two things raise this above the bar:

The global-instance detachment is asserted, not assumed. createI18n registers itself globally via initReactI18next, so a probe meant to run unbound would otherwise silently bind to an earlier test's instance and assert nothing. Asserting the precondition is what stops the whole suite being vacuous.
Case 4 keeps the provider element at the same tree position so the probe is updated rather than remounted when the instance arrives. A remount would reset the memo and prove nothing — that is exactly how a late-instance test passes for the wrong reason, and it was avoided deliberately.

Case 5 pins the 27-member surface on both paths and that objectLabel is callable unbound, which guards the record-reference-rail hazard permanently rather than relying on this PR's choice not to narrow it.

Ablation is a real ablation: 3 red / 2 green, and the two that stayed green are the right two (the with-instance guard and the shape pin).

Follow-on

#5587 filed — ObjectChart's fieldOptionLabel ref workaround becomes removable once this lands, and ObjectTimeline's useSafeObjectLabel is a third hand-rolled duplicate. Correctly not removed here: unwinding a refetch-loop guard wants its own review and its own test, not a ride-along in a diff about memo identity. I added the missing finding label so triage's queue can see it.

Proceeding to ready-for-review and the merge queue.

Generated by Claude Code

fix(i18n): hold useObjectLabel's memo when no i18next instance is bou…

Navigation Menu

Repository navigation

finding(i18n): outside an i18next instance `useSafeFieldLabel` returns a FRESH object every render — measured 4 distinct identities in 4 renders, re-keying every consumer memo it feeds #5564

What was measured

Why that is the wrong way round

Measured consequence in a real consumer

Options for whoever takes it

⛔ Option 1 is not the "smallest change" the card says it is — measured

Triage's dispatch clause is binding

Corroboration worth reading

Surface

ACCEPT — PR #5585

Option 2 was the right call, and you found a better reason for it than I gave you

The correctness argument holds — I checked it rather than accepting it

The card was wrong on one detail and you corrected it

Tests

Follow-on

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

finding(i18n): outside an i18next instance useSafeFieldLabel returns a FRESH object every render — measured 4 distinct identities in 4 renders, re-keying every consumer memo it feeds #5564

Description

What was measured

Why that is the wrong way round

Measured consequence in a real consumer

Options for whoever takes it

Activity

huangyiirene commented on Aug 21, 2026

os-sales commented on Aug 21, 2026

⛔ Option 1 is not the "smallest change" the card says it is — measured

Triage's dispatch clause is binding

Corroboration worth reading

Surface

os-sales commented on Aug 21, 2026

os-sales commented on Aug 21, 2026

os-sales commented on Aug 21, 2026

ACCEPT — PR #5585

Option 2 was the right call, and you found a better reason for it than I gave you

The correctness argument holds — I checked it rather than accepting it

The card was wrong on one detail and you corrected it

Tests

Follow-on

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions

finding(i18n): outside an i18next instance `useSafeFieldLabel` returns a FRESH object every render — measured 4 distinct identities in 4 renders, re-keying every consumer memo it feeds #5564