fix(devx): add a property-chain leg to check-i18n-dead-keys

[claude]

Fixes#6666

A consumer that imports a locale pack object and reads it by property access is invisible to both of the sweep's legs at once. The AST pass classifies t()/tt() call sites and there is no call; the text safety net greps the full dotted key and the namespace segment is bound to a local variable, so the dotted key is never spelled. The key lands in CONFIRMED — the tier documented as the safest thing to delete — while a shipping screen renders it.

textFootprint() now probes each candidate's property chain (the key minus its leading namespace segment, leading dot kept) alongside the full dotted key, in the same single grep -F pass.

The card's tier claim is stale — measured, and stated rather than repeated

The card measured the five keys at CONFIRMED on 813bf8323. On origin/main at 4357ec754 they are NEEDS-REVIEW, demoted by side effect: PR #6662's retirement pin and ar-participle-orthography-6610.test.ts both spell the dotted keys, which gives them a textual footprint. Per the triage ruling that is a mitigation, not a fix — the defect is unchanged, and the before/after evidence below is therefore built on what the hits are rather than on a tier flip.

	before (4357ec754)	after (fb4ccd348)
CONFIRMED	115	115
NEEDS-REVIEW	247	247
candidates	362	362
hits naming the real reader LoadingScreen.tsx	0	7 keys

Nothing was demoted; no pre-existing hit was lost. What changed is that seven keys now name the file that actually renders them, marked (via property chain):

• the five from the card, and
• console.actions.retry / console.actions.retrying — two more the card did not list, read by the same file the same way (LoadingScreen.tsx:118,123).

Both negative controls stay CONFIRMED with an empty hit list.

Three boundaries, each pinned by an ablation

1. Three segments or more, never two. A two-segment key's chain is a single generic word (.ok, .no, .empty). propertyChainProbe() returns null below three segments. Two-segment keys stay on the enumerated importer list, read by hand — which is what refactor(i18n): retire 25 confirmed-dead locale keys across all ten packs #6662 did.
2. The match must end at a property boundary. Otherwise a leaf would demote on a reader of a longer sibling leaf that it merely prefixes.
3. A chain-only hit is suffixed (via property chain), so the report never sends a reader grepping for a dotted key that file does not contain.

The importer list is written into the file

Per the ruling, the pack-object importer class is enumerated in the script header with the two-stage grep that re-derives it. The outboundAgentText.ts caveat is stated explicitly: it indexes its subtree dynamically (ai?.[key]), so neither leg sees the read — its four keys land in NEEDS-REVIEW only because those property names happen to appear as string literals in the OutboundAgentTextKey union above them. By luck, not by design. Replace that union with anything generated and all four drop to CONFIRMED with a live consumer still reading them.

Self-pollution, and why the tests spell keys in segments

textFootprint() greps scripts/ too, so a dotted key written into the test file would make that file a hit for it — the negative controls would stop being reader-less because the test asserting they are reader-less mentioned them. The real-repo controls are therefore joined from segments, and this was verified empirically: with the new test file on disk, console.error.serverUnreachable and console.error.timeout are still CONFIRMED and CONFIRMED is still 115.

Verification — all on fb4ccd348

Whole scripts/__tests__ tree (not a subject-matter subset): Test Files 89 passed (89) · Tests 2512 passed (2512).

Ablations. No build step is involved — vitest loads scripts/*.mjs from source, there is no dist/ for a stale artifact to hide in. Each leg mutated with perl -0pi, the mutation proved on disk by counting the injected and removed text, restored with git checkout HEAD -- against an absolute path, and the restore proved byte-identical (git hash-object equal to the HEAD blob 851663da3 and git diff HEAD empty) before the next reading:

ablation	result
probe always returns null (leg removed)	Tests 4 failed, 23 passed — all four detection assertions red, every negative control still green
three-segment guard relaxed to two	Tests 2 failed, 25 passed
property-boundary check replaced by plain includes	Tests 1 failed, 26 passed

The first row is the discrimination proof: removing the leg loses detection and touches nothing on the negative side.

Gates, each exit code captured before any pipe:

• pnpm type-check:scripts — exit 0
• pnpm check:control-bytes — check-control-bytes: OK (scanned 5804 tracked text file(s); skipped 85 binary)
• pnpm check:shell-escape-residue — OK (4/4 root(s) resolved ...)
• pnpm check:vi-mock-specifiers — OK (4032 tracked source file(s) ...)
• pnpm check:esm-specifiers — exit 0
• node scripts/check-changeset-presence.mjs — No source of a released package changed in this range, so no changeset is owed.No changeset and no label are owed: the guarded surface is derived from published packages and scripts/ is not one.
• node scripts/check-lint-coverage.mjs — lint coverage: 46/46 packages linted, 0 with outstanding errors (0 total)
• eslint, run over the whole population rather than narrowed: eslint --no-inline-config --format json . resolved 4030 files from its own config; both changed files report 0 errors, 0 warnings. The run's own non-zero exit reflects 90 pre-existing errors in files outside the per-package lint scope, unchanged by this diff and confirmed green by the coverage gate above.

Generated by Claude Code

---

Generated by Claude Code

[claude]

Fixes#6666

A consumer that imports a locale pack object and reads it by property access is invisible to both of the sweep's legs at once. The AST pass classifies t()/tt() call sites and there is no call; the text safety net greps the full dotted key and the namespace segment is bound to a local variable, so the dotted key is never spelled. The key lands in CONFIRMED — the tier documented as the safest thing to delete — while a shipping screen renders it.

textFootprint() now probes each candidate's property chain (the key minus its leading namespace segment, leading dot kept) alongside the full dotted key, in the same single grep -F pass.

The card's tier claim is stale — measured, and stated rather than repeated

The card measured the five keys at CONFIRMED on 813bf8323. On origin/main at 4357ec754 they are NEEDS-REVIEW, demoted by side effect: PR #6662's retirement pin and ar-participle-orthography-6610.test.ts both spell the dotted keys, which gives them a textual footprint. Per the triage ruling that is a mitigation, not a fix — the defect is unchanged, and the before/after evidence below is therefore built on what the hits are rather than on a tier flip.

	before (4357ec754)	after (fb4ccd348)
CONFIRMED	115	115
NEEDS-REVIEW	247	247
candidates	362	362
hits naming the real reader LoadingScreen.tsx	0	7 keys

Nothing was demoted; no pre-existing hit was lost. What changed is that seven keys now name the file that actually renders them, marked (via property chain):

• the five from the card, and
• console.actions.retry / console.actions.retrying — two more the card did not list, read by the same file the same way (LoadingScreen.tsx:118,123).

Both negative controls stay CONFIRMED with an empty hit list.

Three boundaries, each pinned by an ablation

1. Three segments or more, never two. A two-segment key's chain is a single generic word (.ok, .no, .empty). propertyChainProbe() returns null below three segments. Two-segment keys stay on the enumerated importer list, read by hand — which is what refactor(i18n): retire 25 confirmed-dead locale keys across all ten packs #6662 did.
2. The match must end at a property boundary. Otherwise a leaf would demote on a reader of a longer sibling leaf that it merely prefixes.
3. A chain-only hit is suffixed (via property chain), so the report never sends a reader grepping for a dotted key that file does not contain.

The importer list is written into the file

Per the ruling, the pack-object importer class is enumerated in the script header with the two-stage grep that re-derives it. The outboundAgentText.ts caveat is stated explicitly: it indexes its subtree dynamically (ai?.[key]), so neither leg sees the read — its four keys land in NEEDS-REVIEW only because those property names happen to appear as string literals in the OutboundAgentTextKey union above them. By luck, not by design. Replace that union with anything generated and all four drop to CONFIRMED with a live consumer still reading them.

Self-pollution, and why the tests spell keys in segments

textFootprint() greps scripts/ too, so a dotted key written into the test file would make that file a hit for it — the negative controls would stop being reader-less because the test asserting they are reader-less mentioned them. The real-repo controls are therefore joined from segments, and this was verified empirically: with the new test file on disk, console.error.serverUnreachable and console.error.timeout are still CONFIRMED and CONFIRMED is still 115.

Verification — all on fb4ccd348

Whole scripts/__tests__ tree (not a subject-matter subset): Test Files 89 passed (89) · Tests 2512 passed (2512).

Ablations. No build step is involved — vitest loads scripts/*.mjs from source, there is no dist/ for a stale artifact to hide in. Each leg mutated with perl -0pi, the mutation proved on disk by counting the injected and removed text, restored with git checkout HEAD -- against an absolute path, and the restore proved byte-identical (git hash-object equal to the HEAD blob 851663da3 and git diff HEAD empty) before the next reading:

ablation	result
probe always returns null (leg removed)	Tests 4 failed, 23 passed — all four detection assertions red, every negative control still green
three-segment guard relaxed to two	Tests 2 failed, 25 passed
property-boundary check replaced by plain includes	Tests 1 failed, 26 passed

The first row is the discrimination proof: removing the leg loses detection and touches nothing on the negative side.

Gates, each exit code captured before any pipe:

• pnpm type-check:scripts — exit 0
• pnpm check:control-bytes — check-control-bytes: OK (scanned 5804 tracked text file(s); skipped 85 binary)
• pnpm check:shell-escape-residue — OK (4/4 root(s) resolved ...)
• pnpm check:vi-mock-specifiers — OK (4032 tracked source file(s) ...)
• pnpm check:esm-specifiers — exit 0
• node scripts/check-changeset-presence.mjs — No source of a released package changed in this range, so no changeset is owed.No changeset and no label are owed: the guarded surface is derived from published packages and scripts/ is not one.
• node scripts/check-lint-coverage.mjs — lint coverage: 46/46 packages linted, 0 with outstanding errors (0 total)
• eslint, run over the whole population rather than narrowed: eslint --no-inline-config --format json . resolved 4030 files from its own config; both changed files report 0 errors, 0 warnings. The run's own non-zero exit reflects 90 pre-existing errors in files outside the per-package lint scope, unchanged by this diff and confirmed green by the coverage gate above.

Generated by Claude Code

---

Generated by Claude Code

[claude]

Fixes#6666

A consumer that imports a locale pack object and reads it by property access is invisible to both of the sweep's legs at once. The AST pass classifies t()/tt() call sites and there is no call; the text safety net greps the full dotted key and the namespace segment is bound to a local variable, so the dotted key is never spelled. The key lands in CONFIRMED — the tier documented as the safest thing to delete — while a shipping screen renders it.

textFootprint() now probes each candidate's property chain (the key minus its leading namespace segment, leading dot kept) alongside the full dotted key, in the same single grep -F pass.

The card's tier claim is stale — measured, and stated rather than repeated

The card measured the five keys at CONFIRMED on 813bf8323. On origin/main at 4357ec754 they are NEEDS-REVIEW, demoted by side effect: PR #6662's retirement pin and ar-participle-orthography-6610.test.ts both spell the dotted keys, which gives them a textual footprint. Per the triage ruling that is a mitigation, not a fix — the defect is unchanged, and the before/after evidence below is therefore built on what the hits are rather than on a tier flip.

	before (4357ec754)	after (fb4ccd348)
CONFIRMED	115	115
NEEDS-REVIEW	247	247
candidates	362	362
hits naming the real reader LoadingScreen.tsx	0	7 keys

Nothing was demoted; no pre-existing hit was lost. What changed is that seven keys now name the file that actually renders them, marked (via property chain):

• the five from the card, and
• console.actions.retry / console.actions.retrying — two more the card did not list, read by the same file the same way (LoadingScreen.tsx:118,123).

Both negative controls stay CONFIRMED with an empty hit list.

Three boundaries, each pinned by an ablation

1. Three segments or more, never two. A two-segment key's chain is a single generic word (.ok, .no, .empty). propertyChainProbe() returns null below three segments. Two-segment keys stay on the enumerated importer list, read by hand — which is what refactor(i18n): retire 25 confirmed-dead locale keys across all ten packs #6662 did.
2. The match must end at a property boundary. Otherwise a leaf would demote on a reader of a longer sibling leaf that it merely prefixes.
3. A chain-only hit is suffixed (via property chain), so the report never sends a reader grepping for a dotted key that file does not contain.

The importer list is written into the file

Per the ruling, the pack-object importer class is enumerated in the script header with the two-stage grep that re-derives it. The outboundAgentText.ts caveat is stated explicitly: it indexes its subtree dynamically (ai?.[key]), so neither leg sees the read — its four keys land in NEEDS-REVIEW only because those property names happen to appear as string literals in the OutboundAgentTextKey union above them. By luck, not by design. Replace that union with anything generated and all four drop to CONFIRMED with a live consumer still reading them.

Self-pollution, and why the tests spell keys in segments

textFootprint() greps scripts/ too, so a dotted key written into the test file would make that file a hit for it — the negative controls would stop being reader-less because the test asserting they are reader-less mentioned them. The real-repo controls are therefore joined from segments, and this was verified empirically: with the new test file on disk, console.error.serverUnreachable and console.error.timeout are still CONFIRMED and CONFIRMED is still 115.

Verification — all on fb4ccd348

Whole scripts/__tests__ tree (not a subject-matter subset): Test Files 89 passed (89) · Tests 2512 passed (2512).

Ablations. No build step is involved — vitest loads scripts/*.mjs from source, there is no dist/ for a stale artifact to hide in. Each leg mutated with perl -0pi, the mutation proved on disk by counting the injected and removed text, restored with git checkout HEAD -- against an absolute path, and the restore proved byte-identical (git hash-object equal to the HEAD blob 851663da3 and git diff HEAD empty) before the next reading:

ablation	result
probe always returns null (leg removed)	Tests 4 failed, 23 passed — all four detection assertions red, every negative control still green
three-segment guard relaxed to two	Tests 2 failed, 25 passed
property-boundary check replaced by plain includes	Tests 1 failed, 26 passed

The first row is the discrimination proof: removing the leg loses detection and touches nothing on the negative side.

Gates, each exit code captured before any pipe:

• pnpm type-check:scripts — exit 0
• pnpm check:control-bytes — check-control-bytes: OK (scanned 5804 tracked text file(s); skipped 85 binary)
• pnpm check:shell-escape-residue — OK (4/4 root(s) resolved ...)
• pnpm check:vi-mock-specifiers — OK (4032 tracked source file(s) ...)
• pnpm check:esm-specifiers — exit 0
• node scripts/check-changeset-presence.mjs — No source of a released package changed in this range, so no changeset is owed.No changeset and no label are owed: the guarded surface is derived from published packages and scripts/ is not one.
• node scripts/check-lint-coverage.mjs — lint coverage: 46/46 packages linted, 0 with outstanding errors (0 total)
• eslint, run over the whole population rather than narrowed: eslint --no-inline-config --format json . resolved 4030 files from its own config; both changed files report 0 errors, 0 warnings. The run's own non-zero exit reflects 90 pre-existing errors in files outside the per-package lint scope, unchanged by this diff and confirmed green by the coverage gate above.

Generated by Claude Code

---

Generated by Claude Code

[claude]

Fixes#6666

A consumer that imports a locale pack object and reads it by property access is invisible to both of the sweep's legs at once. The AST pass classifies t()/tt() call sites and there is no call; the text safety net greps the full dotted key and the namespace segment is bound to a local variable, so the dotted key is never spelled. The key lands in CONFIRMED — the tier documented as the safest thing to delete — while a shipping screen renders it.

textFootprint() now probes each candidate's property chain (the key minus its leading namespace segment, leading dot kept) alongside the full dotted key, in the same single grep -F pass.

The card's tier claim is stale — measured, and stated rather than repeated

The card measured the five keys at CONFIRMED on 813bf8323. On origin/main at 4357ec754 they are NEEDS-REVIEW, demoted by side effect: PR #6662's retirement pin and ar-participle-orthography-6610.test.ts both spell the dotted keys, which gives them a textual footprint. Per the triage ruling that is a mitigation, not a fix — the defect is unchanged, and the before/after evidence below is therefore built on what the hits are rather than on a tier flip.

	before (4357ec754)	after (fb4ccd348)
CONFIRMED	115	115
NEEDS-REVIEW	247	247
candidates	362	362
hits naming the real reader LoadingScreen.tsx	0	7 keys

Nothing was demoted; no pre-existing hit was lost. What changed is that seven keys now name the file that actually renders them, marked (via property chain):

• the five from the card, and
• console.actions.retry / console.actions.retrying — two more the card did not list, read by the same file the same way (LoadingScreen.tsx:118,123).

Both negative controls stay CONFIRMED with an empty hit list.

Three boundaries, each pinned by an ablation

1. Three segments or more, never two. A two-segment key's chain is a single generic word (.ok, .no, .empty). propertyChainProbe() returns null below three segments. Two-segment keys stay on the enumerated importer list, read by hand — which is what refactor(i18n): retire 25 confirmed-dead locale keys across all ten packs #6662 did.
2. The match must end at a property boundary. Otherwise a leaf would demote on a reader of a longer sibling leaf that it merely prefixes.
3. A chain-only hit is suffixed (via property chain), so the report never sends a reader grepping for a dotted key that file does not contain.

The importer list is written into the file

Per the ruling, the pack-object importer class is enumerated in the script header with the two-stage grep that re-derives it. The outboundAgentText.ts caveat is stated explicitly: it indexes its subtree dynamically (ai?.[key]), so neither leg sees the read — its four keys land in NEEDS-REVIEW only because those property names happen to appear as string literals in the OutboundAgentTextKey union above them. By luck, not by design. Replace that union with anything generated and all four drop to CONFIRMED with a live consumer still reading them.

Self-pollution, and why the tests spell keys in segments

textFootprint() greps scripts/ too, so a dotted key written into the test file would make that file a hit for it — the negative controls would stop being reader-less because the test asserting they are reader-less mentioned them. The real-repo controls are therefore joined from segments, and this was verified empirically: with the new test file on disk, console.error.serverUnreachable and console.error.timeout are still CONFIRMED and CONFIRMED is still 115.

Verification — all on fb4ccd348

Whole scripts/__tests__ tree (not a subject-matter subset): Test Files 89 passed (89) · Tests 2512 passed (2512).

Ablations. No build step is involved — vitest loads scripts/*.mjs from source, there is no dist/ for a stale artifact to hide in. Each leg mutated with perl -0pi, the mutation proved on disk by counting the injected and removed text, restored with git checkout HEAD -- against an absolute path, and the restore proved byte-identical (git hash-object equal to the HEAD blob 851663da3 and git diff HEAD empty) before the next reading:

ablation	result
probe always returns null (leg removed)	Tests 4 failed, 23 passed — all four detection assertions red, every negative control still green
three-segment guard relaxed to two	Tests 2 failed, 25 passed
property-boundary check replaced by plain includes	Tests 1 failed, 26 passed

The first row is the discrimination proof: removing the leg loses detection and touches nothing on the negative side.

Gates, each exit code captured before any pipe:

• pnpm type-check:scripts — exit 0
• pnpm check:control-bytes — check-control-bytes: OK (scanned 5804 tracked text file(s); skipped 85 binary)
• pnpm check:shell-escape-residue — OK (4/4 root(s) resolved ...)
• pnpm check:vi-mock-specifiers — OK (4032 tracked source file(s) ...)
• pnpm check:esm-specifiers — exit 0
• node scripts/check-changeset-presence.mjs — No source of a released package changed in this range, so no changeset is owed.No changeset and no label are owed: the guarded surface is derived from published packages and scripts/ is not one.
• node scripts/check-lint-coverage.mjs — lint coverage: 46/46 packages linted, 0 with outstanding errors (0 total)
• eslint, run over the whole population rather than narrowed: eslint --no-inline-config --format json . resolved 4030 files from its own config; both changed files report 0 errors, 0 warnings. The run's own non-zero exit reflects 90 pre-existing errors in files outside the per-package lint scope, unchanged by this diff and confirmed green by the coverage gate above.

Generated by Claude Code

---

Generated by Claude Code

[claude]

Fixes#6666

A consumer that imports a locale pack object and reads it by property access is invisible to both of the sweep's legs at once. The AST pass classifies t()/tt() call sites and there is no call; the text safety net greps the full dotted key and the namespace segment is bound to a local variable, so the dotted key is never spelled. The key lands in CONFIRMED — the tier documented as the safest thing to delete — while a shipping screen renders it.

textFootprint() now probes each candidate's property chain (the key minus its leading namespace segment, leading dot kept) alongside the full dotted key, in the same single grep -F pass.

The card's tier claim is stale — measured, and stated rather than repeated

The card measured the five keys at CONFIRMED on 813bf8323. On origin/main at 4357ec754 they are NEEDS-REVIEW, demoted by side effect: PR #6662's retirement pin and ar-participle-orthography-6610.test.ts both spell the dotted keys, which gives them a textual footprint. Per the triage ruling that is a mitigation, not a fix — the defect is unchanged, and the before/after evidence below is therefore built on what the hits are rather than on a tier flip.

	before (4357ec754)	after (fb4ccd348)
CONFIRMED	115	115
NEEDS-REVIEW	247	247
candidates	362	362
hits naming the real reader LoadingScreen.tsx	0	7 keys

Nothing was demoted; no pre-existing hit was lost. What changed is that seven keys now name the file that actually renders them, marked (via property chain):

• the five from the card, and
• console.actions.retry / console.actions.retrying — two more the card did not list, read by the same file the same way (LoadingScreen.tsx:118,123).

Both negative controls stay CONFIRMED with an empty hit list.

Three boundaries, each pinned by an ablation

1. Three segments or more, never two. A two-segment key's chain is a single generic word (.ok, .no, .empty). propertyChainProbe() returns null below three segments. Two-segment keys stay on the enumerated importer list, read by hand — which is what refactor(i18n): retire 25 confirmed-dead locale keys across all ten packs #6662 did.
2. The match must end at a property boundary. Otherwise a leaf would demote on a reader of a longer sibling leaf that it merely prefixes.
3. A chain-only hit is suffixed (via property chain), so the report never sends a reader grepping for a dotted key that file does not contain.

The importer list is written into the file

Per the ruling, the pack-object importer class is enumerated in the script header with the two-stage grep that re-derives it. The outboundAgentText.ts caveat is stated explicitly: it indexes its subtree dynamically (ai?.[key]), so neither leg sees the read — its four keys land in NEEDS-REVIEW only because those property names happen to appear as string literals in the OutboundAgentTextKey union above them. By luck, not by design. Replace that union with anything generated and all four drop to CONFIRMED with a live consumer still reading them.

Self-pollution, and why the tests spell keys in segments

textFootprint() greps scripts/ too, so a dotted key written into the test file would make that file a hit for it — the negative controls would stop being reader-less because the test asserting they are reader-less mentioned them. The real-repo controls are therefore joined from segments, and this was verified empirically: with the new test file on disk, console.error.serverUnreachable and console.error.timeout are still CONFIRMED and CONFIRMED is still 115.

Verification — all on fb4ccd348

Whole scripts/__tests__ tree (not a subject-matter subset): Test Files 89 passed (89) · Tests 2512 passed (2512).

Ablations. No build step is involved — vitest loads scripts/*.mjs from source, there is no dist/ for a stale artifact to hide in. Each leg mutated with perl -0pi, the mutation proved on disk by counting the injected and removed text, restored with git checkout HEAD -- against an absolute path, and the restore proved byte-identical (git hash-object equal to the HEAD blob 851663da3 and git diff HEAD empty) before the next reading:

ablation	result
probe always returns null (leg removed)	Tests 4 failed, 23 passed — all four detection assertions red, every negative control still green
three-segment guard relaxed to two	Tests 2 failed, 25 passed
property-boundary check replaced by plain includes	Tests 1 failed, 26 passed

The first row is the discrimination proof: removing the leg loses detection and touches nothing on the negative side.

Gates, each exit code captured before any pipe:

• pnpm type-check:scripts — exit 0
• pnpm check:control-bytes — check-control-bytes: OK (scanned 5804 tracked text file(s); skipped 85 binary)
• pnpm check:shell-escape-residue — OK (4/4 root(s) resolved ...)
• pnpm check:vi-mock-specifiers — OK (4032 tracked source file(s) ...)
• pnpm check:esm-specifiers — exit 0
• node scripts/check-changeset-presence.mjs — No source of a released package changed in this range, so no changeset is owed.No changeset and no label are owed: the guarded surface is derived from published packages and scripts/ is not one.
• node scripts/check-lint-coverage.mjs — lint coverage: 46/46 packages linted, 0 with outstanding errors (0 total)
• eslint, run over the whole population rather than narrowed: eslint --no-inline-config --format json . resolved 4030 files from its own config; both changed files report 0 errors, 0 warnings. The run's own non-zero exit reflects 90 pre-existing errors in files outside the per-package lint scope, unchanged by this diff and confirmed green by the coverage gate above.

Generated by Claude Code

---

Generated by Claude Code

[claude]

Fixes#6666

A consumer that imports a locale pack object and reads it by property access is invisible to both of the sweep's legs at once. The AST pass classifies t()/tt() call sites and there is no call; the text safety net greps the full dotted key and the namespace segment is bound to a local variable, so the dotted key is never spelled. The key lands in CONFIRMED — the tier documented as the safest thing to delete — while a shipping screen renders it.

textFootprint() now probes each candidate's property chain (the key minus its leading namespace segment, leading dot kept) alongside the full dotted key, in the same single grep -F pass.

The card's tier claim is stale — measured, and stated rather than repeated

The card measured the five keys at CONFIRMED on 813bf8323. On origin/main at 4357ec754 they are NEEDS-REVIEW, demoted by side effect: PR #6662's retirement pin and ar-participle-orthography-6610.test.ts both spell the dotted keys, which gives them a textual footprint. Per the triage ruling that is a mitigation, not a fix — the defect is unchanged, and the before/after evidence below is therefore built on what the hits are rather than on a tier flip.

	before (4357ec754)	after (fb4ccd348)
CONFIRMED	115	115
NEEDS-REVIEW	247	247
candidates	362	362
hits naming the real reader LoadingScreen.tsx	0	7 keys

Nothing was demoted; no pre-existing hit was lost. What changed is that seven keys now name the file that actually renders them, marked (via property chain):

• the five from the card, and
• console.actions.retry / console.actions.retrying — two more the card did not list, read by the same file the same way (LoadingScreen.tsx:118,123).

Both negative controls stay CONFIRMED with an empty hit list.

Three boundaries, each pinned by an ablation

1. Three segments or more, never two. A two-segment key's chain is a single generic word (.ok, .no, .empty). propertyChainProbe() returns null below three segments. Two-segment keys stay on the enumerated importer list, read by hand — which is what refactor(i18n): retire 25 confirmed-dead locale keys across all ten packs #6662 did.
2. The match must end at a property boundary. Otherwise a leaf would demote on a reader of a longer sibling leaf that it merely prefixes.
3. A chain-only hit is suffixed (via property chain), so the report never sends a reader grepping for a dotted key that file does not contain.

The importer list is written into the file

Per the ruling, the pack-object importer class is enumerated in the script header with the two-stage grep that re-derives it. The outboundAgentText.ts caveat is stated explicitly: it indexes its subtree dynamically (ai?.[key]), so neither leg sees the read — its four keys land in NEEDS-REVIEW only because those property names happen to appear as string literals in the OutboundAgentTextKey union above them. By luck, not by design. Replace that union with anything generated and all four drop to CONFIRMED with a live consumer still reading them.

Self-pollution, and why the tests spell keys in segments

textFootprint() greps scripts/ too, so a dotted key written into the test file would make that file a hit for it — the negative controls would stop being reader-less because the test asserting they are reader-less mentioned them. The real-repo controls are therefore joined from segments, and this was verified empirically: with the new test file on disk, console.error.serverUnreachable and console.error.timeout are still CONFIRMED and CONFIRMED is still 115.

Verification — all on fb4ccd348

Whole scripts/__tests__ tree (not a subject-matter subset): Test Files 89 passed (89) · Tests 2512 passed (2512).

Ablations. No build step is involved — vitest loads scripts/*.mjs from source, there is no dist/ for a stale artifact to hide in. Each leg mutated with perl -0pi, the mutation proved on disk by counting the injected and removed text, restored with git checkout HEAD -- against an absolute path, and the restore proved byte-identical (git hash-object equal to the HEAD blob 851663da3 and git diff HEAD empty) before the next reading:

ablation	result
probe always returns null (leg removed)	Tests 4 failed, 23 passed — all four detection assertions red, every negative control still green
three-segment guard relaxed to two	Tests 2 failed, 25 passed
property-boundary check replaced by plain includes	Tests 1 failed, 26 passed

The first row is the discrimination proof: removing the leg loses detection and touches nothing on the negative side.

Gates, each exit code captured before any pipe:

• pnpm type-check:scripts — exit 0
• pnpm check:control-bytes — check-control-bytes: OK (scanned 5804 tracked text file(s); skipped 85 binary)
• pnpm check:shell-escape-residue — OK (4/4 root(s) resolved ...)
• pnpm check:vi-mock-specifiers — OK (4032 tracked source file(s) ...)
• pnpm check:esm-specifiers — exit 0
• node scripts/check-changeset-presence.mjs — No source of a released package changed in this range, so no changeset is owed.No changeset and no label are owed: the guarded surface is derived from published packages and scripts/ is not one.
• node scripts/check-lint-coverage.mjs — lint coverage: 46/46 packages linted, 0 with outstanding errors (0 total)
• eslint, run over the whole population rather than narrowed: eslint --no-inline-config --format json . resolved 4030 files from its own config; both changed files report 0 errors, 0 warnings. The run's own non-zero exit reflects 90 pre-existing errors in files outside the per-package lint scope, unchanged by this diff and confirmed green by the coverage gate above.

Generated by Claude Code

---

Generated by Claude Code

[claude]

Fixes#6666

A consumer that imports a locale pack object and reads it by property access is invisible to both of the sweep's legs at once. The AST pass classifies t()/tt() call sites and there is no call; the text safety net greps the full dotted key and the namespace segment is bound to a local variable, so the dotted key is never spelled. The key lands in CONFIRMED — the tier documented as the safest thing to delete — while a shipping screen renders it.

textFootprint() now probes each candidate's property chain (the key minus its leading namespace segment, leading dot kept) alongside the full dotted key, in the same single grep -F pass.

The card's tier claim is stale — measured, and stated rather than repeated

The card measured the five keys at CONFIRMED on 813bf8323. On origin/main at 4357ec754 they are NEEDS-REVIEW, demoted by side effect: PR #6662's retirement pin and ar-participle-orthography-6610.test.ts both spell the dotted keys, which gives them a textual footprint. Per the triage ruling that is a mitigation, not a fix — the defect is unchanged, and the before/after evidence below is therefore built on what the hits are rather than on a tier flip.

	before (4357ec754)	after (fb4ccd348)
CONFIRMED	115	115
NEEDS-REVIEW	247	247
candidates	362	362
hits naming the real reader LoadingScreen.tsx	0	7 keys

Nothing was demoted; no pre-existing hit was lost. What changed is that seven keys now name the file that actually renders them, marked (via property chain):

• the five from the card, and
• console.actions.retry / console.actions.retrying — two more the card did not list, read by the same file the same way (LoadingScreen.tsx:118,123).

Both negative controls stay CONFIRMED with an empty hit list.

Three boundaries, each pinned by an ablation

1. Three segments or more, never two. A two-segment key's chain is a single generic word (.ok, .no, .empty). propertyChainProbe() returns null below three segments. Two-segment keys stay on the enumerated importer list, read by hand — which is what refactor(i18n): retire 25 confirmed-dead locale keys across all ten packs #6662 did.
2. The match must end at a property boundary. Otherwise a leaf would demote on a reader of a longer sibling leaf that it merely prefixes.
3. A chain-only hit is suffixed (via property chain), so the report never sends a reader grepping for a dotted key that file does not contain.

The importer list is written into the file

Per the ruling, the pack-object importer class is enumerated in the script header with the two-stage grep that re-derives it. The outboundAgentText.ts caveat is stated explicitly: it indexes its subtree dynamically (ai?.[key]), so neither leg sees the read — its four keys land in NEEDS-REVIEW only because those property names happen to appear as string literals in the OutboundAgentTextKey union above them. By luck, not by design. Replace that union with anything generated and all four drop to CONFIRMED with a live consumer still reading them.

Self-pollution, and why the tests spell keys in segments

textFootprint() greps scripts/ too, so a dotted key written into the test file would make that file a hit for it — the negative controls would stop being reader-less because the test asserting they are reader-less mentioned them. The real-repo controls are therefore joined from segments, and this was verified empirically: with the new test file on disk, console.error.serverUnreachable and console.error.timeout are still CONFIRMED and CONFIRMED is still 115.

Verification — all on fb4ccd348

Whole scripts/__tests__ tree (not a subject-matter subset): Test Files 89 passed (89) · Tests 2512 passed (2512).

Ablations. No build step is involved — vitest loads scripts/*.mjs from source, there is no dist/ for a stale artifact to hide in. Each leg mutated with perl -0pi, the mutation proved on disk by counting the injected and removed text, restored with git checkout HEAD -- against an absolute path, and the restore proved byte-identical (git hash-object equal to the HEAD blob 851663da3 and git diff HEAD empty) before the next reading:

ablation	result
probe always returns null (leg removed)	Tests 4 failed, 23 passed — all four detection assertions red, every negative control still green
three-segment guard relaxed to two	Tests 2 failed, 25 passed
property-boundary check replaced by plain includes	Tests 1 failed, 26 passed

The first row is the discrimination proof: removing the leg loses detection and touches nothing on the negative side.

Gates, each exit code captured before any pipe:

• pnpm type-check:scripts — exit 0
• pnpm check:control-bytes — check-control-bytes: OK (scanned 5804 tracked text file(s); skipped 85 binary)
• pnpm check:shell-escape-residue — OK (4/4 root(s) resolved ...)
• pnpm check:vi-mock-specifiers — OK (4032 tracked source file(s) ...)
• pnpm check:esm-specifiers — exit 0
• node scripts/check-changeset-presence.mjs — No source of a released package changed in this range, so no changeset is owed.No changeset and no label are owed: the guarded surface is derived from published packages and scripts/ is not one.
• node scripts/check-lint-coverage.mjs — lint coverage: 46/46 packages linted, 0 with outstanding errors (0 total)
• eslint, run over the whole population rather than narrowed: eslint --no-inline-config --format json . resolved 4030 files from its own config; both changed files report 0 errors, 0 warnings. The run's own non-zero exit reflects 90 pre-existing errors in files outside the per-package lint scope, unchanged by this diff and confirmed green by the coverage gate above.

Generated by Claude Code

---

Generated by Claude Code

[claude]

Fixes#6666

A consumer that imports a locale pack object and reads it by property access is invisible to both of the sweep's legs at once. The AST pass classifies t()/tt() call sites and there is no call; the text safety net greps the full dotted key and the namespace segment is bound to a local variable, so the dotted key is never spelled. The key lands in CONFIRMED — the tier documented as the safest thing to delete — while a shipping screen renders it.

textFootprint() now probes each candidate's property chain (the key minus its leading namespace segment, leading dot kept) alongside the full dotted key, in the same single grep -F pass.

The card's tier claim is stale — measured, and stated rather than repeated

The card measured the five keys at CONFIRMED on 813bf8323. On origin/main at 4357ec754 they are NEEDS-REVIEW, demoted by side effect: PR #6662's retirement pin and ar-participle-orthography-6610.test.ts both spell the dotted keys, which gives them a textual footprint. Per the triage ruling that is a mitigation, not a fix — the defect is unchanged, and the before/after evidence below is therefore built on what the hits are rather than on a tier flip.

	before (4357ec754)	after (fb4ccd348)
CONFIRMED	115	115
NEEDS-REVIEW	247	247
candidates	362	362
hits naming the real reader LoadingScreen.tsx	0	7 keys

Nothing was demoted; no pre-existing hit was lost. What changed is that seven keys now name the file that actually renders them, marked (via property chain):

• the five from the card, and
• console.actions.retry / console.actions.retrying — two more the card did not list, read by the same file the same way (LoadingScreen.tsx:118,123).

Both negative controls stay CONFIRMED with an empty hit list.

Three boundaries, each pinned by an ablation

1. Three segments or more, never two. A two-segment key's chain is a single generic word (.ok, .no, .empty). propertyChainProbe() returns null below three segments. Two-segment keys stay on the enumerated importer list, read by hand — which is what refactor(i18n): retire 25 confirmed-dead locale keys across all ten packs #6662 did.
2. The match must end at a property boundary. Otherwise a leaf would demote on a reader of a longer sibling leaf that it merely prefixes.
3. A chain-only hit is suffixed (via property chain), so the report never sends a reader grepping for a dotted key that file does not contain.

The importer list is written into the file

Per the ruling, the pack-object importer class is enumerated in the script header with the two-stage grep that re-derives it. The outboundAgentText.ts caveat is stated explicitly: it indexes its subtree dynamically (ai?.[key]), so neither leg sees the read — its four keys land in NEEDS-REVIEW only because those property names happen to appear as string literals in the OutboundAgentTextKey union above them. By luck, not by design. Replace that union with anything generated and all four drop to CONFIRMED with a live consumer still reading them.

Self-pollution, and why the tests spell keys in segments

textFootprint() greps scripts/ too, so a dotted key written into the test file would make that file a hit for it — the negative controls would stop being reader-less because the test asserting they are reader-less mentioned them. The real-repo controls are therefore joined from segments, and this was verified empirically: with the new test file on disk, console.error.serverUnreachable and console.error.timeout are still CONFIRMED and CONFIRMED is still 115.

Verification — all on fb4ccd348

Whole scripts/__tests__ tree (not a subject-matter subset): Test Files 89 passed (89) · Tests 2512 passed (2512).

Ablations. No build step is involved — vitest loads scripts/*.mjs from source, there is no dist/ for a stale artifact to hide in. Each leg mutated with perl -0pi, the mutation proved on disk by counting the injected and removed text, restored with git checkout HEAD -- against an absolute path, and the restore proved byte-identical (git hash-object equal to the HEAD blob 851663da3 and git diff HEAD empty) before the next reading:

ablation	result
probe always returns null (leg removed)	Tests 4 failed, 23 passed — all four detection assertions red, every negative control still green
three-segment guard relaxed to two	Tests 2 failed, 25 passed
property-boundary check replaced by plain includes	Tests 1 failed, 26 passed

The first row is the discrimination proof: removing the leg loses detection and touches nothing on the negative side.

Gates, each exit code captured before any pipe:

• pnpm type-check:scripts — exit 0
• pnpm check:control-bytes — check-control-bytes: OK (scanned 5804 tracked text file(s); skipped 85 binary)
• pnpm check:shell-escape-residue — OK (4/4 root(s) resolved ...)
• pnpm check:vi-mock-specifiers — OK (4032 tracked source file(s) ...)
• pnpm check:esm-specifiers — exit 0
• node scripts/check-changeset-presence.mjs — No source of a released package changed in this range, so no changeset is owed.No changeset and no label are owed: the guarded surface is derived from published packages and scripts/ is not one.
• node scripts/check-lint-coverage.mjs — lint coverage: 46/46 packages linted, 0 with outstanding errors (0 total)
• eslint, run over the whole population rather than narrowed: eslint --no-inline-config --format json . resolved 4030 files from its own config; both changed files report 0 errors, 0 warnings. The run's own non-zero exit reflects 90 pre-existing errors in files outside the per-package lint scope, unchanged by this diff and confirmed green by the coverage gate above.

Generated by Claude Code

---

Generated by Claude Code