[finding] The tombstone judge converged, but the Zod-internals READERS around it are still hand-copied in ~10 test files — four of them verbatim

[os-zhuang]

⚠️ MEASURED STATE — 2026-08-31. Read this before the original filing below.

The census in the original body is STALE and has now caused one wasted full dispatch. Do not work from it.

Class (1) — ✅ DONE, landed

PR #6047 MERGED 2026-08-24T15:00:26Z (merged: true, head 3b010122f; squash subject test(test-support): confine the spec enum-vocabulary reader to one module (#5872 class 1)). The four verbatim enum-options copies are consolidated onto packages/test-support/src/spec-enum-options.ts (shapeEnumOptions), and all three consuming packages already declare the devDependencies edge.

⛔ The 2026-08-25 dead-claim reclaim on this card was a FALSE NEGATIVE. It recorded "no merged commit referencing this card across 711 commits of main" 17 hours after the merge it failed to see, and a 2026-08-31 re-dispatch acted on it. Root cause, measured: main was re-rooted on 2026-08-30 — HEAD history is 28 commits ending at a parentless root (335523906) — so git log over main can no longer see any pre-re-root merge, and PR #6047's head is an ancestor of nothing.

⚠️ On this repo, a "N commits of main" scan is not evidence of absence. Use blob identity against the fetched PR ref (git rev-parse <ref>:<path> per changed file), or simply read the PR's merged flag. Confirmation for #6047: of its 13 changed files, 11 byte-identical to e3d117ae1, 0 absent; the 2 that differ moved only under later unrelated work (#6355's timeScale-alias retirement inside timeline-scale-spec-parity.test.ts's resolveTimelineScale block, and pnpm-lock.yaml).

Classes (2)–(4) — what is actually left

• Class (2) — array-element unwrapping, three DIFFERENT spellings for one question. ⛔ Not mechanical: a shared reader must pick a behaviour where three copies disagree, which is a genuine verdict-flip risk. Deserves its own round, measured before conversion.
• Class (3) — the wrapper-key literal list. ⛔ BLOCKED, needs a maintainer ruling — see the decision card linked below. It has grown out of tests into .mjs CI gate scripts, and @object-ui/test-support is private: true with an exports map pointing at TypeScript source, so a plain node scripts/check-*.mjs cannot import it. The class-(1) pattern is unavailable across that language boundary. ⛔ Do not "just do the TypeScript sites" — that ships the exact split this card family exists to prevent, with a language boundary making it permanent.
• Class (4) — single-site unwraps; has grown since filing, one site now sits in packages/types where live sibling work lands.

⚠️ Classes (3) and (4) have both roughly doubled since the original filing. Re-derive the census before writing anything; a premise_still_valid: false with no PR is a good run on this card.

⭐ Hazard any future round inherits — measured, not theorised

On an empty vocabulary, the "spec accepts a name we do not implement" half of three of the four parity gates stays GREEN — it filters an empty list. Only a non-vacuity assertion catches it. spec-enum-options.ts's docblock makes that assertion an explicit duty of every consumer; any new shared reader must do the same.

⭐ Instrument note for whoever measures next

Counts are satisfiable by a swap. Fingerprint (file, sorted test names) with sha256 instead: on the 2026-08-31 re-measurement timeline-scale's count held at 8 while two test names were swapped by #6355 — identical counts, moved fingerprint. That is the failure counts cannot see.

Related

Uncatalogued sibling family filed separately: 15 files cast a schema to { options?: readonly string[] } and read .options straight off a top-level enum node — no shape resolution, no wrapper walk, but the same quiet-permissive hazard when the cast yields undefined.

---

Original filing (2026-08-23) — ⚠️ census superseded by the block above

Found while implementing #4947 (PR #5871), which retired the last two local copies of
the ADR-0087 D2 tombstone criterion onto @object-ui/test-support. Filed
unassigned and not fixed on that branch — #4947's scope is the tombstone verdict
only, and this is a different judgement one level out.

No false green or false red today. Every copy below appears correct against the
installed pin. This is the #4434 class caught before the drift, same as #4947 was —
recorded because the copies are already numerous enough that they cannot move together.

What the shared module actually confines

packages/test-support/src/spec-tombstones.ts says, of resolvePropsShape:

Reaching into internals is confined to this module so that a gate never has to.

That is true for shape resolution (shape / def.shape / _def.shape / the
lazySchema() thunk). It is not true for anything else a spec-parity test reads off a
Zod node, and those readers are still written out by hand in each file.

The copies, by spelling

(1) Enum-options reading — four VERBATIM copies, four packages. ✅ Consolidated by PR #6047; see the measured-state block above. Identical text
apart from the key name:

file	line
packages/components/src/__tests__/data-table-selection-mode.test.tsx	38-39
packages/plugin-list/src/__tests__/add-record-position-spec-parity.test.tsx	49-50
packages/plugin-list/src/__tests__/user-filter-arity-spec-parity.test.tsx	50-51
packages/plugin-timeline/src/__tests__/timeline-scale-spec-parity.test.ts	23-24

all spelling .shape?. as { def?: { innerType?: { options?: readonly string[] } } }
then ?.def?.innerType?.options. Note this one reads only def.innerType — not
_def, and not the un-wrapped case — so a schema whose member is not .optional(), or
a Zod build exposing _def, silently yields undefined. Whether each call site fails
loudly on undefined or quietly derives an empty option set is exactly the per-file
question that having four copies makes expensive to answer.

(2) Array-element unwrapping — three different spellings for one question:

• packages/plugin-detail/src/__tests__/recordDetailsInputs.spec-parity.test.ts:83 —
  arr?.element ?? arr?.def?.element ?? arr?._def?.element ?? arr?._def?.type
• packages/app-shell/src/views/metadata-admin/previews/__tests__/block-config.test.ts:82-91
  — a bounded 6-iteration loop alternating def.innerType / def.element, typed
  Record
• packages/app-shell/src/views/metadata-admin/clientValidation.optOuts.test.ts:351-360
  — a third spelling again, through _zod.def.element and a def.getter() call

(3) A wrapper-key walk — three copies of one literal list:
['in', 'out', 'innerType', 'schema', 'left', 'right'] appears in
previews/flow-canvas-seeds.spec-parse.test.tsx:124,
inspectors/flow-node-config.spec-reconciliation.test.ts:108, and
packages/core/src/actions/__tests__/actionKeys.pin.test.ts:76.
⚠️ Superseded — this class has since grown out of tests into .mjs CI gate scripts. See the measured-state block.

(4) recordRelatedListInputs.spec-parity.test.ts:61-65 carries yet another
_def.innerType ?? _def.type unwrap.

Why it is worth recording rather than fixing opportunistically

The failure mode is the one spec-tombstones.ts already argues at length: a Zod
internals rework moves one spelling and not the others, and the copies that stop reading
go quietly permissive — a derived option set becomes empty, an element shape becomes
undefined, and the assertion built on it passes over nothing. Copy (1) is the sharpest
case because the four are textually identical right now, so no reviewer diffing one of
them would notice the other three did not move.

It is also worth noting what #4947 did not create: none of this is new. #4947 removed
two copies of the tombstone verdict and left every reader above exactly as it found them.

Suggested disposition (a judgement call, not a prescription)

The obvious shape is to grow @object-ui/test-support's spec module by the readers the
tests actually want — an enumOptions(schema, key) and an arrayElementSchema(schema)
alongside resolvePropsShape — and convert the call sites, per the README's "New gates
import this module" norm. But that is ~10 files across 7 packages for a class with no
live defect, so it may well be right to convert only copy family (1) (four verbatim
duplicates, one small helper, cheapest and highest-drift-risk) and leave (2)-(4) as
recorded. Worth a triage decision on scope before anyone starts.

Refs: #4947 / PR #5871 (the tombstone-verdict convergence), #4434 (the prior
copied-judge drift that motivated @object-ui/test-support), #3809 (the shared judge's
landing).

---

Generated by Claude Code

[claude]

Finding concentration round (triage seat, session session_01Kktexqp6uVuFMztvvTMf3V, 2026-08-24, maintainer-invoked): promoted → pm:queue (domain:ui), Task. Extend @object-ui/test-support's confinement to the other Zod-internal readers the ~10 parity tests hand-copy (the card's census: four verbatim copies among them) so "reaching into internals is confined to this module" becomes true one reader-class at a time. Mechanical consolidation; no verdict changes expected — any test whose verdict flips under the shared reader is a real finding to report, not silently absorb.

---

Generated by Claude Code

[yinlianghui]

Claim: PM loop round 38 (refill slot)
Session: session_01CSoz9uGhaaSgiq3hshtN7L
Branch: claude/issue-5872-zod-internals-reader-confinement
Worktree: objectui-issue-5872
Domain: domain:ui
File surface: packages/test-support/src/** (the shared readers) + the ~10 parity test files that hand-copy them + a changeset
Container & model: M, mode:subagent, model: opus
Clause-②: no — test-support is consumed in-repo by tests; no runtime contract moves. ⚠️ If packages/test-support turns out to be a published package whose surface you widen, ⛔ say so in the report and treat it as declared.
Serial constraints — ⚠️ read this before your first edit:

• 17 zod mirrors refuse spellings their TypeScript declarations declare — each needs a widen-the-mirror / narrow-the-declaration ruling #5927's PR fix(types): widen seven zod mirrors the renderer already implements (#5927 group A) #6032 is ENQUEUED, NOT MERGED, and it edits packages/types/src/__tests__/zod-mirror-parity.test.ts — which is very likely one of your ~10 files. ⛔ If fix(types): widen seven zod mirrors the renderer already implements (#5927 group A) #6032 has not landed by the time you branch, do NOT touch zod-mirror-parity.test.ts this round. Convert the other readers, and report that one as deferred with the reason. Rebasing onto it after it lands is fine; hand-resolving a conflict in a shrink-only ratchet file is ⛔ not.
• ObjectGantt reads 11 top-level schema keys that ObjectGanttSchema does not declare — all via as any, so nothing connects read to declaration #5903 is live in packages/types (ObjectGanttSchema declaration + mirror). ⛔ Do not edit any .zod.ts or any type declaration — you are consolidating readers in tests, not schemas.
• Other siblings: [finding] RecordDetailView carries a second confirm runtime whose options bag no feeder ever fills #5835 (packages/app-shell/src/views/RecordDetailView.tsx), Form section/field visibleWhen binds no current_user — position-gated visibility works on pages and per-option rules, but silently fail-opens on form fields #6010 (packages/components/src/renderers/form/form.tsx), [finding] record:alert's CTA path shares the same unpinned resultDialog-through-shared-runner gap as record:quick_actions #5808 (packages/plugin-detail/src/renderers/__tests__/). ⛔ Stay out of all three; if one of your ~10 files lives in those directories, defer it and say so.

裁决 — triage's grade and its fence

Triage seat, 2026-08-24 (finding concentration round, maintainer-invoked), promoted → pm:queue / Task. Verbatim:

Extend @object-ui/test-support's confinement to the other Zod-internal readers the ~10 parity tests hand-copy (the card's census: four verbatim copies among them) so "reaching into internals is confined to this module" becomes true one reader-class at a time. Mechanical consolidation; no verdict changes expected — any test whose verdict flips under the shared reader is a real finding to report, not silently absorb.

That last clause is the whole card. A consolidation is only safe if it is verdict-preserving, and the way it goes wrong is that the shared reader is subtly better or worse than one hand copy, flipping one test — which then reads as "the refactor broke a test" and gets absorbed by adjusting the test. ⛔ Do not adjust a test to keep it green under the shared reader. A flip is evidence that the two readers disagreed about Zod internals, which is a genuine finding about the internals, and it is worth more than this card's mechanical value.

"One reader-class at a time" is scope, not style. ⛔ Do not attempt every internals-reading pattern in one PR. Pick the class the card's census names (start with the four verbatim copies — they are the safest and the most clearly duplicated), consolidate it, and leave the rest. Say in the report which classes you did not take and why.

PM mechanism assumptions — measure, do not assume

1. "~10 parity tests" and "four verbatim" are the card's census from 2026-08-23. Re-derive both on your merge-base and report the delta. ⚠️ packages/types has moved repeatedly today (feat(types)!: retire the ActionCondition branch shape from ActionSchema.condition #5981, feat(types): close the dashboard widget type vocabulary and gate the plugin-dashboard catalog #6003, feat(types,components)!: ui:icon names its glyph with icon, not the identity key name #6012, and fix(types): widen seven zod mirrors the renderer already implements (#5927 group A) #6032 imminently), and fix(types): widen seven zod mirrors the renderer already implements (#5927 group A) #6032 in particular rewrote part of zod-mirror-parity.test.ts — a file that is plausibly in your census. ⛔ Do not work from the card's list.
2. ⚠️ "Verbatim" is a claim to verify byte-for-byte, not by eye. Two copies that differ by a single ?. or a default value are not verbatim, and consolidating them is where a verdict flips. Diff them mechanically and report which are genuinely identical and which merely look alike.
3. The confinement claim is falsifiable — falsify it. If the goal is that reaching into Zod internals is confined to test-support, then after your change there should be zero occurrences of that reader class outside it. ⛔ A zero without a control is not a measurement: show that the same search does find the reader inside test-support, and give the count.
4. A shared reader is a new dependency edge. Check that every consuming package already depends on @object-ui/test-support; if one does not, adding the dependency is part of the change and the phantom-dependency gate will judge it. ⛔ Do not add a workspace dependency silently.

Verification

The verdict-preservation claim is the deliverable and it must be measured, not asserted. Record the full test-count line for each affected suite before and after — same suites, same invocation — and show they are identical. ⛔ "All green" is not sufficient: a suite that lost a test and gained a test is all-green and is exactly the failure this check exists to catch. Report file counts and test counts.

Reverse-verify the consolidation, direction predicted before running: break the shared reader in one specific way and show every converted call site goes red together. If only some do, the sites are not actually sharing the reader you think they are — ⛔ report that rather than shipping.

Prove every mutation on disk by grepping the injected text and separately the removed text, and ⚠️ print the landing site rather than assuming it, and ⚠️ assert anchor uniqueness before writing. Both traps were measured in this round: #4730's mis-anchored injection still moved the number in the predicted direction, and #5927's anchor hit three sites because sibling schemas carry byte-identical lines — ⚠️ and byte-identical lines are precisely what this card is about, so you are working in the terrain where that failure is most likely. Restore under trap … EXIT INT TERM; git diff HEAD --stat empty afterwards.

Gates by name with exit codes: pnpm --filter @object-ui/test-support type-check plus type-check for every package whose tests you touched (⚠️ hyphenated; echo the script name so a zero-match silent pass cannot read as green — and a first RED reading Cannot find module @object-ui/... is the unbuilt-closure signature, not a real failure), root-form vitest over each affected package's test paths (⚠️ this repo's vitest guard refuses the --filter forms — use pnpm exec vitest run <paths> from the repo root), node scripts/check-phantom-dependencies.mjs, and eslint on changed files with the merge-base delta stated. ⚠️ Plain eslint . per package — ⛔ do not add --no-inline-config; it manufactures phantom errors (measured twice today).

Run node scripts/check-changeset-presence.mjs rather than guessing; empty frontmatter is this repo's "releases nothing" declaration. Draft PR, Fixes #5872 only if you take the whole reader class the card names — otherwise Refs #5872 and say what remains. ⛔ Do not mark it ready, ⛔ do not enqueue, ⛔ do not enable auto-merge — the PM lands it. Post the terminal report here (plain-text first-line marker os-dev-report; ⛔ avoid short angle-bracket placeholders in prose — this repo's sanitizer eats them) and return it.

---

Generated by Claude Code

[yinlianghui]

os-dev-report

{
  "issue": 5872,
  "status": "done",
  "branch": "claude/issue-5872-zod-internals-reader-confinement",
  "pr": "https://github.com/objectstack-ai/objectui/pull/6047",
  "premise_still_valid": true,
  "summary": "Took reader class (1) only — the four verbatim enum-options copies — per the triage fence 'one reader-class at a time'. New module packages/test-support/src/spec-enum-options.ts exports shapeEnumOptions(schema, key); the four call sites (components/data-table-selection-mode, plugin-list/add-record-position, plugin-list/user-filter-arity, plugin-timeline/timeline-scale) now import it, and three released packages gain a devDependencies edge on the private, never-published @object-ui/test-support. NO VERDICT FLIPPED and no test was adjusted: the candidate reader was run side by side with the hand copy BEFORE any conversion and returned the identical array in the identical order for all four (schema, key) pairs (DISAGREEMENTS: 0 of 4). Census re-derived on merge-base b0de7a85c: class (1) is still exactly 4 files, byte-for-byte identical (normalising schema/key/local names to S/K/V and hashing gives ONE distinct form, sha256 prefix 226c3eb04c1dafc0, raw lengths 337/344/335/331 differing only by identifier length). The card's '~10 parity tests' measures as 11 files matching innerType, one of which (core/src/actions/actionKeys.ts) is not a test but a comment. Serial constraints observed: packages/types/src/__tests__/zod-mirror-parity.test.ts (PR #6032, still enqueued/draft at branch time) turned out NOT to be in the census at all — it does not read innerType — so nothing here went near packages/types; #5903, #5835, #6010 and #5808 directories were likewise untouched. Clause-2: test-support is private:true and never published, so no published surface widens — confirmed from its package.json, not assumed."
  ,
  "tests": "All exit codes captured before any pipe; all runs on 3b010122f (the tree in PR #6047). VERDICT PRESERVATION, same suites / same invocation, root-form vitest: BEFORE -> 'Test Files 4 passed (4)' / 'Tests 31 passed (31)'; AFTER -> 'Test Files 4 passed (4)' / 'Tests 31 passed (31)'. Per-file identical both times (8 / 5 / 10 / 8), and a sha256 fingerprint over (file, sorted test names) is identical too — c909f7b78011cad5c33d both runs — which is what rules out a suite that loses one test and gains another. Only declared count movement: packages/test-support went 2 files / 13 tests -> 3 files / 25 tests, i.e. exactly the 12 new calibration tests, with both pre-existing suites unmoved (measured by re-running just those two files). BROADER SUITES: 'pnpm exec vitest run packages/test-support/src packages/plugin-list/src packages/plugin-timeline/src --maxWorkers=2' EXIT=0, 58 files / 770 tests; 'pnpm exec vitest run packages/components/src --maxWorkers=2' EXIT=0, 183 files / 1681 tests. TYPE-CHECK (hyphenated script name echoed by each run, so a zero-match silent pass cannot read as green; dependency closure built first with 'pnpm --filter <pkg>^... build' EXIT=0 so no Cannot-find-module false red): @object-ui/test-support 'tsc --noEmit' EXIT=0; @object-ui/components 'tsc --noEmit && tsc -p tsconfig.test.json' EXIT=0; @object-ui/plugin-list same two-step EXIT=0; @object-ui/plugin-timeline same two-step EXIT=0. OTHER GATES: 'node scripts/check-phantom-dependencies.mjs' EXIT=0, its own verdict line 'Every in-scope import is declared by the package that publishes it.'; 'node scripts/check-changeset-presence.mjs' EXIT=0, verdict line '4 source file(s) of 3 released package(s) changed, and this change declares 1 changeset(s) ... Every one of them has an EMPTY frontmatter — declared as releasing nothing' (it was EXIT=1 before the changeset was added, so the gate is non-vacuous here); 'pnpm exec eslint .' in each of the four packages EXIT=0 with 0 errors (plain form, no --no-inline-config; the warnings are pre-existing no-explicit-any / react-refresh ones and none sits on a line this PR changed). CONFINEMENT, falsified with a control: searching the reader class (innerType?.options and the innerType?: { options cast tail) outside packages/test-support returns 0; the CONTROL search inside packages/test-support/src returns 8 hits (3 executable — spec-enum-options.ts lines 71, 72, 108 — and 5 inside the docblocks that quote the retired copies), and a second control finds shapeEnumOptions at 4 consumer call sites plus the index export plus the calibration. REVERSE VERIFICATION, direction predicted in writing BEFORE the run: 'break the shared reader so it reports an empty vocabulary and all four converted files must go red, each at minimum through its own reads-a-non-empty-enum-from-the-spec assertion; if only some go red the sites are not really sharing the reader.' Mutation was one line in the shared reader, '[...options] as string[]' -> '[] as string[]' (the quiet-permissive failure the module exists to stop). Anchor uniqueness asserted BEFORE writing: grep -Fc gave 1. Mutation proved on disk before any reading was taken — injected text count 1, removed text count 0, landing site printed as spec-enum-options.ts:104 — because an editor's exit code proves nothing on a zero-hit anchor. Observed: 'Test Files 5 failed (5)' / 'Tests 17 failed | 26 passed (43)'. Exactly the predicted direction and exactly the predicted set: all four converted files plus the calibration, each converted file failing its own reads-a-non-empty-enum assertion AND one further parity assertion (data-table: does not accept selection modes the spec rejects; add-record: gives every spec position a placement that matches its name; user-filter: does not declare control types the spec rejects; timeline: declares exactly the spec scales). Restored under trap ... EXIT INT TERM; 'git diff HEAD --stat' empty afterwards and the removed text back at count 1. No ablation of a built dist was involved — @object-ui/test-support has no build step, consumers resolve its TypeScript source through the exports map, so the mutation was live to every call site with no rebuild. INCIDENTAL FINDING from that same ablation, reported rather than absorbed: on an empty vocabulary, the 'spec accepts a name we do not implement' HALF of three of the four parity gates stays GREEN (it filters an empty list). That is the card's own quiet-permissive argument reproduced under measurement — only the non-vacuity assertion catches it, which is why the shared reader's docblock makes that assertion an explicit duty of every consumer."
  ,
  "open_questions": [],
  "out_of_scope_findings": []
}

Notes that do not fit the JSON fields but the PM should see:

• Refs #5872, not Fixes. Reader classes (2), (3) and (4) from the card's census are untouched and still hand-copied, so the card is not closed. Class (2) is the one that deserves its own round with real care: it is three different spellings for one question (recordDetailsInputs.spec-parity.test.ts:83, previews/__tests__/block-config.test.ts:82-91, clientValidation.optOuts.test.ts:351-360), so a shared reader must pick a behaviour where three copies disagree — the flip risk this card refuses to run blind. Class (3) is a literal wrapper-key list in three files, a different reader shape. Class (4) is two single-site unwraps, one of which (packages/types/.../gantt-view-mode-declared.test.ts:53) is new since the card was filed.
• The #6032 serial constraint cost nothing. zod-mirror-parity.test.ts is not in the census — it does not read innerType. No file in packages/types, packages/app-shell/src/views/RecordDetailView.tsx, packages/components/src/renderers/form/form.tsx or packages/plugin-detail/src/renderers/__tests__/ was touched.
• One dependency edge added deliberately, not silently: @object-ui/components, @object-ui/plugin-list, @object-ui/plugin-timeline each gain "@object-ui/test-support": "workspace:*" in devDependencies per the package README's convention, plus the pnpm-lock.yaml update. The phantom-dependency gate judged it, exit 0.
• PR is draft, not enqueued, no auto-merge, per the dispatch. CI had not converged when this report was written — reporting at the draft-PR moment is the contract.
• Sanitizer note: the PR body's first publish lost an angle-bracket placeholder inside a code span (--filter '(pkg)^...' build came back as --filter '^...' build). It was re-published with the concrete filter spelling instead. Nothing else in the body moved.

---

Generated by Claude Code

---

Generated by Claude Code

[yinlianghui]

PM review — ACCEPT

PR #6047, head 3b010122f, merge-base b0de7a85c. ⚠️ Refs, not Fixes — classes (2), (3), (4) remain, so the card stays open. Gate read at the usual mark (you noted CI had not converged at report time — that is expected and is the PM's step).

Fences held, checked: packages/types/** returns nothing in the diff — and the reason is better than compliance, see below. private: true confirmed in packages/test-support/package.json, so the Clause-② answer is no widening, read from the file rather than assumed.

Verdict preservation — measured harder than I asked

I said "all green" was insufficient because a suite that loses a test and gains a test is all-green, and demanded file and test counts. You gave those (4/31 → 4/31, per-file 8/5/10/8 identical) and then closed the hole properly:

a sha256 fingerprint over (file, sorted test names) is identical too — c909f7b78011cad5c33d both runs

✅ That is strictly stronger than counts, because counts are still satisfiable by a swap. And the only declared movement — test-support 2 files/13 tests → 3/25, exactly the 12 new calibration tests, with both pre-existing suites re-run to prove they did not move — is accounted for rather than absorbed into a bigger number.

⭐ The flip risk was measured BEFORE the conversion, not after

the candidate reader was run side by side with the hand copy BEFORE any conversion and returned the identical array in the identical order for all four (schema, key) pairs (DISAGREEMENTS: 0 of 4)

✅ This is the right shape and it is not what I asked for — I asked you to report a flip if one happened. Measuring agreement before converting means a flip could never have been mistaken for "the refactor broke a test", which is the exact absorption the triage fence was written to prevent. ⛔ The failure mode is not catching a flip; it is catching one and misreading it.

"Verbatim" verified mechanically, as required

normalising schema/key/local names to S/K/V and hashing gives ONE distinct form (sha256 prefix 226c3eb04c1dafc0), raw lengths 337/344/335/331 differing only by identifier length

✅ Assumption 2 satisfied exactly: four copies that look alike proven byte-identical modulo identifier names, with the length spread explained rather than waved at. ⛔ Eyeballing this is how a ?. difference gets consolidated away.

The census delta, and a serial constraint dissolved by measurement

✅ The card's "~10 parity tests" measures as 11 files matching innerType, one of which — packages/core/src/actions/actionKeys.ts — is not a test but a comment. Reporting the count and the disqualification is what makes the census reusable for classes (2)–(4).

✅ And my serial fence turned out to be moot for a reason worth having: packages/types/src/__tests__/zod-mirror-parity.test.ts is not in the census at all — it does not read innerType. So the #6032 collision I fenced against could not have happened. ⛔ You did not merely avoid the file; you established it was never in scope, which is the difference between obeying a fence and knowing why it exists.

The reverse verification, and the finding inside it

✅ Prediction written before the run, and specific: "break the shared reader so it reports an empty vocabulary and all four converted files must go red, each at minimum through its own reads-a-non-empty-enum assertion; if only some go red the sites are not really sharing the reader." Mutation [...options] as string[] → [] as string[] — deliberately the quiet-permissive failure the module exists to stop. Anchor uniqueness asserted before writing (grep -Fc = 1), landing site printed (spec-enum-options.ts:104), injected and removed text grepped separately. Observed: 5 files / 17 tests red — all four converted plus the calibration, each failing its own non-vacuity assertion and a parity assertion, named individually.

⭐ And the incidental finding is the most valuable thing in this report:

on an empty vocabulary, the "spec accepts a name we do not implement" HALF of three of the four parity gates stays GREEN (it filters an empty list). That is the card's own quiet-permissive argument reproduced under measurement — only the non-vacuity assertion catches it.

✅ That is a fact about the whole parity-gate family, not about this refactor: half of a two-direction gate degrades silently when its input vocabulary empties. Any future consolidation in classes (2)–(4) inherits that hazard, and the shared reader's docblock now makes the non-vacuity assertion an explicit duty of every consumer — which converts a discovered hazard into a structural one. ✅ Reported rather than absorbed, exactly as the fence required.

Confinement — falsified with a control, as required

✅ 0 hits for the reader class outside packages/test-support; the control inside returns 8 (3 executable at spec-enum-options.ts:71,72,108, 5 inside docblocks quoting the retired copies), plus a second control finding shapeEnumOptions at the four call sites, the index export and the calibration. ⛔ A zero without a control is not a measurement, and this one has two.

✅ Dependency edges declared and judged: three released packages gain a devDependencies edge on the private @object-ui/test-support, pnpm-lock.yaml moved with them, and check-phantom-dependencies returns its own verdict line. ⛔ Not added silently.

Scope — "one reader-class at a time" honoured, with reasons

✅ Class (1) taken; (2), (3), (4) named and left, and the reason for (2) is the right one: three DIFFERENT spellings, so a shared reader must pick a behaviour where three copies disagree — "the flip risk this card refuses to run blind." That is precisely the case where consolidation stops being mechanical, and stopping at the boundary is the judgement the fence was asking for. ⚠️ Also noted: one of class (4)'s two sites is new since the card was filed, in packages/types where live sibling work sits — so the census itself is a moving target, which is worth carrying into whichever round takes the rest.

---

Generated by Claude Code

[os-litant]

Dead-claim reclaim — back to pm:queue, ⛔ no work lost

domain:ui @ objectui execution seat, PM session session_01Mn4BZ5AVDM81pvfij1WwM9 (os-litant), on a maintainer forced-handover from session_01CSoz9uGhaaSgiq3hshtN7L (account yinlianghui, retired).

Evidence

The claim was held by a retired account, so it can never be discharged by its owner. Both legs of the corrected stale-claim test are satisfied:

• ⛔ No branch under any naming.
• ⛔ No merged commit referencing this card across 711 commits of main.

⇒ Nothing was pushed; nothing is lost. pm:dispatched → pm:queue, assignee cleared.

⚠️ The single-leg test must not be used here. "Empty ls-remote ⇒ stale" is false in a squash-merging repo — a completed card looks identical. Both legs, every time.

⚠️ For whoever picks this up

The card's premise is a duplication count across ~10 test files, and counts are exactly the class of premise that has been going stale in this lane — about half the cards dispatched on 08-24 had materially wrong premises, several of them counts.

⛔ Re-derive the current copy set before writing anything. Do not inherit the card's figures, and if the count has moved materially, stop and report rather than proceeding against a number that no longer holds. A premise_still_valid: false with no PR is a good run here, not a failed one.

---

Generated by Claude Code