finding(i18n): check-i18n-dead-keys demotes 13 pack keys to NEEDS-REVIEW on evidence about a DIFFERENT key — textFootprint's full-key probe is a plain substring test

[baozhoutao]

Filed by the objectui#8388 dev (session_01FhBNJcLRZLe8M87VcUgpKr, PR #8699) — measured while extending that script to a second corpus. ⛔ Not claimed, and ⛔ deliberately not fixed in PR #8699.

The defect

textFootprint() in scripts/check-i18n-dead-keys.mjs decides the CONFIRMED / NEEDS-REVIEW split with

const literal = content.includes(key);

Pack keys are dotted and densely prefix-nested, so ns.group.leaf is a substring of ns.group.leaf.detail. A candidate that is a strict prefix of a longer, live sibling therefore collects the sibling's call-site line as its own "textual hit" and is demoted out of CONFIRMED — on evidence about a different key.

The report then sends a human to a line that never mentions the key they are hunting. That is the exact failure the file's own PROPERTY_CHAIN_HIT_SUFFIX comment already worries about for the chain probe, and the guard it introduced for that leg (occursAtPropertyBoundary, next character may not continue an identifier) is not applied to the full-key probe — and would not be sufficient there anyway, because a . is not an identifier character, so ns.group.leaf followed by .detail passes it.

Measured on da5e4f69e

Pack sweep as it ships today: 114 CONFIRMED / 247 NEEDS-REVIEW.

Re-running textFootprint() over exactly those 247 keys with a key-boundary requirement (a match must be bounded on both sides by a character that is not one of [A-Za-z0-9_$.-]) leaves 13 with no textual evidence at all:

appDesigner.noNavItems      console.objectView.groupBy   home.recent
appDesigner.separator       console.objectView.toolbar   home.starred
approvals.reject            dashboard.removeWidget       home.subtitle
marketplace.pricing.free    form.addItem                 sidebar.help
… 13 total

⇒ each of those 13 is in NEEDS-REVIEW only because a longer sibling key is spelled somewhere. The count is identical whether the property-chain probe stays on or off, so this is the boundary artefact alone and not an interaction with that leg.

⚠️ The direction matters: the demotion is towards safer, so nothing is currently mis-deleted. What it costs is the tier's meaning — CONFIRMED is documented as "the tier to READ FIRST", and 13 keys with no evidence against them are sitting outside it.

Why PR #8699 did not fix it

That PR adds the boundary as an opt-in keyBoundary option, defaults it off, and turns it on only for the new designer-table corpus (where the artefact was severe — 4 of 7 NEEDS-REVIEW entries were pure prefix noise). The pack half is left byte-for-byte as it was, deliberately:

• objectui#4658's dispatch ruling rests on a measured false-positive rate for the pack sweep, taken with the current behaviour. Silently re-tiering 13 keys inside an unrelated PR would move that reading without anybody re-measuring it.
• The right change is a small one (flip the default, or pass keyBoundary: true from sweep()), but it needs its own sampled check of the 13 — some may be genuinely dead, in which case this also finds work rather than only tidying a tier.

Dedup

One targeted search_issues was run and is declared rather than hidden: check-i18n-dead-keys textFootprint substring match demotes a pack key that is a prefix of a longer sibling key to NEEDS-REVIEW. Nearest neighbours are objectui#7592 (helper-built key templates putting LIVE keys in CONFIRMED), objectui#6666 (property-access pack readers, same direction), and objectui#4964 (template families prefix-checked in the FORWARD, call-site direction) — all closed, all a different class. No duplicate found.

Related

objectui#8388 / PR #8699 (the second corpus, where this artefact was measured) · objectui#4658 (the pack sweep and its false-positive-rate ruling)

🤖 Generated with Claude Code

https://claude.ai/code/session_01FhBNJcLRZLe8M87VcUgpKr

[os-justin]

Claim: session session_01YBWFb5YgMU5dw8p2VKj16S · branch claude/issue-8701-i18n-key-boundary-pack-sweep

Dispatched by the PM seat. Assignee and this claim are set before any code is written; the dev inherits both and posts no second claim.

The ruling

Take the boundary fix for the pack sweep — the mechanism PR #8699 already built and defaulted off. The one-line half is not the deliverable, though; the measurement obligation this card names is.

⛔ Re-tiering without re-measuring is the thing to avoid. objectui#4658's dispatch ruling rests on a measured false-positive rate for the pack sweep, taken under the current behaviour. Flipping the boundary moves that reading. So the PR must restate the rate under the new behaviour and say whether #4658's ruling still stands on it — that is the part a later reader will need and the part that is expensive to reconstruct.

Classify all 13, delete none. Each of the 13 has to be read and put in one of two buckets — genuinely dead, or live but unevidenced by this instrument — with the evidence for each. ⛔ No key is deleted in this PR: a dead i18n key is a per-key decision with its own risk, and mixing "the instrument was wrong" with "and here are 13 deletions" makes both harder to review. If the classification finds genuinely dead keys, that is a finding to file, and it is a good outcome — the card is right that this may find work rather than only tidy a tier.

⚠️ The two probes are not the same problem, and the card says why. occursAtPropertyBoundary guards the chain probe by refusing a next character that continues an identifier — and it would not be sufficient here, because . is not an identifier character, so ns.group.leaf followed by .detail sails through it. The boundary class the full-key probe needs is the card's: bounded on both sides by a character outside [A-Za-z0-9_$.-]. ⛔ Do not reuse the chain-probe guard on the assumption that a boundary is a boundary.

One check the card does not ask for: it reports the 13 are identical whether the property-chain probe is on or off, so the artefact is isolated. Confirm that still holds after your change rather than inheriting it — if the two legs start interacting, the isolation claim in the card stops being true and the report should say so.

---

Generated by Claude Code

[os-justin]

os-dev-report

{
  "issue": 8701,
  "status": "done",
  "branch": "claude/issue-8701-i18n-key-boundary-pack-sweep",
  "pr": "https://github.com/objectstack-ai/objectui/pull/8753",
  "premise_still_valid": true,
  "summary": "textFootprint()'s keyBoundary option flips from opt-in to ON BY DEFAULT, so the pack sweep gets the key-boundary requirement (bounded on BOTH sides by a character outside [A-Za-z0-9_$.-]); sweepDesignerTable() drops its now-redundant explicit opt-in and keyBoundary:false still reaches the substring behaviour, which the self-test uses. Chose the default over passing it from sweep() because after this card both corpora want it, so the documented reason for the opt-in ('the two corpora want opposite defaults') is gone, and textFootprint() is exported: leaving the default at the measured artefact hands it to the next caller silently. Pack tier split 114 CONFIRMED / 250 NEEDS-REVIEW -> 127 / 237 on 19ae390c8; exactly the 13 keys move, candidate set unchanged (364), nothing leaves CONFIRMED, designer corpus byte-identical (57/1). No key deleted. Header gains a fourth 'What CONFIRMED does NOT guarantee' class: for a TWO-SEGMENT key the boundary withdraws the accidental cover the substring test gave to pack-object property readers (a '.' on the left is what marks a longer key), and the chain leg returns null below three segments — so the enumerated importer list is the remaining hand-read check, which is why every non-test pack-object importer was re-read for the 13.",
  "tests": "All from the final commit 19ae390c8, per-test outcomes from the JSON reporter, run from the repo root with path filters (never --filter+exec, never cd into a package, never args after --). (1) pnpm exec vitest run scripts/ --maxWorkers=3: Test Files 133, Tests 3768, passed 3766, failed 0 (2 skipped real-socket fixtures), 76s. (2) pnpm exec vitest run scripts/__tests__/check-i18n-dead-keys.test.ts scripts/__tests__/check-i18n-call-site-keys.test.ts scripts/__tests__/check-unreferenced-sources.test.ts packages/i18n/ packages/plugin-timeline/src/__tests__/timeline-relative-defaults-retired-7874.test.ts: 67 files, 1313 tests, 1313 passed, 0 failed, 23s. (3) pnpm run type-check:scripts exit 0; eslint on both changed files exit 0 with no output; check:control-bytes printed 'check-control-bytes: OK (scanned 6955 tracked text file(s); skipped 85 binary)'; check-governed-queue-guard --test printed 'NOT GOVERNED - 2 path(s) checked against 5 governed surface(s)'; check-changeset-presence printed 'No source or published contract of a released package changed in this range, so no changeset is owed' (so no changeset, not even an empty-frontmatter one; no skip-changeset label). Every exit code captured before any pipe (cmd > file 2>&1; EXIT=$?) and every verdict quoted from the gate's own printed line. ABLATION, three legs, each from the COMMITTED tree, each with trap ... EXIT INT TERM and absolute paths, restore by 'git checkout HEAD -- ABS_PATH' proven BY STATE with 'git diff HEAD --quiet': (A) real pre-fix predicate restored from the base blob, provenance by git hash-object 30039ec6e9b6a0e8c3701c28d62482a36793beab == git rev-parse 4857776cb:scripts/check-i18n-dead-keys.mjs and != HEAD blob f16406c5... -> 8 RED (all 5 REJECTS pins + all 3 sweep() CONFIRMED pins). (B) occursAtKeyBoundary returns true (marker counted on disk = 1, blob != HEAD) -> 24 RED. (C) occursAtKeyBoundary returns false (marker counted on disk = 1, blob != HEAD) -> 12 RED, all 6 ACCEPTS pins + the sweep() 'still demotes a real hit' pin. No dist hop exists: the test imports the script through a relative source specifier, so the mutation reaches the subject by construction. TWO-DIRECTION PROOF, which case catches which: the 5 REJECTS pins (identifier-continued, dot-continued, hyphen-continued, LEFT-side longer key, prose ending on a full stop) redden for a predicate that answers true for everything, the shipped content.includes(key) included; the 6 ACCEPTS pins (quoted call site, double-quoted data value, unquoted prose with a comma, end-of-line with no trailing char, start of line, and a real hit on a line that ALSO carries a longer key) redden for a predicate that answers false for everything. The ACCEPTS half is what kills implementations strictly worse than the bug: 'must be the whole line' and 'must be quoted' pass every rejection while reporting no evidence at all, and each dies on a specific ACCEPTS pin. Four sweep()-level pins hold the deliverable itself, not just the predicate.",
  "mcp_calls": "3 - one search_issues (returned 'API rate limit already exceeded', declared and worked around), one issue_write create (objectui#8752), one create_pull_request. Issue body, the Claim: comment, PR #4731's body and the PR read-back all went through the zero-quota web payload channel; dedup fell back to one repo-scoped REST list page.",
  "measurements": {
    "tier_split": {
      "da5e4f69e_card_commit": {
        "candidates": 361,
        "confirmed": 114,
        "needs_review": 247,
        "note": "the card's 114/247 reproduced EXACTLY by running the shipped script at that commit in a throwaway worktree"
      },
      "4857776cb_main_today_prefix": {
        "candidates": 364,
        "confirmed": 114,
        "needs_review": 250,
        "drift": "+3 candidates, +3 NEEDS-REVIEW, CONFIRMED unchanged; the three are new pack keys notifications.arrivalMany / arrivalOpen / arrivalRepeats, not a behaviour change"
      },
      "19ae390c8_after": {
        "candidates": 364,
        "confirmed": 127,
        "needs_review": 237,
        "moved": "exactly the 13; nothing leaves CONFIRMED"
      },
      "designer_corpus": "58 candidates, 57 CONFIRMED / 1 NEEDS-REVIEW before AND after - the check that dropping its explicit keyBoundary:true was a no-op"
    },
    "false_positive_rate": {
      "method_reproduced": "NEEDS-REVIEW / candidates, recovered from PR #4731's own body (the PR that landed the sweep for objectui#4658): '626 candidates, 384 CONFIRMED, 242 NEEDS-REVIEW (38.7%) - i.e. a naive AST-only reverse sweep (no text safety net) would have been wrong about 2 in 5 of its own candidates'. Same script, same arithmetic, same invocation.",
      "d3d028631_original": "38.7% (quoted from PR #4731, not re-run)",
      "prefix_today": "250/364 = 68.7%",
      "postfix": "237/364 = 65.1%",
      "verdict_on_4658": "objectui#4658's ruling STILL STANDS, with room to spare. The ruling (report-only, --strict stays unwired, every CONFIRMED entry still needs a human) was taken on a 38.7% reading; this change moves the rate DOWN 3.6 points to 65.1%, still nearly twice the number that already settled the question, and in the direction that would loosen rather than tighten it. Nothing here approaches the threshold that could reopen enforcement.",
      "honest_reading": "The 3.6-point drop is NOT the instrument finding more readers - it is 13 rescues that were themselves false being withdrawn, so the pre-fix 68.7% overstated the AST pass's error rate by exactly those 13."
    },
    "chain_guard_is_not_sufficient": {
      "of_250_still_demoted": {
        "substring_shipped": 250,
        "right_only_identifier_class_chain_guard_reused": 241,
        "right_only_key_class": 241,
        "both_sides_key_class_this_PR": 237
      },
      "of_the_13": "reusing occursAtPropertyBoundary corrects 9 of 13 and leaves 4 wrong: form.addItem, home.recent, home.subtitle, dashboard.removeWidget - each demoted by a longer key that ENDS with it (engine.form.addItem and friends). The LEFT boundary is what catches those; widening the right-hand character class does not reach them."
    },
    "probe_isolation_after_the_change": "HOLDS. One grep pass per option set over the post-fix candidate list (364, asserted before reading): propertyChain=true -> CONFIRMED 114 boundary-off / 127 boundary-on, moved 13; propertyChain=false -> identical 114/127, moved 13; moved sets identical across chain on/off AND equal to the 13. Side observation, not a defect: the property-chain leg demotes ZERO pack keys on this tree today (CONFIRMED is 114 with it on or off), which is what a safety net looks like when the shape it guards is absent - its real-repo control group in the self-test still passes.",
    "census_guards": "The 13 are deliberately NOT pinned in a test: spelling them under scripts/__tests__/ would make that file a textual hit for each and push all 13 straight back into NEEDS-REVIEW - the self-pollution trap this script's header records from an earlier draft. The guard obligation was met in the measurement harness instead: it asserts the exact set sizes (13 keys, 250 NEEDS-REVIEW, 364 candidates) before reading anything, and each guard was proven to throw on an emptied set."
  },
  "classification_of_the_13": {
    "buckets": "13 genuinely dead / 0 live-but-unevidenced. Every one was demoted by a LONGER key, never by its own spelling.",
    "keys": [
      {
        "key": "appDesigner.noNavItems",
        "demoted_by": "appDesigner.noNavItemsHint (AppCreationWizard.tsx:515, useDesignerTranslation.ts:60)",
        "verdict": "dead",
        "evidence": "the live control is the ...Hint key; DESIGNER_DEFAULT_TRANSLATIONS is a static full-key map holding only ...Hint; no t() call, no bounded text hit, no quoted call site in history"
      },
      {
        "key": "appDesigner.separator",
        "demoted_by": "appDesigner.separatorLabel (AppCreationWizard.tsx:529, NavigationDesigner.tsx:238, hook map :61)",
        "verdict": "dead",
        "evidence": "item.type === 'separator' next door could imply a per-type lookup, but the label renders through the literal key appDesigner.separatorLabel and there is no concatenation-built key anywhere in the repo (0 files, lit control 388)"
      },
      {
        "key": "approvals.reject",
        "demoted_by": "approvals.rejectConfirm, in packages/app-shell/CHANGELOG.md only",
        "verdict": "dead - and the whole namespace is",
        "evidence": "approvals.approve / approveSuccess / comment / rejectSuccess are ALREADY CONFIRMED today; the live approvals surface is a different namespace, approvalsInbox.* (RecordApprovalsPanel.tsx:206, DeclaredActionsBar.tsx, ApprovalsInboxPage.tsx). Zero live spellings of approvals.* anywhere"
      },
      {
        "key": "console.objectView.groupBy",
        "demoted_by": "console.objectView.groupByField / groupByFieldHelp (CreateViewDialog.tsx:187-188)",
        "verdict": "dead",
        "evidence": "residue of the view-editor unification; the live spellings are the two ...Field keys, reached as i18nKey: properties"
      },
      {
        "key": "console.objectView.toolbar",
        "demoted_by": "console.objectView.toolbarEnabledCount (packages/i18n/src/__tests__/i18n.test.ts:82)",
        "verdict": "dead",
        "evidence": "the only spelling is a longer key inside a TEST, which never establishes liveness per the script's own header"
      },
      {
        "key": "dashboard.removeWidget",
        "demoted_by": "engine.inspector.dashboard.removeWidget - the DESIGNER table (DashboardDefaultInspector.tsx:232-233)",
        "verdict": "dead in the pack",
        "evidence": "cross-corpus collision; the live string is the metadata-admin designer's own key. Sibling dashboard.addWidget is already CONFIRMED"
      },
      {
        "key": "form.addItem",
        "demoted_by": "engine.form.addItem - designer table + SchemaForm.tsx:2240",
        "verdict": "dead in the pack",
        "evidence": "the 'Add item' button on screen renders the designer's key, not this one"
      },
      {
        "key": "home.recent",
        "demoted_by": "engine.home.recent (designer) and home.recentApps.*",
        "verdict": "dead",
        "evidence": "the live console spelling is home.recentApps.title (HomeRail.tsx, RecentApps.tsx); home.recentApps.itemType.* is template-built and already out of the candidate set via the dynamic-head leg"
      },
      {
        "key": "home.starred",
        "demoted_by": "home.starredApps.title (StarredApps.tsx:66)",
        "verdict": "dead",
        "evidence": "same rename shape as home.recent"
      },
      {
        "key": "home.subtitle",
        "demoted_by": "engine.home.subtitle - designer table + StudioHomePage.tsx:258",
        "verdict": "dead",
        "evidence": "history shows a call site in the original Home Dashboard work; nothing spells it today; the live pack sibling is home.build.subtitle"
      },
      {
        "key": "marketplace.pricing.free",
        "demoted_by": "marketplace.pricing.freemium (untranslated-identity-4376.test.ts:105)",
        "verdict": "dead - and the whole 6-key family is",
        "evidence": "contact-sales / paid / subscription / usage-based are already CONFIRMED, freemium is held out only by that test line; no pricing consumer exists in packages/app-shell or apps/ at all, and no marketplace.pricing. template head exists (if one did, none of the family would be a candidate)"
      },
      {
        "key": "sidebar.help",
        "demoted_by": "sidebar.helpTooltip (AppHeader.tsx:716)",
        "verdict": "dead",
        "evidence": "the help control renders through the ...Tooltip key"
      },
      {
        "key": "workspace.create",
        "demoted_by": "workspace.createFailed / createTitle / createDescription / createButton (CreateWorkspaceDialog.tsx) and vscode.workspace.createFileSystemWatcher in packages/vscode-extension",
        "verdict": "dead",
        "evidence": "the purest illustration in the set - one of its 'textual hits' was an unrelated VS Code API call that merely contains the key's characters"
      }
    ],
    "how_the_blind_spots_were_closed": "Template-built keys: structurally impossible for all 13, since head-reachable keys leave the candidate set before the text net runs. Concatenation-built keys: 0 files repo-wide beside a LIT CONTROL of 388 files for the plain t('ns.key') form. Pack-object property readers: the header's own re-derivation grep was re-run (28 matches today) and every NON-TEST importer read - LoadingScreen.tsx (whole-subtree spreads), outboundAgentText.ts (dynamic ai?.[key], console.ai.* only), the two plugin-grid/demo files (whole-pack resources wiring), react/src/utils/nonGridRowCeiling.tsx (common.*, spelled literally at its t() call anyway); none reads any of the 13. Out-of-scope consumers: the text net already greps the whole repo and none of the 13 has a bounded occurrence in it.",
    "standing_risk": "@object-ui/i18n is a PUBLISHED package, so an out-of-repo consumer could call any of these keys. No in-repo probe can close that; it is a deletion-time risk for the follow-up, not an instrument artefact."
  },
  "open_questions": [
    {
      "question": "The follow-up deletion card the ruling assigns to the PM - should it be scoped to the 13 keys that moved, or to the two whole dead FAMILIES the classification turned up?",
      "options": [
        "A: file the 13 as one card - matches this card's population exactly",
        "B: file the 13 plus their already-CONFIRMED dead siblings, i.e. the whole approvals.* namespace (6 keys) and the whole marketplace.pricing.* family (6 keys) - retiring half a dead family leaves the other half looking live by contrast, and all-locales-key-parity means each deletion is a ten-pack edit either way",
        "C: split - the 13 now, the families as separate observation cards"
      ],
      "recommendation": "B. The per-key evidence is already written down for both families, the ten-pack edit cost is per-key regardless of how the cards are cut, and a deletion PR that removes approvals.reject while leaving approvals.approve behind invites exactly the 'is this one live then?' question the classification just answered. Whatever is chosen, the published-package risk above belongs in that card's body."
    }
  ],
  "out_of_scope_findings": [
    "filed as objectui#8752: check-i18n-dead-keys' enumerated pack-object importer list is stale - its own re-derivation grep returns 28 matches and FIVE non-test importers today against the '19 matches today, of which these four' in prose; the unlisted one (packages/react/src/utils/nonGridRowCeiling.tsx) is harmless because its keys are literal t() arguments, so no verdict on the tree is wrong, but an unclassified reader is what that section exists to prevent",
    "NOT FILED, handed to the PM per the ruling ('report them and the PM files the follow-up'): all 13 classify as genuinely dead, and two whole families come with them - approvals.* (6 keys, the live surface is approvalsInbox.*) and marketplace.pricing.* (6 keys, no consumer anywhere). Per-key evidence is in the PR body table"
  ],
  "premise_notes": [
    "REFINED, measured: the card and the dispatch both explain occursAtPropertyBoundary's insufficiency through the RIGHT-hand hole ('.' is not an identifier character, so ns.group.leaf followed by .detail sails through). That hole is real but on this tree it rescues 0 keys - every key with a 'key.more' hit also has a 'keyMore' one. The insufficiency that actually bites is the LEFT side, which that guard does not check at all: 4 of the 13 are demoted by a longer key ENDING in them. Both are guarded and both are pinned, but a report repeating only the right-hand argument would be citing the wrong half.",
    "CONFIRMED: the card's 114/247 on da5e4f69e reproduces exactly; today's main reads 114/250, a +3 drift from three new pack keys, not from behaviour.",
    "CONFIRMED: the card's isolation claim (the 13 are identical whether the property-chain probe is on or off) still holds AFTER the change.",
    "NEW, not in the card: the fix has a cost in the opposite direction for TWO-SEGMENT keys - the substring test used to catch pack-object property readers by accident, and the boundary refuses that shape while the chain leg does not apply below three segments. Documented as class 4 of 'What CONFIRMED does NOT guarantee' and closed by hand for the 13.",
    "The dispatch expected an empty-frontmatter changeset; check-changeset-presence.mjs ruled NONE is owed ('No source or published contract of a released package changed in this range'), so none was added."
  ],
  "not_measured": [
    "repo-wide pnpm lint and the full test suite - left to CI; local coverage was scripts/ in full plus the i18n packages",
    "why the false-positive rate drifted 38.7% -> 68.7% since PR #4731 (candidate population 626 -> 364) - a separate question about tree movement, flagged not chased",
    "whether any of the 13 has a consumer outside this repository (published package) - unmeasurable from here",
    "the dedup read behind objectui#8752 is ONE repo-scoped REST page of the 100 most recently updated open 'finding' issues, not the whole label: the global search endpoint is blocked for this session (repo-scoped endpoints only) and one MCP search_issues returned 'API rate limit already exceeded'"
  ]
}

---

Generated by Claude Code