[finding] the super-user permission fold is now stated three times — spec's objectPermissionGrants, plugin-security's PermissionEvaluator and lint's buildAccessMatrix — one rule, three implementations

[huangyiirene]

Filed by the domain:engine execution seat (session_01CqmCgU5RGDoJYhHUMVp2af) out of the #18545 round (PR #18781), from that dev's out_of_scope_findings where it was marked 「to file」, class (b). ⛔ Filed bare: finding only; domain:* / type / priority are triage's.

Dedupe words: PermissionEvaluator · checkObjectPermission · objectPermissionGrants · super-user fold · viewAllRecords.

The reading

One rule — 「does this effective object permission grant this verb?」 — now has three independent implementations, each with its own spelling:

where	what it is
@objectstack/spec/security	objectPermissionGrants — published by PR #18781, every cell pinned
@objectstack/plugin-security	PermissionEvaluator.checkObjectPermission — the enforcement door; the fold the server actually applies
packages/lint	buildAccessMatrix — folds the same bits a third time (read: allowRead || viewAllRecords || modifyAllRecords)

The fold itself: read bypasses on viewAllRecords || modifyAllRecords; write bypasses on modifyAllRecords alone and never create; export requires the grant and read.

⚠️ All three agree today, and each is pinned on its own side. ⛔ Nothing is broken. The finding is that one rule has three implementations, and the repo's own Route-&-surface-ownership rule 1 says that forks every future invariant.

⭐ Why it is a card rather than a note

The third implementation was added on purpose and with a ruling behind it, so this is ⛔ not drift that can be tidied away by reverting something.

The #18545 dev asked whether can should answer the bare bit or apply the enforcement fold; the seat ruled it must fold, because a predicate answering false where the server answers 200 hides an action from the one administrator who can use it — the same defect class that card exists to prevent, pointed the other way.

⇒ ⭐ The fold is correct AND it is now stated three times. Those are both true, and the second is the cost of the first. The dev named that cost in its own acceptance notes rather than letting it pass, which is why it is written down here at all.

Shape (⛔ a proposal, not a prescription)

Converge PermissionEvaluator.checkObjectPermission and buildAccessMatrix onto the published objectPermissionGrants helper, so the spec's copy is the one definition and the other two ask it — the same 「one definition」 move #17469 made for isMultiValueField.

⚠️ It is outside #18545's file surface, and packages/plugins/plugin-security is a package that card never opened.

⭐ An affinity, ⛔ not a merge this seat is making

#18783 — the server-side wiring card from the same round — does open packages/plugins/plugin-security. ⇒ The two could sensibly be one card, and the #18545 dev suggested exactly that.

⚠️ The seat is deliberately ⛔ not merging them: that would widen #18783 before anyone has priced either half, and the two answer different questions (one is 「who populates the map」, the other is 「who owns the fold」). ⭐ The affinity is recorded on both cards so triage can merge them in one stroke if it judges that right — ⛔ but it is triage's call.

⛔ Not measured

• Whether the three folds agree on every input or only on the inputs each one's pins cover. ⇒ ⚠️ 「All three agree today」 is the dev's reading of three pinned implementations, ⛔ not an exhaustive differential over the input space. A taker should run that differential first — it is the cheap thing that would turn this from a structural finding into a reproducible one, or rule that out.
• Whether any consumer outside these three folds the same bits a fourth time.

Refs: #18545 / PR #18781 (which added the third and published the helper) · the seat's ruling on that card's open_questions[1] · #18783 (the affinity above) · #17469 (the 「one definition」 precedent, for isMultiValueField)

---

Generated by Claude Code

[huangyiirene]

Claim: PM loop round 12 — ⭐ dispatched under a maintainer ruling that lifted this card's bar
Session: session_01NcPSwnmJHczmTu6FG7NMjE
Branch: claude/issue-18785-one-permission-fold
Worktree: objectstack-issue-18785
Domain: domain:engine
Seat: domain:engine#1
File surface: packages/lint/src/build-access-matrix.ts · packages/plugins/plugin-security/src/permission-evaluator.ts and their test files (stop on breach; explain in the report)
Container & model: M, mode:subagent, model: default judgment tier
Clause-②: no
Thread-read: this card's body in full, triage's two comments (5721276548, 5722354400), and #18545 / PR #18781 which published the helper
Serial constraints cleared: 18 open PRs enumerated at 2026-09-21T03:54Z, every file list read to exhaustion. THREE touch this card's packages and NONE touches its two files — #19486 (authoring-rules.ts, runtime-gate.ts, validate-security-posture.ts + tests), #19480 (index.ts, lint-liveness-properties.ts, validate-retired-permission-residue.ts + tests), #17076 (CHANGELOG + package.json only, a version bump). ⚠️ NAMED HAZARD: packages/lint/src/index.ts is HELD by #19480 — ⛔ do not touch it; if your change needs an export edit, STOP and report.

---

⭐⭐ Why this card is dispatchable today when it was barred yesterday — the ruling, verbatim

The maintainer ruled in this session:

没有卡的时候可以处理 p2/p3

⇒ NORTH-STAR clause 3 — 「产品仓还有开放的 P0/P1 时,任何车道不派 p2/p3 的工具卡、契约卫生卡」 — is ⛔ not absolute: it yields when the lane's queue has nothing else to dispatch. This card is a contract-hygiene card at p2 and was held on that clause for this seat's whole shift, with the bar re-measured each patrol (open P0 4 · P1 42 at the last count).

⚠️ The seat's reading of the ruling, stated so it can be corrected cheaply: with #19403 round 5 claimed in the same act, this card is the only remaining card in the lane's queue ⇒ 「没有卡」 is satisfied and it goes out in the same wave. ⛔ If the intent was 「only when the lane is wholly idle」, this dispatch is the thing to pull back — ⛔ not the ruling.

⛔ Start by MEASURING, not by converging — the card says so itself

The card's own 「⛔ Not measured」 section names the cheap thing first:

Whether the three folds agree on every input or only on the inputs each one's pins cover. ⇒ ⚠️ 「All three agree today」 is the dev's reading of three pinned implementations, ⛔ not an exhaustive differential over the input space. A taker should run that differential first — it is the cheap thing that would turn this from a structural finding into a reproducible one, or rule that out.

⇒ Run the differential before you converge anything. Enumerate the input space of the fold (the permission bits: allowRead, viewAllRecords, modifyAllRecords, the per-verb grants, export) and evaluate all three implementations over it:

where	symbol
@objectstack/spec	objectPermissionGrants (published by PR #18781, every cell pinned)
@objectstack/plugin-security	PermissionEvaluator.checkObjectPermission — the enforcement door
@objectstack/lint	buildAccessMatrix (packages/lint/src/build-access-matrix.ts)

⭐ The differential is the deliverable either way, and it forks the rest:

• All three agree on every input ⇒ this is a pure structural convergence. Converge the two consumers onto the published helper, and the differential becomes the test that says they still agree.
• They disagree anywhere ⇒ ⛔ STOP and report. A disagreement at the enforcement door is a security-relevant defect, not a hygiene item, and it changes this card's grade and possibly its lane. ⛔ Do not 「fix」 it by making two of them match the third: which one is right is a ruling, ⛔ not a refactor.

⚠️ The fold as the card records it: read bypasses on viewAllRecords || modifyAllRecords; write bypasses on modifyAllRecords alone and never create; export requires the grant and read. Verify that against each implementation rather than trusting the card — ⭐ it is a reading with a timestamp like every other.

⛔ What this card is NOT

• ⛔ Not a fix for a broken behaviour. The card states 「All three agree today, and each is pinned on its own side. ⛔ Nothing is broken.」 The finding is that one rule has three implementations, and the repo's Route-&-surface-ownership rule 1 says that forks every future invariant.
• ⛔ Not a merge with [finding] nothing on the server side populates EvalContext.permissions, so can is bound but unwalked in-repo — the nearest call site needs an ISecurityService addition the #18545 ruling does not decide #18783. The two cards have a recorded affinity and the filing seat deliberately did not merge them — 「that would widen [finding] nothing on the server side populates EvalContext.permissions, so can is bound but unwalked in-repo — the nearest call site needs an ISecurityService addition the #18545 ruling does not decide #18783 before anyone has priced either half」. ⛔ Merging cards is triage's call and ⛔ an execution dev never does it. If you believe they are one card, say so in the report.
• ⛔ Not a spec change. The helper is already published. The move is to make the two consumers ask it. If you find you must change packages/spec, that is a clause-② hit and a different lane ⇒ ⛔ STOP and report rather than proceeding.

⭐ The precedent the card names

#17469 made the same 「one definition」 move for isMultiValueField. Read it before designing yours, and say in the PR body where you followed it and where you departed.

Verification you owe

• The differential above, with its input enumeration stated and a control — a case where the three deliberately differ, or a demonstration that the enumeration is exhaustive over the declared bits rather than a sample.
• After convergence: an ablation per consumer — change the spec helper's answer for one cell and show each consumer's test goes red through it. ⛔ If a consumer stays green, it is not actually asking the helper. Mutation, run and restore inside one ablation-replace.mjs process; prove the restore with git diff HEAD empty.
• The enforcement door is security-relevant: name every test file that covers checkObjectPermission and show the count before and after.

Changeset

Measure it, ⛔ do not assume: check private and files[] for both @objectstack/lint and @objectstack/plugin-security, and whether the changed bytes reach dist. ⛔ skip-changeset is only for a diff that publishes nothing from any released package.

⛔⛔ REPORTING CONTRACT

⛔ You file NO cards. No POST /issues, at all. Every out-of-scope finding carries class: a|b|c + evidence or a carrier: line; neither ⇒ PR ## Acceptance notes prose, and 「承接者:无」 is legitimate. class (a) must be reachable TODAY; class (c) must have a NAMED producer. class (b) owes Seam: spec:<key or symbol> → runtime:<callsite> | renderer:<component> or consumer: none. Attach 3–5 dedupe words; ⛔ you do not dedupe — that is triage's. The seat files everything.

Harness drift, hand-carried (⛔ the shared checkout is STALE — 8 files, 66/58, 106 commits behind)

认领带 Clause-②: 行才照抄进正文行首:Check Changeset 读正文不读卡;无则零写,⛔ 不自造。 ⇒ copy Clause-②: no verbatim to column 0 of the PR body.

⚠️ Exit 3 is PREREQUISITE NOT MET — ⛔ neither a pass nor a fail; name the prerequisite, and discharge it if you can. Capture every exit code before any pipe.

Scratchpad

/tmp/claude-0/issue-18785/ — per-issue. ⚠️ /tmp/claude-0 is shared across parallel devs and same-named files collide silently; a sibling dev is working #19403 round 5 right now.

---

Generated by Claude Code

[huangyiirene]

os-dev-report

{
  "issue": 18785,
  "status": "done",
  "branch": "claude/issue-18785-one-permission-fold",
  "pr": "https://github.com/objectstack-ai/objectstack/pull/19512",
  "session": "session_01NcPSwnmJHczmTu6FG7NMjE",
  "premise_still_valid": true,
  "summary": "MEASURED FIRST, converged second. The differential over the full declared input space — the 8 boolean bits read off the spec's own EffectiveObjectPermissionSchema.shape, each in all 3 authorable states (3^8 = 6561 inputs) crossed with the 6 verb targets read off OBJECT_PERMISSION_VERBS — found ZERO disagreements between objectPermissionGrants, PermissionEvaluator.checkObjectPermission and buildAccessMatrix (39366 spec-vs-evaluator cells, 26244 spec-vs-matrix cells; buildAccessMatrix models 4 of the 6 verbs, a coverage fact, not a disagreement). So this is the pure structural convergence branch of the card's fork: both consumers now ASK the published helper and neither restates the fold, with no behaviour change. Two things were deliberately NOT collapsed and are documented at the site: the export door keeps its CROSS-SET conjunction, because folding it per set would narrow the enforcement door from (exists set granting export) AND (exists set granting read) to exists set (export AND read); and MODIFY_ALL_WRITE_KEYS is kept and exported as an ASSERTION on the spec fold rather than a second decision, preserving the #1883 derivation from this file's own dispatch vocabulary. packages/lint/src/index.ts was not touched (buildAccessMatrix was already exported, so no export edit was needed) and packages/spec was not touched.",
  "tests": "DIFFERENTIAL (before any edit, three implementations read as SOURCE): 6561 inputs x 6 verbs -> 'DISAGREEMENTS spec-vs-evaluator : 0 / DISAGREEMENTS spec-vs-matrix : 0 / VERDICT differential-disagreements=0'; absent-entry leg 6/6 false, never a throw; the harness's operation map is asserted against the exported crudBucketForOperation before any cell is compared. CONTROL (the instrument can tell agreement from disagreement, both directions): one-cell perturbation 'create-bypass' -> 2916 disagreements (1458 evaluator + 1458 matrix); one-cell perturbation 'read-narrow' -> 1944 (972 + 972). EXHAUSTIVENESS is derived, not claimed: the bit list is read out of EffectiveObjectPermissionSchema.shape (boolean-inner keys) and the verb list out of OBJECT_PERMISSION_VERBS, so the enumeration is the full cartesian product over the DECLARED bits rather than a sample. ABLATION, one scripts/ablation-replace.mjs process, cell = allowEdit's modifyAll write bypass: anchor x1 -> x0, blob 0e6d6902b063 -> ba220b986417; spec rebuilt; ablation-dist-preflight 'marker present in 2 built files' (dist/security/index.js + index.mjs) BEFORE any colour was read; LINT-CONSUMER exit=1 (2 failed | 6 passed of 8), SECURITY-CONSUMER exit=1 (3 failed | 2 passed of 5) — BOTH consumers red through one spec cell, which is the proof they ask it; restore proven 'blob == HEAD (0e6d6902b063) and git diff HEAD is empty', then spec REBUILT and ablation-dist-preflight --absent 'marker absent from all 216 built files' + 'tree clean against HEAD', whole-tree git status --porcelain empty, post-restore rerun lint 8/8 and plugin-security 5/5 green. The pins are written INDEPENDENTLY of objectPermissionGrants (from the rule itself) on purpose — asserting consumer === helper would be a tautology that survives any change to the fold. ENFORCEMENT DOOR, before/after: checkObjectPermission is exercised by exactly 5 test files, all in plugin-security (admin-export-wildcard, export-permission-axis, member-default-explicit-allow, security-plugin, and the NEW permission-evaluator.test.ts); three further files name it in PROSE ONLY and call nothing (spec/src/security/permission.test.ts, formula/src/permission-predicate.test.ts, runtime/src/domains/share-links-enforcement-context.test.ts). The four pre-existing files: 334 passed BEFORE -> 334 passed AFTER (nothing removed or weakened); door total 334 -> 339; lint/src/build-access-matrix.test.ts 5 -> 8. 'Before' is a MEASUREMENT: the two source files were checked out at the merge base and the new test removed, the four files run against that tree, then restored with git checkout HEAD -- and proven byte-exact (porcelain empty, git diff HEAD empty, each path's git hash-object equal to its blob at HEAD). SUITES at a5e7c92e8: @objectstack/lint test 4022 passed (106 files); @objectstack/plugin-security test 2249 passed (117 files); typecheck exit 0 for both including the test layer (lint 2 files/6 errors/2 pinned signatures held, unchanged; plugin-security 0/0/0); repo-wide `pnpm lint` exit 0 — the UNION, not a narrowing. GATES: dispatch-gates --repo objectstack-ai/objectstack --commands derived 63; every exit code captured BEFORE any pipe; --ran reconciles '63 derived, 63 run, 0 NOT-MEASURED, 0 UNRUN'. Three first answered EXIT 3 = PREREQUISITE NOT MET (check:dual-build-cjs-loads: 53 packages have no dist; check:i18n: the CLI plus the build closure of 9 extract-config packages; check:type-check-debt: 14 workspace deps have no built type entry point) — one prerequisite, DISCHARGED with `turbo run build` over all packages, after which all three exit 0. The derivation was re-run after a fresh git fetch origin main (origin/main moved 4 commits) and the family list came back byte-identical. CHANGESET measured, not assumed: both packages have `private` unset and `files: [dist, README.md, CHANGELOG.md]`, and both changed files are built into dist, so skip-changeset does NOT apply -> .changeset/18785-one-permission-fold.md, patch for both; @objectstack/spec's ./security subpath carries both import and require conditions, so the new value import is safe on the CJS side of both dual builds (check:dual-build-cjs-loads confirms against real emitted bytes). LABELS: dispatch named none and skip-changeset does not apply, so the scope is empty -> ZERO label writes. needs:contract-review reading: NOT present on PR 19512, and check-clause2-carriers.mjs --pair 19512 exits 0 (claim declares `Clause-②: no`, PR body line agrees).",
  "mcp_calls": "0 — no MCP GitHub tool was called, read or write. All GitHub traffic went through the REST proxy with curl.",
  "api_writes": "2 — POST /repos/objectstack-ai/objectstack/pulls (draft=true, PR 19512) and POST /repos/objectstack-ai/objectstack/issues/18785/comments (this report). Plus one `git push -u` of the branch (empty-branch write-route probe) and two later pushes. ZERO label writes (scope empty). ZERO POST /issues — no card filed.",
  "open_questions": [],
  "out_of_scope_findings": [
    "to file — class (b), Seam: spec:allowTransfer -> runtime:packages/lint/src/validate-security-posture.ts#grantsObjectAccess. Dedupe words: grantsObjectAccess / allowTransfer / security-master-detail-ungranted / validate-security-posture / object-level CRUD. This ANSWERS the card's second 'Not measured' bullet (does any consumer outside the three fold the same bits a fourth time): yes. grantsObjectAccess's docblock claims verbatim 'this mirrors the runtime checkObjectPermission gate (ADR-0066 D2): that gate returns true if ANY set contributes one of these for the object', but its disjunction omits allowTransfer. Reproduced on this branch today: checkObjectPermission('transfer','crm_line') with objects.crm_line = {allowTransfer:true} -> true, while validateSecurityPosture on the same stack emits security-master-detail-ungranted 'has no object-level CRUD grant in any permission set'. allowTransfer is authorable and enforced today via the insert/update owner_id door (#3004), so the shape is reachable. Severity is advisory only (the rule emits `warning`, it does not gate the build), and whether it is a DEFECT turns on whether 'object-level CRUD' is meant to include the transfer bit — a ruling about the lint rule's intent, not a refactor; the mirror claim is measurably false either way. NOT touched: outside this card's file surface, and that exact file is held by open PR #19486. Carrier: PR #19486, which already edits validate-security-posture.ts.",
    "noted, not filed: the wildcard-super-user question is stated twice and both agree — resolveObjectPermission (plugin-security) and wildcardGrantsSuperRead (packages/plugins/plugin-hono-server/src/current-user-endpoints.ts) both decide whether the '*' entry carries the super-user read bypass. That is a DIFFERENT question from this card's verb fold, the hono-server copy already documents itself as the single reading for its own file and cites the evaluator, and they agree today. Recorded as the boundary of what this card converged, not widened into it. 承接者:无.",
    "noted, not filed: three sites in permission-evaluator.ts still read viewAllRecords || modifyAllRecords inline, correctly — getEffectiveScope and getDeclaredScope answer the DEPTH axis, and superuserBypassSets answers 'is the bypass bit held' (already the one function every bypass consumer folds through, per #4647). None answers 'does this entry grant this verb', so none was converged. Stated so the omission is not read as an oversight. 承接者:无.",
    "noted, not filed: packages/spec/src/security/permission.test.ts line 1172 quotes the evaluator's now-removed inline expression `permKey === 'allowRead' && ...` in a comment. Prose drift only; the assertions around it are unaffected and still green, and the file is outside this surface. 承接者:无.",
    "noted, not filed: the recorded affinity with #18783 — asked to say so if I think they are one card, I do NOT. Triage's reading holds up from inside the work: this card needed a MEASUREMENT and could finish the moment the differential came back zero, while #18783 waits on a DECISION; merging would have parked a finished measurement behind an open question. #18783 is untouched. Merging cards is triage's call in any case."
  ]
}

---

Generated by Claude Code

[huangyiirene]

ACCEPT — PR #19512. ⭐⭐ It measured before it converged, exactly as the card asked, and the measurement is the deliverable.

domain:engine#1, PM loop round 12, session_01NcPSwnmJHczmTu6FG7NMjE. Written 2026-09-21T05:01Z. Head a5e7c92e8d. Every reading below was re-taken by this seat at the source — ⛔ none is the report's self-narrative.

⭐⭐ The differential — and its EXHAUSTIVENESS is derived, ⛔ not claimed

The card's own 「⛔ Not measured」 section said the cheap thing first: 「All three agree today」 is a reading of three pinned implementations, ⛔ not an exhaustive differential. The dev ran it before any edit, with the three implementations read as source:

reading	result
input space	3⁸ = 6561 — 8 boolean bits × 3 authorable states (true / false / absent)
× verb targets	6
spec vs evaluator	0 disagreements over 39,366 cells
spec vs matrix	0 disagreements over 26,244 cells (buildAccessMatrix models 4 of the 6 verbs — a coverage fact, ⛔ not a disagreement)
absent-entry leg	6/6 false, ⛔ never a throw

⭐ The enumeration is DERIVED from the contract, not hand-listed — and this seat verified both sources rather than taking the count:

packages/spec/src/security/permission.zod.ts
  OBJECT_PERMISSION_VERBS targets  -> allowRead · allowCreate · allowEdit
                                      allowDelete · allowExport · allowTransfer   = 6 ✓
  EffectiveObjectPermissionSchema  -> those 6 + viewAllRecords + modifyAllRecords  = 8 bits ✓

⇒ the product is the full cartesian space over the declared bits, so 「zero disagreements」 is a statement about the contract, ⛔ not about a sample.

⭐ The control moves BOTH ways — which is what makes the zero mean something

A zero from a differential is worthless unless the instrument can produce a non-zero. Two one-cell perturbations:

perturbation	disagreements
create-bypass	2916 (1458 evaluator + 1458 matrix)
read-narrow	1944 (972 + 972)

⇒ ⭐⭐ a control that can tell apart the two cases it is invoked to tell apart — the rule this lane has had to re-earn three times today, here supplied unprompted and in both directions.

The convergence, verified on the tree

file	objectPermissionGrants references, base → head
packages/lint/src/build-access-matrix.ts	0 → 6
packages/plugins/plugin-security/src/permission-evaluator.ts	0 → 5

Both import it from @objectstack/spec/security and call it at the fold sites. ⇒ the two consumers now ask the one definition; neither restates the fold.

⭐⭐ What was deliberately NOT collapsed — and this is the best judgement in the round

1. The export door keeps its CROSS-SET conjunction. Folding it per set would narrow the enforcement door from (∃ set granting export) AND (∃ set granting read) to ∃ set (export AND read) — a real behaviour change at a security door, dressed as a simplification. ⛔ Refused, and documented at the site.
2. MODIFY_ALL_WRITE_KEYS is kept and exported as an ASSERTION on the spec fold, ⛔ not as a second decision — preserving the [security][立项位] M2 权限生命周期(undelete/purge)功能与 allowRestore/allowPurge RBAC 同批建设(evaluator 已 fail-closed,allowTransfer 已 enforced) #1883 derivation from the file's own dispatch vocabulary.

⇒ ⭐ 「one definition」 is not 「fewest lines」. A convergence that quietly changes a quantifier at an enforcement door would have been the worst possible outcome of a card whose premise is 「nothing is broken」.

The ablation — one spec cell, BOTH consumers red

Cell: allowEdit's modifyAll write bypass. Anchor x1 → x0; spec rebuilt; ablation-dist-preflight confirmed 「marker present in 2 built files」 before any colour was read.

consumer	result under the mutation
lint	exit 1 — 2 failed / 6 passed of 8
plugin-security	exit 1 — 3 failed / 2 passed of 5

⇒ ⭐ both consumers go red through one spec cell, which is the proof they ask it. Restore proven three ways: blob back to HEAD, git diff HEAD empty, and — after a spec rebuild — --absent confirming 「marker absent from all 216 built files」 plus a clean whole-tree porcelain; post-restore re-run 8/8 and 5/5 green.

⭐ And the pins are written independently of objectPermissionGrants, from the rule itself, on purpose: 「asserting consumer === helper would be a tautology that survives any change to the fold.」 That is the one design error this card could have shipped, named and avoided.

The enforcement door, before and after — and 「before」 is a MEASUREMENT

checkObjectPermission is exercised by exactly 5 test files, all in plugin-security; three further files name it in prose only and call nothing — a distinction drawn rather than assumed.

The four pre-existing files: 334 passed BEFORE → 334 passed AFTER — ⛔ nothing removed, nothing weakened. Door total 334 → 339; build-access-matrix.test.ts 5 → 8.

⭐ 「Before」 was obtained by checking the two source files out at the merge base, removing the new test, running the four files against that tree, then restoring and proving byte-exactness (porcelain empty, git diff HEAD empty, each path's git hash-object equal to its blob at HEAD). ⇒ ⛔ not a remembered number.

Gates, changeset, route

• @objectstack/lint 4022 passed (106 files); @objectstack/plugin-security 2249 passed (117 files); typecheck exit 0 for both including the test layer; repo-wide pnpm lint exit 0 — the union, ⛔ not a narrowing.
• 63 derived / 63 run / 0 NOT MEASURED / 0 UNRUN, every code captured before any pipe. ⭐ Three gates first answered exit 3 (PREREQUISITE NOT MET — 53 packages with no dist; the CLI plus 9 extract-config closures; 14 workspace deps with no built type entry) and all three were DISCHARGED by one turbo run build, after which all three exit 0. ⇒ exit 3 treated as a prerequisite to satisfy, ⛔ not a verdict to report.
• ⭐ origin/main moved 4 commits mid-run; derivation re-run after a fresh fetch came back byte-identical.
• Changeset measured: both packages have private unset and files: [dist, README.md, CHANGELOG.md], both changed files build into dist ⇒ skip-changeset does ⛔ not apply ⇒ patch for both. And @objectstack/spec's ./security subpath carries both import and require conditions, so the new value import is safe on the CJS side of both dual builds — confirmed against real emitted bytes by check:dual-build-cjs-loads.
• check-governed-merges --pr 19512 exit 0 (0 of 5 paths, 394 lines); check-clause2-carriers --pair 19512 exit 0.
• ⛔ packages/lint/src/index.ts untouched — the serial hazard the dispatch named, and buildAccessMatrix was already exported so no edit was needed. packages/spec untouched.
• Zero label writes, zero POST /issues, zero MCP calls.

⭐ The card's SECOND 「not measured」 bullet is now answered — YES

Whether any consumer outside these three folds the same bits a fourth time.

It does. grantsObjectAccess (packages/lint/src/validate-security-posture.ts) carries a docblock claiming verbatim to mirror checkObjectPermission, and its disjunction omits verb targets the runtime gate honours.

⚠️ Verifying rather than transcribing, this seat found the hole one bit wider than reported: the dev named allowTransfer; reading OBJECT_PERMISSION_VERBS at its source shows the disjunction omits allowExport AND allowTransfer — two of the six verb targets. Filed as #19515, bare, with the Seam: line, the reproduction, the severity (warning, ⛔ does not gate the build) and both fix shapes — because 「is object-level CRUD meant to include those bits」 is a ruling about the rule's intent, ⛔ not a refactor. ⭐ The mirror claim is false either way; only which artefact changes is open. Carrier: PR #19486, which already edits that file.

⇒ ⭐ The dev did ⛔ not touch it — outside its surface and held by an open PR its dispatch order named. Correct on both counts.

The affinity question, answered from inside the work

Asked to say so if it thought #18785 and #18783 were one card, the dev said no, with a reason worth keeping: 「this card needed a MEASUREMENT and could finish the moment the differential came back zero, while #18783 waits on a DECISION; merging would have parked a finished measurement behind an open question.」 ⇒ triage's original split holds up from inside the work. ⛔ #18783 untouched, and ⛔ merging cards remains triage's call.

Route

Not governed ⇒ ordinary queue. ⚠️ ZERO check-runs on this head at this writing — ⛔ zero is not green, it is the absence of a reading, and Vercel=success on the commit-status surface does ⛔ not cover it. ⛔ Not flipped ready, ⛔ not armed. The watch treats NORUNS as its own phase.

---

Generated by Claude Code

[huangyiirene]

Release: session session_01NcPSwnmJHczmTu6FG7NMjE · 因 (reason): PR #19512 landed and Fixes #18785 closed this card, so the claim held for it is spent — written in the LANDED act, per the standing rule, on its fourth clean use · 去向 (destination): none — the card is closed completed and ⛔ nothing is handed on. The follow-up this round produced is #19515, filed separately and ⛔ not a remainder of this card.

---

LANDED — PR #19512. ⭐ One rule, one definition, and the convergence is proved on the trunk rather than asserted.

domain:engine#1, PM loop round 12, written 2026-09-21T06:09Z. Both readings are this seat's own, on a freshly fetched origin/main.

reading	result
squash on origin/main	ef256e6a8f9bf5744e071631f4337ded9c3b3b76 — %p word count 1, --is-ancestor origin/main ⇒ YES
content, POSITIVE	both consumers ASK the helper on the trunk — build-access-matrix.ts 6 references, permission-evaluator.ts 5 (base: 0 and 0)
content, NEGATIVE CONTROL	the inline fold that was replaced, permKey === 'allowRead' && …, counts 0 in the evaluator
⭐ LIT CONTROL — what was deliberately NOT collapsed	allowExport still appears 4 times in the evaluator ⇒ the cross-set conjunction survived the convergence, which is the whole point

⇒ the two restatements are gone and the one thing that must not have been folded is still there. A negative control alone could not have told those apart.

What this card actually bought

1. ⭐⭐ A measurement, not a refactor. The differential ran before any edit over the full declared input space — 3⁸ inputs × 6 verb targets, the enumeration derived from EffectiveObjectPermissionSchema.shape and OBJECT_PERMISSION_VERBS rather than hand-listed — and returned 0 disagreements across 65,610 cells. With two one-cell perturbations producing 2916 and 1944 disagreements, so the zero is a discrimination.
2. The convergence then became safe to make, and it was made without touching packages/spec, without touching packages/lint/src/index.ts (the serial hazard the dispatch named, held by open PR feat(lint): security-role-word crosses to the runtime publish gate, whole (#19370) #19486), and without narrowing the export door.
3. ⭐ An ablation on one spec cell reds BOTH consumers — 2 failures in lint, 3 in plugin-security — which is the only real proof that they ask it. And the pins are written independently of the helper on purpose, because 「asserting consumer === helper would be a tautology that survives any change to the fold」.
4. The enforcement door was measured before and after — 334 passed → 334 passed on the four pre-existing files, with 「before」 obtained by checking the sources out at the merge base and proving byte-exact restoration. ⇒ ⛔ nothing was removed or weakened.

⭐ The card's second 「not measured」 bullet, answered and filed

Whether any consumer outside these three folds the same bits a fourth time.

Yes — grantsObjectAccess. Filed as #19515, bare, with its Seam: line, the reproduction, the severity (warning, ⛔ does not gate the build) and both fix shapes. ⚠️ Verifying rather than transcribing, this seat found the hole one bit wider than reported: the disjunction omits allowExport AND allowTransfer, two of the six verb targets, while its docblock claims verbatim to mirror checkObjectPermission. ⭐ The mirror claim is false either way — only which artefact changes is open, and that is a ruling about the rule's intent, ⛔ not a refactor. Carrier: PR #19486, which already edits that file.

The affinity with #18783, answered from inside the work

The dev was asked to say so if it judged them one card. It said no: 「this card needed a MEASUREMENT and could finish the moment the differential came back zero, while #18783 waits on a DECISION; merging would have parked a finished measurement behind an open question.」 ⇒ triage's original split holds up. ⛔ #18783 untouched; merging cards remains triage's call.

Closeout

Fixes #18785 closed this card on the merge (completed) — ⛔ an auto-close does not remove the state label. Taken off by hand in the same act: pm:dispatched removed, assignee cleared, read back clean ⇒ priority:p2 · domain:engine, 0 assignees. ⛔ Grading labels are triage's and untouched.

⇒ Regions packages/lint/src/build-access-matrix.ts and packages/plugins/plugin-security/src/permission-evaluator.ts are RELEASED.

---

Generated by Claude Code