fix(security)!: under a wall the security catalog is the platform operator's — tenants are refused catalog writes and may only assign (ADR-0131 D3)

[objectstack-fleet]

Part of #15204

Clause-②: yes (narrowing)

Stage W of #15204 (claim amendment 6112750171): the walled catalog refusal. This PR builds the refusal ADR-0131 D3 names. The C3 card's seeder and catalog-object retirement stays open on #15204 and is not part of this PR.

The ruling this executes

ADR-0131 D3, verbatim: "Shared-DB multi-tenant (group / isolated): tenant admins cannot create or edit catalog items; the operator defines the catalog for every tenant (managed packages, Studio). Creation through Setup or the API is refused with a message naming the posture and the capability." ADR-0131 D7 adds that the catalog pages are "read-only under a wall". The server refuses whatever the UI shows. This is the refusal D3 names, not a new gate. Catalog writes stay with manage_metadata holders under single (#22621 → A, unchanged). Under a wall the authority is the platform operator's (#22801 → A).

Measured first, on main at 40bc51f

The harness was a booted examples/app-crm stack (bootStack, multiTenant: 'posture-only', which gives a real non-degraded posture). It had two organizations created over POST /auth/organization/create. Organization A's owner was granted a set carrying manage_metadata, scoped to organization A. That owner is not the platform operator (isPlatformAdmin: false). Organization B's member held an environment set, w_shared, scoped to organization B.

Door	isolated	group
A: PUT /meta/permission/w_shared (adds manage_users)	200	200
B's next resolution (resolveUserAuthzGrants)	gains manage_users	gains manage_users
A: PUT /meta/position/:name and PUT /meta/permission/:name (create)	200	200
A: DELETE /meta/permission/:name	200	200
A: PUT /meta/capability/:name	403 (capability is code-only)	403
A: POST /security/_activation/permission/w_shared	403 (operator-only already)	403
A with CRUD on sys_permission_set: POST /data/sys_permission_set	201, lands in the environment ledger	n/a
A with CRUD on sys_position: POST /data/sys_position	201, an organization row with no definition	n/a
A: POST /data/sys_user_permission_set (assignment)	201	n/a

Cross-tenant write: yes. One organization's PUT /meta changed what another organization's members resolve. The precondition is that a principal below the operator rung holds manage_metadata. A stock organization_admin does not hold it, and the anti-escalation deny in default-permission-sets.ts refuses that admin the catalog tables at the data door. The data door's edit of an existing environment set answered 404, because the wall hides the organization-less projection row. So the data door created environment sets but did not reach another organization's set.

What changed

• One predicate: walledCatalogWriteVerdict in @objectstack/metadata-core, beside metaWriteCapabilityVerdict. Under group / isolated, a write of a position, permission or capability is admitted only for isSystem or the PLATFORM_ADMIN posture rung. The activation gate already reads that rung for the same reason. It is derived from the platform-admin anchors, never from a held capability. Under single, or with no tenancy service, the predicate is inert. URL spellings are folded through canonicalMetaUrlType, the same fold saveMetaItem applies, so /meta/permissions/x is judged as permission.
• The /meta doors ask it right after the capability gate. On RestServer that is save, reset, publish and rollback (walledCatalogRefusal), answering 403 FORBIDDEN. On the runtime dispatcher it is PUT, answering 403 PERMISSION_DENIED. The posture is read only for a catalog type. The dispatcher reads it through readEffectiveTenancyPosture, factored out of the activation gate unchanged. REST reads it through the existing resolveAnonymousFormTenancy wiring, so the posture and the protocol come from one kernel. A tenancy service that is registered but cannot answer is a 503, never a permit.
• The data door: walled-catalog-data-door.ts (plugin-security), a middleware on sys_position, sys_permission_set and sys_capability. It runs inside the security middleware and before the permission-set write-through, so a refused write is never translated into a metadata save. Assignments (sys_user_position, sys_user_permission_set) are not judged.
• ADR-0087: the semantic entry walled-catalog-tenant-writes-retired. The changeset carries FROM → TO.
• Docs: system-context census row 55b, with the counts regenerated.

Why not saveMetaItem

The amendment named "the metadata door's write path, so every door that saves through saveMetaItem inherits it". Measured: saveMetaItem receives no caller principal, only an actor string. It cannot tell the operator rung from a tenant without a contract change on every caller. So the predicate sits where the caller is known: each door, after its capability gate. That is the activation gate's two-step shape (refuseUngrantedActivationAuthoring, then refuseUngrantedActivationWrite). It is one predicate with one sentence, and no door restates it.

The clone door (#22881, stage CD)

CD landed first (45e206e4), and this PR merged it in (d2a6934e5). The clone door creates environment permission sets, so this PR adds the same call there:

• permission-set-clone.ts asks walledCatalogWriteVerdict({ type: 'permission', ... }) right after metaWriteCapabilityVerdict, with the posture read through readEffectiveTenancyPosture.
• The call comes before any catalog read, so a refused caller learns nothing about the source or the target name. The answer is 403 PERMISSION_DENIED, the dispatcher PUT's envelope.
• Pins:
  • permission-set-clone.test.ts: isolated and group are refused with nothing saved; the operator rung is admitted; a single control; the capability gate still runs first.
  • Two dogfood legs: organization A's clone answers 403 with no definition saved, and an operator control answers 200.

Re-judged pins

• position-write-through.test.ts: the old walled pin was "an organization administrator's create lands its organization row, and no definition". That is the behaviour D3 refuses, so it now pins the refusal. The promotion pin seeds the pre-existing row-only position with a system write: that population is feat(core,objectql,plugin-security,plugin-sharing): the catalog is read from the registry; assignment tables reference it by name (ADR-0131 D2/D3/D4) #15196 Q3 = A's.
• builtin-positions.boot.test.ts: under a wall, the golden's "control: create a position row" is now 403 PERMISSION_DENIED. The assignment cases do not move.

Tests (first round at de89dd413; patch round 1 at code head 26c70abf8; every later commit up to 637f1269b is text only: docs, the changeset, the spec entry's strings, one .describe() string and its regenerated reference page)

• metadata-core 17 files / 386 tests, plus 8 new predicate cases. plugin-security 198 files / 4057 passed. rest 278 files / 5542 passed. runtime 355 files / 5831 passed. spec src/migrations 205 passed. Build and typecheck green for metadata-core, rest, runtime and plugin-security.
• New pins:
  • wall-catalog-refusal.dogfood.test.ts: isolated and group, plus a single control, 13 tests.
  • walled-catalog-data-door.test.ts: real ObjectQL and SecurityPlugin, 12 tests.
  • meta-walled-catalog-refusal.test.ts: the dispatcher, 11 tests.
• Ablation, REST door: walledCatalogRefusal was forced to admit, rest rebuilt, and the marker confirmed in dist/ by ablation-dist-preflight. The 4 /meta dogfood pins went red, on both postures. The data-door and control pins stayed green, as expected. Restored to the HEAD blob and rebuilt; the marker is absent from dist/.
• Ablation, data door: the middleware registration was removed. 6 walled refusal pins went red, and the assignment, operator and single controls stayed green. Restored to the HEAD blob.
• dispatch-gates --ran on 5354b1601: 125 derived, 124 run green, 1 not measured. check:dual-build-cjs-loads exited 3 (PREREQUISITE NOT MET) because packages outside this diff are unbuilt here.
• Ablation, clone door: the if (!walled.allowed) refusal was made unreachable. The isolated and group clone refusal pins went red (2 failed, 20 passed). Restored to the HEAD blob.
• permission-set-clone.test.ts with meta-walled-catalog-refusal.test.ts: 33 passed. The dogfood files wall-catalog-refusal and permission-set-clone-door: 21 passed. check:system-context-census: OK, 129 sites (the clone door's isSystem read moved it from 128).
• Dogfood suite in full on the merged head: 251 files passed, 1 skipped (the report of patch round 1 on refactor(plugin-security,platform-objects,spec): retire the catalog seeders, the per-organization catalog machinery and the four catalog objects; Setup creation is an environment write under single and refused under a wall (ADR-0131 D2/D3/D5/D13) #15204, 6114447595).

Acceptance notes

• Not filed (no reach measured): under a wall, an organization-scoped manage_metadata holder also writes the environment's objects, views and flows through /meta. D3 governs only the catalog. No card is filed, because only the catalog was measured.
• Not covered here: under a wall, ADR-0090 D12's delegated authorEnvironmentSets scope can no longer author environment sets at the data door. D3 refuses tenants catalog creation, and a delegate is a tenant. The delegated assignment and binding paths are untouched. sys_position_permission_set is not judged: under ADR-0131 D3 its rows bind nothing, and it retires with the catalog objects.

---

Generated by Claude Code

[github-actions]

📓 Docs Drift Check

This PR changes 5 package(s): @objectstack/metadata-core, @objectstack/plugin-security, @objectstack/rest, @objectstack/runtime, @objectstack/spec, touching 30 documentable anchor(s).

15 hand-written doc(s) NAME something this change touched and may need an implementation-accuracy re-verification:

• content/docs/api/client-sdk.mdx (via meta.publishItem (sdk, the route ledger binds it to POST /api/v1/meta/:type/:name/publish, selected by route anchor /:type/:name/publish), meta.rollbackItem (sdk, the route ledger binds it to POST /api/v1/meta/:type/:name/rollback, selected by route anchor /:type/:name/rollback), publishItem (sdk, the bare tail of client method meta.publishItem, bound to POST /api/v1/meta/:type/:name/publish), rollbackItem (sdk, the bare tail of client method meta.rollbackItem, bound to POST /api/v1/meta/:type/:name/rollback))
• content/docs/automation/flows.mdx (via PLATFORM_ADMIN (literal, a string literal in PLATFORM_OPERATOR_POSTURE))
• content/docs/concepts/metadata-lifecycle.mdx (via sys_permission_set (symbol, a field of const object WALLED_CATALOG_OBJECT_TYPES), /security/permission-sets/:name/clone (route, a path literal in surface))
• content/docs/deployment/environment-variables.mdx (via sys_permission_set (symbol, a field of const object WALLED_CATALOG_OBJECT_TYPES), sys_position (symbol, a field of const object WALLED_CATALOG_OBJECT_TYPES), PLATFORM_ADMIN (literal, a string literal in PLATFORM_OPERATOR_POSTURE))
• content/docs/kernel/contracts/metadata-service.mdx (via /:type/:name/publish (route, bridged from symbol walledCatalogRefusal — its route source's handler names it), /:type/:name/rollback (route, bridged from symbol walledCatalogRefusal — its route source's handler names it))
• content/docs/permissions/authentication.mdx (via PLATFORM_ADMIN (literal, a string literal in PLATFORM_OPERATOR_POSTURE))
• content/docs/permissions/authorization.mdx (via sys_capability (symbol, a field of const object WALLED_CATALOG_OBJECT_TYPES), sys_permission_set (symbol, a field of const object WALLED_CATALOG_OBJECT_TYPES), sys_position (symbol, a field of const object WALLED_CATALOG_OBJECT_TYPES), PLATFORM_ADMIN (literal, a string literal in PLATFORM_OPERATOR_POSTURE))
• content/docs/permissions/capabilities.mdx (via sys_capability (symbol, a field of const object WALLED_CATALOG_OBJECT_TYPES))
• content/docs/permissions/delegated-administration.mdx (via sys_permission_set (symbol, a field of const object WALLED_CATALOG_OBJECT_TYPES), sys_position (symbol, a field of const object WALLED_CATALOG_OBJECT_TYPES))
• content/docs/permissions/permission-metadata.mdx (via PLATFORM_ADMIN (literal, a string literal in PLATFORM_OPERATOR_POSTURE))
• content/docs/permissions/permission-sets.mdx (via sys_capability (symbol, a field of const object WALLED_CATALOG_OBJECT_TYPES), sys_permission_set (symbol, a field of const object WALLED_CATALOG_OBJECT_TYPES), PLATFORM_ADMIN (literal, a string literal in PLATFORM_OPERATOR_POSTURE), /security/permission-sets/:name/clone (route, a path literal in surface))
• content/docs/permissions/positions.mdx (via sys_position (symbol, a field of const object WALLED_CATALOG_OBJECT_TYPES))
• content/docs/permissions/system-context.mdx (via createWalledCatalogDataDoorRefusal (symbol, a top-level function), handleMetadataRequest (symbol, a top-level function), handlePermissionSetCloneWrite (symbol, a top-level function), refuseUngrantedActivationWrite (symbol, a top-level function), sys_capability (symbol, a field of const object WALLED_CATALOG_OBJECT_TYPES), sys_permission_set (symbol, a field of const object WALLED_CATALOG_OBJECT_TYPES), sys_position (symbol, a field of const object WALLED_CATALOG_OBJECT_TYPES), walledCatalogRefusal (symbol, a method of class RestServer), walledCatalogWriteVerdict (symbol, a top-level function))
• content/docs/protocol/backward-compatibility.mdx (via sys_position (symbol, a field of const object WALLED_CATALOG_OBJECT_TYPES))
• content/docs/protocol/kernel/error-handling.mdx (via /:type/:name/publish (route, bridged from symbol walledCatalogRefusal — its route source's handler names it))

⛔ 14 release-owned page(s) also name something this change touched. These are read-only:

• content/docs/releases/implementation-status.mdx (via RestServer (symbol, a top-level class))
• content/docs/releases/index.mdx (via sys_position (symbol, a field of const object WALLED_CATALOG_OBJECT_TYPES))
• content/docs/releases/v12.mdx (via RestServer (symbol, a top-level class), sys_permission_set (symbol, a field of const object WALLED_CATALOG_OBJECT_TYPES))
• content/docs/releases/v13.mdx (via sys_permission_set (symbol, a field of const object WALLED_CATALOG_OBJECT_TYPES), sys_position (symbol, a field of const object WALLED_CATALOG_OBJECT_TYPES))
• content/docs/releases/v14.mdx (via sys_position (symbol, a field of const object WALLED_CATALOG_OBJECT_TYPES))
• content/docs/releases/v15.mdx (via sys_capability (symbol, a field of const object WALLED_CATALOG_OBJECT_TYPES), sys_permission_set (symbol, a field of const object WALLED_CATALOG_OBJECT_TYPES), sys_position (symbol, a field of const object WALLED_CATALOG_OBJECT_TYPES), PLATFORM_ADMIN (literal, a string literal in PLATFORM_OPERATOR_POSTURE))
• content/docs/releases/v16.mdx (via RestServer (symbol, a top-level class), PLATFORM_ADMIN (literal, a string literal in PLATFORM_OPERATOR_POSTURE))
• content/docs/releases/v17/17-0.mdx (via AdminScopeSchema (symbol, a top-level const), sys_permission_set (symbol, a field of const object WALLED_CATALOG_OBJECT_TYPES), /:type/:name/publish (route, bridged from symbol walledCatalogRefusal — its route source's handler names it))
• content/docs/releases/v17/17-1.mdx (via sys_permission_set (symbol, a field of const object WALLED_CATALOG_OBJECT_TYPES), sys_position (symbol, a field of const object WALLED_CATALOG_OBJECT_TYPES), PLATFORM_ADMIN (literal, a string literal in PLATFORM_OPERATOR_POSTURE), /:type/:name/publish (route, bridged from symbol walledCatalogRefusal — its route source's handler names it))
• content/docs/releases/v17/17-2.mdx (via sys_position (symbol, a field of const object WALLED_CATALOG_OBJECT_TYPES), /:type/:name/publish (route, bridged from symbol walledCatalogRefusal — its route source's handler names it))
• content/docs/releases/v17/17-3.mdx (via RestServer (symbol, a top-level class))
• content/docs/releases/v17/17-4.mdx (via /:type/:name/publish (route, bridged from symbol walledCatalogRefusal — its route source's handler names it))
• content/docs/releases/v17/17-5.mdx (via RestServer (symbol, a top-level class), sys_permission_set (symbol, a field of const object WALLED_CATALOG_OBJECT_TYPES), PLATFORM_ADMIN (literal, a string literal in PLATFORM_OPERATOR_POSTURE))
• content/docs/releases/v17/17-6.mdx (via sys_permission_set (symbol, a field of const object WALLED_CATALOG_OBJECT_TYPES))

content/docs/releases/ is RELEASE-OWNED (AGENTS.md "Documentation Guardrails"): release
notes are written centrally at release time, and a code PR that edits them is the exact PR
that guardrail exists to stop. They are still audited — read-only. If one of them is actually
wrong, file an issue or open a dedicated docs-only PR; do not edit it here.

What this run could not see

• 7 name(s) were too generic to anchor anything (single lowercase words)
• the SDK route bridge reached 54 of 206 client-bound route-ledger rows — the other 152 have no registrar path: tail to select them, so pages documenting THEIR client methods cannot appear above, on this or any run. Of those 152: 0 are remediable by widening that discovery convention (an in-repo file declares the path; the convention did not scan it); 55 are structural — on a ledger where NOT ONE row is declared in-repo, so no discovery change reaches them at any price; 97 are undecided (no in-repo declaration, on a ledger that has other in-repo registrars — absence and an unreadable spelling are not distinguishable here). The rows themselves: node scripts/docs-audit/affected-docs.mjs --bridge-coverage
• a page that states a rule by its inputs shares no identifier with the emitter that implements the rule, so an emitter-only diff cannot list it — not on this run and not on any run. Measured on fix(driver-sql): emit varchar(maxLength) for a text field a declared index keys on #11430: content/docs/protocol/objectql/types.mdx documents the text-family column mapping by the ObjectQL type names it maps FROM (text / textarea / html) while the diff changed createColumn; it went unlisted, and it was the page that diff falsified, in four places. No shared token exists to detect this on, so a rule your change carries has to be re-read by hand in the pages that restate it.
• a key NAME is not a key, so the hand re-read the line above prescribes can land on the wrong schema. The same spelling is authorable on one governed type and a [REMOVED] tombstone on another for each of active, aria, joins, objects, template, tools and version (censused on [finding] tools is a key on BOTH AgentSchema (tombstoned, dead) and SkillSchema (live, cloud-attested), so a name-based search attributes skill examples to the agent key — it produced a false stop-the-line alarm on PR #19059 #19093 over the liveness ledger's governed types, top-level keys); nothing in a search result distinguishes the two, so a grep hit on a LIVE example reads as evidence about the DEAD key. Measured on fix(spec): the agent.tools liveness row says dead — it claimed live on a key the schema tombstoned #19059: content/docs/ai/agents.mdx was reported as contradicting the agent.tools tombstone over its tools: example at :161, which is inside the defineSkill({ block opened at :155 — the page was already correct. Settle ownership by PARSING the value against both schemas, never by the name: that literal PASSES SkillSchema, and as an AgentSchema it FAILS at tools with the tombstone prescription. ⛔ These names are not the whole class — a key retired through a .strict() guidance map leaves no tombstone in the walked shape and none of them here (tool.category, live as AIToolDefinition.category).

Coarse fallback — 148 page(s) merely mention a changed package (the pre-#9192 predicate, kept for the deliberately-wide backstop): node scripts/docs-audit/affected-docs.mjs --json 018b077fbf8876c0fd5f46e66dcc551a4e74c895 → packageMentionDocs.

Which tree this was computed on

This run read content/docs from 74a360f307087ca6031391273c8f8e1e216841db — the merge of head b1f2335b29d12e5494247d1a2bab6796da6fad8c into base 018b077fbf8876c0fd5f46e66dcc551a4e74c895, which is what actions/checkout gives a pull_request run. Not the PR head.

A worktree cut from an older main holds a different content/docs, so re-deriving there can legitimately return a different list — that is a different tree, not a wrong row. To answer on the same tree:

# while this PR is open — GitHub drops the merge commit once it closes
git fetch origin 74a360f307087ca6031391273c8f8e1e216841db && git checkout 74a360f307087ca6031391273c8f8e1e216841db
# afterwards, rebuild it from the two parents, which stay fetchable
git fetch origin 018b077fbf8876c0fd5f46e66dcc551a4e74c895 b1f2335b29d12e5494247d1a2bab6796da6fad8c && git checkout -B drift-repro 018b077fbf8876c0fd5f46e66dcc551a4e74c895 && git merge --no-ff b1f2335b29d12e5494247d1a2bab6796da6fad8c

node scripts/docs-audit/affected-docs.mjs --json 018b077fbf8876c0fd5f46e66dcc551a4e74c895

⚠️ That checkout carried uncommitted changes, so the commit above does not fully identify what was read.

Advisory only, and a precision-first one (#9192): a page is listed because it names a
symbol, wire route or SDK method this diff touched — not because it mentions a changed
package. Each row says which anchor put it there, so a wrong row is reportable rather than
merely annoying. To re-verify, run the docs-accuracy-audit workflow scoped to these files:
node scripts/docs-audit/affected-docs.mjs 018b077fbf8876c0fd5f46e66dcc551a4e74c895 → pass the list as
args.docs, on the commit named under Which tree this was computed on.

[objectstack-fleet]

Contract review

Served-tier: CONTRACT_REVIEW_TIER
Head-sha: 5354b160163455c1558da61a18a66c905b4a310e
Local-runs: none

First record on this PR (stage W of #15204, the walled catalog refusal, ADR-0131 D3). Reviewed read-only: the PR body and its 19-file list (+1,196 / −29), the net diff of 5354b160 against main (base 45e206e4, the three-dot diff after the two origin/main merges 376af552 and d2a6934e), card #15204's body and all 138 comments through 6114486311 (the stage-W seat note 6112750171 with its five measurement questions, the measurement report 6113821871, the seat note 6113841670 with release-cut condition (f) and the patch round, the patch-round report 6114447595, the rulings 6095027530 (#22621 → A) and the triage pointer 6070695107), ADR-0131 D3 / D4 / D7 / D11 / D13, ADR-0090 D12 and ADR-0087 (the launch-window section and the disposition addendum) as served on main, and on main the activation gate, the four REST /meta write doors, the dispatcher's /meta branch, permission-set-clone.ts, security-plugin.ts's middleware roster, resolve-authz-context.ts's rung derivation, the two re-judged test files and the ObjectQL middleware contract. main read at 63739521d8, which contains the PR base.

Gate verdicts on the head, read at 2026-10-11T22:36Z. 42 check-runs across two workflow rounds, every one completed: 38 success, 4 skipped (the second round's Auto Label and Check PR Size, which the first round ran green, and the two roster skips Console Pin Gate and Packed-tarball smoke (opt-in)). Nothing is pending. Green among them: Lint & Repo Gates (which runs check:system-context-census), Check Changeset (which runs check-adr-0087-registration.mjs against the merge base and the no-major guard), TypeScript Type Check with its four sub-gates, Test Core 1/6 to 6/6, Dogfood Regression Gate 1/3 to 3/3, Build Core, Build Docs, Temporal Conformance, Governed Surface Queue Guard, Spec property liveness, Check Documentation Links, and the five PR-form guards. No review is submitted on the PR; it is a draft, mergeable true, no auto-merge armed. Where a judgment below says "pinned", the pin is one of these green runs, not a local run.

① Derived judgments

Every accept-set and public-surface change the diff implies, each named right or wrong.

1. The doors (brief point a). On main exactly six caller-facing doors write a position, permission or capability item, read off the callers of metaWriteCapabilityVerdict and of saveMetaItem plus the engine writers of the three tables: REST PUT (save), DELETE (reset), POST …/publish and POST …/rollback on /meta/:type/:name; the dispatcher's PUT (its DELETE / POST / PATCH on an item answer 405 with an Allow header, so it has one write door); the clone door POST /security/permission-sets/:name/clone; and the data door on sys_position, sys_permission_set, sys_capability. The diff adds walledCatalogWriteVerdict after the capability gate at every one of them: the four REST sites (rest-server.ts save, reset, publish, rollback, each before resolveProtocol), the dispatcher's PUT branch (before anything is read or written), the clone door (before any catalog read), and a middleware registered per catalog object. The activation door (/security/_activation/...) already refuses below the operator rung on main (refuseUngrantedActivationWrite) and is left alone — right. The data-door middleware judges insert, update, delete, which is the whole write vocabulary the ObjectQL middleware chain carries (find, findOne, count, aggregate are the rest; no bulk or upsert operation exists), so no write verb slips past it — right. It is registered after the security middleware (:2323, which also hosts the ADR-0090 D12 delegated-admin gate) and after the position catalog refusal, before the permission-set write-through and the position write-through, so a refused write is never translated into a metadata save — right, and the ablation the dev reports (registration removed, six refusal pins red, controls green) is the measured form of the same claim. Engine writers that remain are permission-set-projection.ts (projector, SYSTEM_CTX), position-environment-backfill.ts (migration), the two write-throughs (behind the new middleware) and the registry's registerItem calls in builtin-positions.ts / builtin-capabilities.ts (boot declarations) — all system writes, all admitted by isSystem. No door is left out. Two adjacent doors I considered and judge outside D3's catalog-write set are in ③ (items 8 and 9).

2. The predicate (brief point b). walledCatalogWriteVerdict admits everything outside SECURITY_CATALOG_TYPES, admits when the posture is undefined or not wall-enforcing (postureEnforcesWall from @objectstack/spec/security, the activation gate's own test), admits isSystem === true, admits callerPosture === 'PLATFORM_ADMIN', refuses the rest — right. Inert under single and with no tenancy service: pinned (meta-write-capability.test.ts, the dispatcher's single and no-service controls, the dogfood single control where a manage_metadata holder who is not the platform administrator still saves a position, [Decision] 切换后谁能在 Setup 里新建/编辑岗位和权限集：保持今天的权限（持 manage_metadata 者），还是放开给组织管理员（安全边界） #22621 → A unchanged). The operator rung is the context's posture, which resolve-authz-context.ts derives through derivePosture({ isPlatformAdmin: hasPlatformAdminGrant }), and hasPlatformAdminGrant is set by exactly two routes: the configured OS_PLATFORM_OWNER_EMAIL match, and the unscoped admin_full_access grant only while legacyGrantAnchorRetired is false, i.e. never under a wall. No held capability and no position name reaches it; a context carrying no rung is refused (pinned) — right, and it is the activation gate's ec?.posture === 'PLATFORM_ADMIN' test verbatim. URL spellings: the predicate and isSecurityCatalogType fold through canonicalMetaUrlType from @objectstack/spec/meta-spelling, the one function metadata-protocol/src/protocol.ts:342 applies at the saveMetaItem boundary (spec/src/meta-spelling/metadata-url-spelling.ts:161, re-exported by shared), so /meta/permissions/x is judged as permission — right, pinned on the dispatcher. A registered tenancy service that cannot answer: on REST the read goes through resolveAnonymousFormTenancy, which returns undefined only for the never-registered brand and otherwise throws AuthzStoreUnavailableError('tenancy'), which the route's handleRouteError answers with the error's own status; on the dispatcher and the clone door readEffectiveTenancyPosture throws the same error, which is the activation gate's body moved into a named function and called from its old site — right: a throw is never a permit. At the data door the posture is the security plugin's one start-time read (this.tenancyPosture), which resolves a tenancy service that answers nothing to single; that is not a 503, and the PR body's "503, never a permit" sentence is true of the three transport doors, not of the data door. I judge it right nonetheless: that read is the same one that decides whether Layer 0 walls any object at all, so a deployment whose plugin reads single enforces no wall for this refusal to protect, and reading the posture twice through two mechanisms in one plugin would be the drift the activation gate's comment warns against. Named, not failed.

3. The answers (brief point c). REST /meta answers 403 FORBIDDEN; the dispatcher PUT, the clone door and the data door answer 403 PERMISSION_DENIED. Read against main: each of the four REST doors' own capability refusal already answers res.status(403).json({ error: { code: 'FORBIDDEN', … } }); the dispatcher's capability refusal is deps.error(verdict.message, 403), whose code buildApiError derives from the status through standardErrorCodeForHttpStatus, PERMISSION_DENIED ("this transport's pinned spelling", the branch's own comment); the clone door uses the same deps.error; the data door throws PermissionDeniedError, the plugin's one refusal class. So the split is the pre-existing per-transport envelope, and the diff puts each new refusal in exactly the envelope the door's sibling refusal already uses — right. A refusal that invented a third spelling, or made the two transports agree by changing one of them, would have been the defect; neither happened. The changeset and the semantic entry name both spellings.

4. The accept-set narrowing beyond the measured cross-tenant write (brief point d). ADR-0090 D12's delegated scope may "optionally author environment-owned sets"; its authorEnvironmentSets arm lives in delegated-admin-gate.ts (:906) inside the security middleware, and admits a sys_permission_set write at the data door only. After this PR such a delegate under a wall passes the delegated gate and is refused by the walled middleware behind it. Within D3: yes — D3's rule is by rung and posture ("tenant admins cannot create or edit catalog items; the operator defines the catalog for every tenant"), a delegate is a principal below the operator rung inside one organization, and ADR-0131 is the later record. Does the changeset's FROM → TO say it: in population terms, yes — its data-door bullet reads "FROM admitted to any caller whose grants admit the table TO 403 PERMISSION_DENIED", and the delegated gate is what admits the table to a delegate; by name, no — neither the changeset nor the semantic entry's surface spells authorEnvironmentSets or ADR-0090 D12; only the PR body's Acceptance notes do. Judged right for the verdict (the accept-set change is declared and the level is BREAKING already); the missing name is carried in ③ item 4 for the seat.

5. Assignments and engine writes (brief point e). WALLED_CATALOG_OBJECT_TYPES is exactly the three catalog objects (pinned, "assignment objects are never judged"); sys_user_position and sys_user_permission_set have no registration. Controls pinned: the walled tenant still assigns a catalog position (walled-catalog-data-door.test.ts, both postures), still assigns a permission set through POST /data/sys_user_permission_set on the booted stack (dogfood, 201), and the boot golden's assignment cases do not move. Engine writes: isSystem admitted at the predicate (pinned), a system sys_position insert admitted at the data door under a wall (pinned), the position write-through's "a system write is never translated" pin untouched, the operator saves and clones under both postures (dogfood controls, 200). The grant-by-name door (feat(plugin-security,plugin-auth,verify,platform-objects)!: a permission-set grant names its set by name — no writer fills permission_set_id (#15204 stage 6b-2-pre) #22863) writes GRANT_OBJECT = 'sys_user_permission_set' and only reads the catalog; the suggestion-confirm door is already behind assertDeploymentAuthority and since stage 1 writes a junction row that grants nothing — right.

6. The ADR-0087 entry and the changeset (brief point f). 18.walled-catalog-tenant-writes-retired.ts carries the five SemanticMigration fields (id, surface, replacement, reason, acceptanceCriteria); surface avoids backticks as the generator requires; entries are discovered one-file-per-entry under entries/semantic/, and Check Changeset verified the id resolves and is new in the diff. The changeset: five packages at minor (metadata-core, rest, runtime, plugin-security, spec), Clause-②: yes (narrowing) on its own line, the **BREAKING** banner, the marker adr-0087: registered walled-catalog-tenant-writes-retired, a FROM → TO per door, an Unchanged paragraph, and the remedy (the operator defines in Studio or a managed package; tenants assign; a custom door asks the predicate after metaWriteCapabilityVerdict) — right; ② judges the levels. The dev's deviation 4 (the gate refuses a FROM → TO body beside not-required, so the change is registered) is the ADR-0087 addendum's own rule ("no-migration-prescription is refused by a body that carries a migration prescription") — right.

7. Census row 55b and the counts (brief point g). The regenerated counts move 124 → 129 reads, which is the five isSystem reads the diff adds (the predicate's input, walledCatalogRefusal, the dispatcher branch, the data-door middleware, the clone door); files 58 → 59 (one new file, walled-catalog-data-door.ts; the other four files were already anchored), symbols 106 → 109 (walledCatalogWriteVerdict, walledCatalogRefusal, createWalledCatalogDataDoorRefusal; the dispatcher's and the clone door's reads sit in symbols already anchored), declarations 28 → 29 (the predicate's isSystem?: boolean), and the row's anchors name those three symbols. The arithmetic closes and Lint & Repo Gates holds it — right. The two new scripts/adr-anchors/*.json files anchor ADR-0131 on the predicate's file and the data door's file — right.

8. The two re-judged pins (brief point g). position-write-through.test.ts: the old pin "a walled posture: an organization administrator's create lands its organization row, and no definition" is the behaviour D3 refuses; the new pin asserts 403 PERMISSION_DENIED, zero rows, saveMetaItem not called and zero environment rows — strictly more is asserted, nothing is dropped: the D3 refusal, not a weakening. Its promotion pin now seeds the pre-existing row-only position with an isSystem insert instead of a tenant create and keeps every downstream assertion (catalog does not resolve, the metadata-door definition promotes it, permissionSets read back); that population is feat(core,objectql,plugin-security,plugin-sharing): the catalog is read from the registry; assignment tables reference it by name (ADR-0131 D2/D3/D4) #15196 Q3 = A's, as the comment says — right. builtin-positions.boot.test.ts: REFUSAL_GOLDEN's "control: create a position row" was { ok: true } for every scenario; refusalGoldenFor now overrides it to { ok: false, code: 'PERMISSION_DENIED', status: 403 } when scenario.posture !== 'single' (the Scenario.posture union is 'single' | 'isolated', so the walled scenario and its three variants), and leaves the assignment rows, the reference_not_found cases and the two built-in-row refusals untouched — the D3 refusal, not a weakening. group is not in that harness and is pinned in walled-catalog-data-door.test.ts instead.

9. Public surface. @objectstack/metadata-core gains four exports (walledCatalogWriteVerdict, isSecurityCatalogType, SECURITY_CATALOG_TYPES, PLATFORM_OPERATOR_POSTURE), named in the changeset's remedy; @objectstack/runtime gains readEffectiveTenancyPosture from activation-gate.ts, a factoring of the gate's own body whose old call site now calls it (the refusal path, the AuthzStoreUnavailableError wrap and the undefined-posture early return are byte-for-byte the same decisions) — right, additive, and the message strings of the activation gate are unchanged.

10. The dispatcher reads the posture only for a catalog type (isSecurityCatalogType(type) guards the readEffectiveTenancyPosture call) and REST's helper returns { allowed: true } before the tenancy read for any other type, so a view or object save pays no tenancy read and cannot meet a tenancy 503 it did not meet before — right, pinned ("a type outside the catalog is not judged").

11. The capability gate still answers first at every door (pinned on the dispatcher, the clone door and the predicate's placement in REST after metaSaveVerdict), so a caller without manage_metadata learns the capability sentence, not the posture — right, and it is the activation gate's two-step shape the dispatch named.

② Semver level

• The changeset declares minor for @objectstack/metadata-core, @objectstack/rest, @objectstack/runtime, @objectstack/plugin-security and @objectstack/spec, with Clause-②: yes (narrowing). The diff publishes: new exports in metadata-core and runtime (additive, minor); an accept-set narrowing at the REST, dispatcher, clone and data doors ((narrowing) is BREAKING; ADR-0087's launch-window section rules that pre-GA "a metadata-facing retirement or break ships minor, carrying the **BREAKING** banner and its ADR-0087 disposition entry", and check-changeset-no-major.mjs refuses a major); a new semantic entry in spec (minor). The level matches what the diff publishes, the banner and the marker are present, and the Clause-②: yes (narrowing) line is in the changeset body and on the PR body's third line (the dev's measured answer in 6114447595 that readClause2Line scans every line stands; Check Changeset is green). The PR title carries the !.
• Packages touched without a changeset: @objectstack/qa/dogfood (private test package) and content/docs — owe none.
• Three wording imprecisions in the FROM → TO, none of which moves the level or the accept set, carried to ③ item 4: (i) for capability at the /meta doors the FROM clause "admitted to any manage_metadata holder" overstates — the dev's own table measured PUT /meta/capability/:name as 403 on main (code-only type), so at /meta the capability row changes the sentence a tenant reads, not the accept set; the capability narrowing is real at the data door (sys_capability insert, pinned); (ii) the semantic entry's surface says "PUT, DELETE, publish and rollback … on @objectstack/rest and the runtime dispatcher", while the dispatcher has PUT alone (the changeset's own parenthetical says it right); (iii) the ADR-0090 D12 delegate is not named (①.4).

③ Boundary flags

Every dev flag and every open_questions entry, answered or escalated.

1. open_questions[0] (6113821871): which PR carries the clone door's call → A. The seat answered A in 6113841670 (CD landed first at 45e206e4; this PR, landing second, adds the call); the patch round did it (26c70abf8), with the isolated / group refusal pins, the operator and single controls, the capability-first pin, two dogfood legs and an ablation. Closed by this head.
2. Deviation 3 (both reports): the refusal is not in saveMetaItem. Measured and accepted by the seat (6113841670): saveMetaItem carries an actor string and no principal, so the dispatch's placement hint was falsified and the predicate sits where the caller is known. I confirm the reading on main (SaveMetaItemRequestSchema, spec/src/api/protocol.zod.ts:646: type, name, item, parentVersion, actor, force, mode, packageId — no principal, no posture). One predicate, asked at every door, no door restating it — the activation gate's shape. Right; closed.
3. Deviations 1, 2 (report form) and the patch round's two (body not PATCHed by the dev; the released-line answer's line position): process, not contract. The PR body has since been rewritten by the seat and now says CD landed first, names this PR as the carrier of the clone call, and dates the tests to 5354b1601; the dev's "stale body" out-of-scope finding is closed.
4. For the seat, text only, not a condition of this verdict: ②'s three wording imprecisions. A one-line addition naming ADR-0090 D12's authorEnvironmentSets delegate in the changeset's data-door bullet and in the entry's surface would spare an operator with a delegated scope the inference; the capability-at-/meta FROM and the entry's "runtime dispatcher" clause would read exactly with "where it was admitted" and "PUT on the runtime dispatcher". Should the seat take them, a text-only push changes the head and a delta record is owed on it.
5. Out-of-scope finding (both reports), carried to the maintainer by 6113841670: under a wall an organization-scoped manage_metadata holder also writes the environment's objects, views and flows through /meta, which reach every organization. D3 governs the catalog only, so this PR rightly does not touch it. Escalated, already on the seat's desk; not this stage's.
6. Out-of-scope finding: sys_position_permission_set is not judged. Right: since stage 1 (feat(spec,core)!: positions declare their permissionSets; the authorization resolver reads the security catalog and the activation ledger #22723) the binding is the position's definition (PositionSchema.permissionSets), a junction row grants nothing (6102862135), and the object retires with the catalog under D13. A tenant writing a row that binds nothing is not a catalog write. Closed.
7. The released line (6114447595): 17.7.0 is exposed by code read. Not a contract item of this diff; it is release-cut condition (f)'s substance (6113841670) and the seat's to carry. I note that the condition names the four door families and this head refuses at all four.
8. A door I considered and judge outside D3's set, named so the seat can disagree: POST /meta/_migrate-stored. manage_metadata-gated on REST, it replays the ADR-0087 chain over stored rows of every type, catalog rows included. It authors nothing: the bytes it writes are the chain's, identical for every caller, and the same rewrite happens on load. Not a create, change or delete of a catalog item in D3's sense; no action asked.
9. Escalated to the same carrier as item 5: the package door. handlePackagesRequest guards its writes with requireManageMetadata and reads no posture rung; a managed package declares positions, permission sets and capabilities, which boot registers as sealed catalog items (registerItem, isSystem). So under a wall the same population this PR refuses at /meta — a sub-operator manage_metadata holder — can still bring catalog definitions in through a package install, which D3 names as the operator's instrument. This was outside stage W's five measurement questions (6112750171 named /meta, the activation door and CD) and is the package-door face of item 5's question (what manage_metadata below the operator rung may author under a wall), so it belongs with that decision rather than in this PR. Not measured by me; a code read.
10. The data door's posture read (①.2). The PR body's sentence "a tenancy service that is registered but cannot answer is a 503, never a permit" holds for REST, the dispatcher and the clone door; at the data door the posture is the plugin's start-time read, shared with Layer 0. Judged right in ①.2; recorded here so nobody reads the sentence as covering four doors.
11. Size and tier. 19 files, +1,196 / −29: not governed (check-governed-merges paths untouched; the seat's 6113841670 says the same), under the 3,000-line class. Part of #15204, no closing keyword, assignee set, draft. ⛔ Not armed, not approved, not readied by this review; the seat's pre-landing check and the queue follow.

Implemented-by: session_015JgoXpZUGtNxnMDvRZsKzP
Reviewed-by: session_01Rerax7QTjKMPCUZxQUtPFR

VERDICT: PASS

---

Generated by Claude Code

[objectstack-fleet]

Contract review

Served-tier: CONTRACT_REVIEW_TIER
Head-sha: 637f1269b2ccbf52ffe706a0389f43b16f75b860
Local-runs: none

Delta record on this PR (stage W of #15204, the walled catalog refusal, ADR-0131 D3), after the same-form PASS 6114668919 on 5354b160163455c1558da61a18a66c905b4a310e. Reviewed read-only: the compare 5354b16016...637f1269b2 (3 commits, 7 files), the PR body, its 24-file list (+1,207 / −35), the net diff of 637f1269b2 against main (base 45e206e4; main read at 9fdb12752c), card #15204's body and all 147 comments through 6114897832 (among them the stage-W seat note 6112750171, the measurement report 6113821871, the seat note 6113841670, the patch-round report 6114447595, the seat note 6114679100 that sent this text round, the dev's addenda 6114710994 and 6114885744, the W2 seat note 6114823168, the ruling 6095027530 and the triage pointer 6070695107), ADR-0131 D3 / D4 / D7, ADR-0090 D12 and ADR-0087 (the launch-window section and the disposition addendum) as served on main, and on main: the activation gate, delegated-admin-gate.ts (the authorEnvironmentSets arm, assertSetAuthoring at :906), the metadata protocol's code-only guards (saveMetaItem :20021-20030, deleteMetaItem :26347-26353) and the publish / rollback envelope codes, rest-server.ts's door order, content/docs/permissions/capabilities.mdx :225-240, and the spec reference generator (packages/spec/scripts/build-docs.ts, lib/schema-section.ts :531-546, lib/escape-mdx.ts). The regenerated content/docs/references/security/permission.mdx was read on the head.

Gate verdicts on the head, read 2026-10-11T23:21Z. 34 check-runs: 29 completed — 27 success, 2 skipped (the roster skips Console Pin Gate and Packed-tarball smoke (opt-in)) — and 5 still in_progress: Test Core (1/6), (2/6), (3/6), (4/6) and (5/6), named here as pending; their conclusions are the gate verdicts when they land, and this record does not pre-empt them. Green among the completed: Type Check · source gates (the job that runs @objectstack/spec's check:docs, check:upgrade-guide, check:spec-changes and check:authorable-surface — the gates that read this delta's regenerated reference page, the entry's surface string and the schema), Lint & Repo Gates (check:system-context-census, check:doc-authoring and the docs guards), Check Changeset (check-adr-0087-registration.mjs against the merge base, the no-major guard, the Clause-② reader), TypeScript Type Check with Type Check · consumer gates, · debt ledger and · workspace, Build Core, Build Docs, Dogfood Regression Gate 1/3 to 3/3 and its aggregate, Dogfood Verify CLI, Temporal Conformance, Test Core (6/6), Governed Surface Queue Guard, Spec property liveness, Check Documentation Links, Flag docs affected by code changes, and the five PR-form guards. The delta changes no code the pending shards execute; the two spec suites that read its strings (src/migrations, 205 tests; src/security, 319) the dev reports green locally on 2243159ea1 and 637f1269b2 (6114710994, 6114885744), and the same shards were green on 5354b16016 for the same code. No review is submitted on the PR; it is a draft, mergeable true, mergeable_state blocked while they run. Where a judgment below says "pinned", the pin is one of the green runs above, not a local run.

① Derived judgments

Every accept-set and public-surface change the diff implies, each named right or wrong. Items 1 to 11 of 6114668919's ① are adopted on this head: the compare lists none of the files they judge, so the predicate, the four REST doors, the dispatcher PUT, the clone door, the data-door middleware and its registration, the activation-gate factoring, the re-judged pins, the census row and the two anchors are byte-identical between the two heads. The delta's own judgments follow.

1. The delta is text only (brief point a). compare/5354b16016...637f1269b2: ahead_by 3, behind_by 0 (linear on the reviewed head, no rebase), commits 2243159ea1, a337a2fa62, 637f1269b2, 7 files: .changeset/15204-walled-catalog-refusal.md (+3 / −2), content/docs/permissions/authorization.mdx (+3 / −1), content/docs/permissions/delegated-administration.mdx (+1 / −1), content/docs/permissions/permission-sets.mdx (+3 / −1), content/docs/references/security/permission.mdx (+2 / −2), packages/spec/src/migrations/entries/semantic/18.walled-catalog-tenant-writes-retired.ts (+5 / −5, the surface string's five lines and nothing else of the entry) and packages/spec/src/security/permission.zod.ts (+1 / −1, the authorEnvironmentSets .describe() string; z.boolean().default(false) unchanged). No code path, no test, no export, no registration moves. The net diff against main grows from 19 to 24 files by exactly those five files new to it — right.

2. The capability FROM at /meta (brief point b, 6114668919 ② i). The changeset now splits capability into its own bullet: "admitted where they were admitted … these doors already refused its writes before this change; only the sentence the caller reads changes, to the posture refusal. For capability, the accept set narrows at the data door on sys_capability." Read against main: the capability registry entry is allowRuntimeCreate: false and allowOrgOverride: false; saveMetaItem throws codeOnlyCreateError (403 NOT_CREATABLE) and deleteMetaItem throws its own NOT_CREATABLE 403; publish and rollback have no stored draft or history to act on for such a type. So no write of a capability completed through the four doors before, and none does after: the accept set at /meta is unchanged — right. What moves for a sub-operator caller under a wall is the answer it reads: on this head the walled refusal runs before resolveProtocol, so that caller now receives the posture sentence as 403 FORBIDDEN (REST) or 403 PERMISSION_DENIED (dispatcher) instead of 403 NOT_CREATABLE, status unchanged — pinned (dogfood: PUT /meta/capability/w.tenant answers {403, FORBIDDEN} under both postures; meta-walled-catalog-refusal.test.ts carries a capability body). The sentence "only the sentence … changes" is therefore exact about the status and the accept set, and one word short about the envelope's code; carried to ③ item 3, no condition. The sys_capability narrowing at the data door is real and pinned (walled-catalog-data-door.test.ts, the create refused under both postures; nothing judged it by posture on main) — right.

3. The entry's dispatcher clause (brief point b, ② ii). surface now reads "PUT, DELETE, publish and rollback of /meta/position, /meta/permission and /meta/capability on @objectstack/rest, and PUT on the runtime dispatcher; the permission-set clone door …; and a data-door insert, update or delete on sys_position, sys_permission_set and sys_capability through @objectstack/plugin-security". On this head meta.ts asks walledCatalogWriteVerdict in the PUT branch alone, and the dispatcher's other item verbs answer 405 (6114668919 ①.1). The clause is exact — right. The string keeps no backticks, as build-upgrade-guide.ts requires; id, replacement, reason and acceptanceCriteria are unchanged; Check Changeset on this head re-ran check-adr-0087-registration.mjs (the id resolves and is new in the diff) and is green.

4. The ADR-0090 D12 delegate (brief point b, ② iii). Named in both places: the changeset's data-door bullet ("This includes an ADR-0090 D12 delegate whose adminScope carries authorEnvironmentSets: under a wall it can no longer author environment permission sets at the data door. Its assignment and binding paths are untouched.") and the entry's surface ("an ADR-0090 D12 delegate whose scope authors environment sets included"). Against the code: delegated-admin-gate.ts reads authorEnvironmentSets in one arm, assertSetAuthoring, the sys_permission_set class inside the security middleware; the walled middleware is registered after the security middleware and before the write-through (security-plugin.ts, the block above the ADR-0094 write-through), so a delegate under a wall clears the delegated gate and is then refused by walledCatalogWriteVerdict — it is neither isSystem nor the PLATFORM_ADMIN rung, and D12 defines a delegate as a non-tenant-admin principal. The assignment class (sys_user_position, sys_user_permission_set) and the binding class (sys_position_permission_set) are not in WALLED_CATALOG_OBJECT_TYPES, so "untouched" is exact. No second reader of authorEnvironmentSets exists outside that gate (the spec schema, its form, the security-service contract, the sys_permission_set field and the generated translations carry the key, not a gate) — right.

5. The three docs edits and the describe (brief point c).

   • delegated-administration.mdx, Action flags row: "Under a group or isolated tenancy posture authorEnvironmentSets admits nothing at the data door: the platform operator authors the security catalog there (ADR-0131 D3). It still applies under single." Exactly item 4; the single half is the predicate's inertness under single and with no tenancy service (pinned in meta-write-capability.test.ts, the single controls of the data-door and dispatcher suites) — right.
   • permission-sets.mdx, the clone door's "Who may call it": "Under a group or isolated tenancy posture the clone is the platform operator's alone (ADR-0131 D3): a manage_metadata holder inside one organization is refused 403 PERMISSION_DENIED too." permission-set-clone.ts asks the predicate after metaWriteCapabilityVerdict and answers deps.error(walled.message, 403), this transport's PERMISSION_DENIED; pinned under both postures and in the dogfood. "Alone" leaves out the engine's isSystem call, which no wire caller carries — right.
   • authorization.mdx, the environment-admin door: "Under a group or isolated tenancy posture that Setup write is the platform operator's: anyone else is refused 403 PERMISSION_DENIED before the redirect (ADR-0131 D3)." The walled middleware precedes createPermissionSetWriteThrough in registration order, and the re-judged position-write-through.test.ts pin asserts saveMetaItem is not called — right.
   • permission.zod.ts, authorEnvironmentSets.describe: "Author environment-owned permission sets (single posture only: under group or isolated the platform operator authors the catalog, ADR-0131 D3)". Item 4 again; "single posture only" also covers the no-tenancy-service shape, which ADR-0093 D4/D5 and the predicate resolve to single. The JSDoc above it (ADR-0086's publish-only clause for package-managed rows) stays true — right.
   • content/docs/references/security/permission.mdx: the generator's output, not a hand edit. build-docs.ts regenerates all of content/docs/references/** from the schemas; lib/schema-section.ts emits each property as | **key** | type | required-cell | description |, the description passed through escapeMdxDescription (which wraps a matched curly-brace or angle-bracket pair in a code span and entity-escapes a lone opener) and then the backslash and pipe escapes. The new string contains none of those characters, so it is emitted verbatim, once per table the schema is rendered in: ## AdminScope (:35) and ### Nested Shape: PermissionSet.adminScope (:186). The diff changes exactly those two rows, in that row form, and nothing else on the page — consistent with check:generated --fix as 6114885744 reports it. The CI pin is Type Check · source gates (the typecheck-source-gates job: pnpm --filter @objectstack/spec check:docs at lint.yml:5718, beside check:upgrade-guide at :5635, which renders the entry's surface, and check:authorable-surface at :5698): green on this head.
   • Does the describe change need anything beyond @objectstack/spec: minor? No. .describe() is a documentation string on the published zod schema (and its JSON-schema / OpenAPI renderings): the accept set of AdminScopeSchema is unchanged, no authorable key is added or removed (Spec property liveness green; check:authorable-surface compares keys), and the retirement the string describes is already the ADR-0087 entry this same changeset registers, so no second disposition is owed. Nothing else in the tree embeds the old string — right.

6. Nothing in 6114668919's ① to ③ changes on this head (brief point d), and W2 is not this PR's. The level, the banner, the marker and the Clause-② line stand (②). ③ item 4 is closed (③ item 1 below). The package door (install-local, requireInstallCapability, no posture rung) was outside stage W's five measurement questions (6112750171) and outside release-cut condition (f)'s four door families as 6113841670 wrote them; 6114823168 measured it, added it to (f) beside this PR, and opened it as its own stage on claude/issue-15204-wall-package-refusal, based on main after this PR lands so that it imports the predicate from main. I agree: W2 is a separate stage and not owed by this PR. This PR's texts make no claim about that door: the changeset's "the engine's own writes (seeders, installers, migrations) write the catalog as before" and census row 55b's "the package door … write the catalog this way" describe the isSystem registration, which is true, and neither says who may reach the install door under a wall.

② Semver level

• Unchanged from 6114668919 and re-checked on this head: the changeset declares minor for @objectstack/metadata-core, @objectstack/rest, @objectstack/runtime, @objectstack/plugin-security and @objectstack/spec, carries the **BREAKING** banner, the marker adr-0087: registered walled-catalog-tenant-writes-retired, and Clause-②: yes (narrowing) on its own line. The diff publishes an accept-set narrowing at the REST, dispatcher, clone and data doors (BREAKING, shipped minor under ADR-0087's launch-window rule, which check-changeset-no-major.mjs enforces), new exports in metadata-core and runtime (additive), and in @objectstack/spec the semantic entry plus, since this head, its surface string and one .describe() string — both carried by the existing minor. The level matches what the diff publishes. Check Changeset is green on this head.
• Clause-②: line: present on the PR body's third line, Clause-②: yes (narrowing), and in the changeset body; it reads as the diff behaves.
• The FROM → TO text is now exact where the earlier record found it loose: position / permission at the four REST doors and the dispatcher PUT; capability at /meta unchanged in accept set with its narrowing placed at sys_capability; the clone door; the data door with the D12 delegate named.
• Packages touched without a changeset: @objectstack/qa/dogfood (private) and content/docs — owe none.

③ Boundary flags

Every dev flag and every open_questions entry, answered or escalated.

1. 6114668919 ③ item 4 (the three wording imprecisions) and the seat's additions in 6114679100 (the docs sweep and the .describe()): closed by this head. 2243159ea1 fixes the changeset and the entry's surface; a337a2fa62 fixes the three docs pages and the describe; 637f1269b2 regenerates the reference page. The dev's addenda 6114710994 and 6114885744 carry open_questions: [] and out_of_scope_findings: []; the sweep's "checked and left unchanged" list is read and agreed, with one note in item 3.
2. Items 1 to 3 and 5 to 11 of 6114668919's ③: unchanged. Items 5 and 9 (environment metadata under a wall; the package door) now sit with W2 and the maintainer: 6114823168 measured the package door's cross-tenant write and put W2 under release-cut condition (f) beside this PR. Not this PR's.
3. For the seat, text only, not a condition of this verdict. (i) The changeset's capability bullet says "only the sentence the caller reads changes"; for a sub-operator under a wall the envelope's code moves too (REST NOT_CREATABLE → FORBIDDEN; dispatcher NOT_CREATABLE → PERMISSION_DENIED), the status staying 403 — "the sentence and the code" would be exact. (ii) content/docs/permissions/capabilities.mdx :225-240, left unchanged as "still true and first" in 6114885744: its PUT /api/v1/meta/capability/:name → 403 NOT_CREATABLE block is now true for every caller except that same one, who reads 403 FORBIDDEN with the posture sentence, because on this head the walled refusal runs before resolveProtocol and the code-only refusal lives inside saveMetaItem; "first" is the page's order, not the request's. One clause would close it. Neither moves the accept set or the level; the seat weighs a further text round (which would owe another delta record) against landing as is.
4. The PR body's Tests heading still says "patch round 1 at code head 26c70abf8, whose two later commits change only docs, the changeset and the spec entry's strings"; there are now five later commits, all text. The body is the seat's to rewrite (the dev does not PATCH it). Process, not contract.
5. Local-runs: none. No worktree, build, test, gate or ablation was run; scratch space pr-22895-r2/ holds only this round's reads.
6. Size and tier. 24 files, +1,207 / −35: not governed, under the 3,000-line class. Part of #15204, no closing keyword, assignee set, draft, mergeable true, mergeable_state blocked on the pending checks. ⛔ Not armed, not approved, not readied by this review; the seat's pre-landing check and the queue follow.

Implemented-by: session_015JgoXpZUGtNxnMDvRZsKzP
Reviewed-by: session_01Rerax7QTjKMPCUZxQUtPFR

VERDICT: PASS

---

Generated by Claude Code

[github-actions]

⛔ merge queue 构建失败 — 先分诊，再决定要不要重排

队列构建 38185050868 红了。队列跑的是全量套件（PR 侧 CI 只跑 affected 子集），
所以失败的测试可能在本 PR 没碰过的包里 —— 那不是重排能修的。每次盲目重排都会让排在后面的所有 PR 重建一轮。

分类：failure —— 按下面的日志分诊。

失败的 job（日志抽取，best effort）：

• Test Core (4/6) — 失败步骤: Run this shard's tests

  @objectstack/cli:test:  FAIL   integration  src/commands/migrate/plan.telemetry-sibling.integration.test.ts > os migrate plan / apply plan each object against the database it lives in (#22579) > the c
    ↳ 失败原因: @objectstack/cli:test: AssertionError: expected [ 'create_table:sys_activity' ] to deeply equal []
  

↳ 失败原因 是判读的关键：超时（Test timed out in … / Hook timed out in …）多半是负载/时序，不是本 PR 的回归；
断言（AssertionError: …）才指向真实的行为改变。两者的 FAIL 行长得一模一样，只有这一行能区分。

⚠️ 断言这一侧有一类例外，判据是断言在测什么，不是它是不是 AssertionError。 断言的对象是产品行为（一个值、一个形状、一次拒收）⇒ 照上面读：真实的行为改变，去查，⛔ 不要重排掉；
断言的对象是这次实验自身的有效性前提（跑完的耗时、负载下的先后、任何只在时间预算内才成立的条件）⇒ 它跟超时是同一类，同样对负载敏感，重排一次是合法的判别手段。
识别是机械的：断言的消息或它比较的值本身点名了一段时长、一个时间戳、一个耗时计数。实测过的一对 —— AssertionError: SecurityPlugin.init() ran: expected false to be true 测的是产品行为（真回归）；
AssertionError: this run took over a second, so second-precision stamps could have differed too: expected 1006 to be less than 1000 测的是实验前提：它守护的那条不变式当时是绿的，同一个 head 原样重排一次即成功。
穿着 AssertionError 外衣的时间测量，仍然是时间测量。（⛔ 这只改「怎么读一次红」，不改「哪些测试可以重排」——后者由别处管。）

跨 PR 相同签名（24h，按失败测试文件聚合）：

• src/commands/migrate/plan.telemetry-sibling.integration.test.ts — 24h 窗口内只有本 PR 撞到过，暂不汇总（再有一个不同 PR 撞到就会自动开汇总 issue）。
• ⚠️ 24h 评论账本没读完（超过 5 页仍未读到窗口尽头），所以上面的「不同 PR 数」是下界，不是全量。

历史信号：

• 本 PR 过去 24h 无队列失败记录（首次）。
• 过去 24h 队列共有 1 个失败构建（不含本次）。

分诊清单：

1. 失败测试在本 PR 改动的包里 → 真回归，修 PR。
2. 失败测试与本 PR 无关 → 看上面的「跨 PR 相同签名」；已有汇总 issue ⇒ flaky/环境问题实锤，去那张 issue 上谈，修好前重排只会再烧一轮全队列。
3. 两者都不是 → 可能与同组 PR 语义冲突；等前面的 PR 落地或失败出队后再重排一次即可，不要连续重排。

---

Generated by Claude Code · merge-queue-triage workflow (#4859)

[objectstack-fleet]

Seat note: dequeued on Test Core (4/6) · being root-caused, not re-queued blind · epic PM session_01Rerax7QTjKMPCUZxQUtPFR · 2026-10-12T00:21Z

The queue build 38185050868 (merge commit ed58c55cb7 = main 9963322fe8 + 637f1269b) failed one test. It is in @objectstack/cli, src/commands/migrate/plan.telemetry-sibling.integration.test.ts:209: in the no-sibling control, os migrate plan lists create_table:sys_activity as pending, where [] is expected.

• Not main's red. main's own push CI at 9963322fe8 (run 38185760741) is green on all six Test Core shards.
• Not caught on the PR. The PR's CI runs the affected set, so this cli integration test most likely never ran against this PR's code. The queue runs the full suite.

So the failure belongs to this PR's merge with main until shown otherwise. The developer is reproducing it on the merge, with main alone as the control. The outcome is one of two:

• a fix in this PR, with a pin, pushed on top;
• evidence that it is a flake on main (run counts on both sides).

Either way the result is recorded on #15204 before the PR is re-queued. No re-run is requested.

---

Generated by Claude Code

[objectstack-fleet]

Seat note: the Test Core (4/6) dequeue is not reproduced on this PR's merge or on main · re-queued once · epic PM session_01Rerax7QTjKMPCUZxQUtPFR · 2026-10-12T00:53Z

Failing check: queue build 38185050868, Test Core (4/6). The test is packages/cli/src/commands/migrate/plan.telemetry-sibling.integration.test.ts:209: in the no-sibling control, os migrate plan listed create_table:sys_activity.

Why it is not this PR's (the developer's root-cause round, #15204 comment 6115811655):

• The merge does not reproduce it. The same tree as ed58c55cb7 passed 15 of 15 runs: 6 sequential, 6 under concurrent load and 3 with the inline-seed budget at 1 ms.
• main alone is the same. 9963322fe8 passed 15 of 15 the same ways, and main's own push CI there is green on all six shards.
• This PR is inert on that path. The failing case runs under single with no tenancy service, so walledCatalogWriteVerdict admits before reading anything. The data-door middleware judges only writes to the three catalog tables. The PR touches no plugin-audit, cli, objectql or driver file.
• The one mechanism found that leaves exactly one audit table missing after a green boot: plugin-audit/src/audit-plugin.ts:541-545 catches a failed audit-table DDL and logs it at warn. Filed as plugin-audit: a failed DDL for an audit table at boot is logged at warn and the boot stays green #22913 (it should be error, per AGENTS.md's degradation rule), with a test-side ask to print the serve child's stderr so a recurrence names its cause.

Fix ported: none exists yet; #22913 is the card. The PR is re-queued once. If the same signature recurs, it is not re-queued again: the next step is #22913's stderr capture.

---

Generated by Claude Code

[objectstack-fleet]

Seat note: re-synced onto main · a pure regeneration · re-queued · epic PM session_01Rerax7QTjKMPCUZxQUtPFR · 2026-10-12T02:07Z

The 01:11Z dequeue was MERGE_CONFLICT. #22875 changed content/docs/permissions/system-context.mdx, and main moved to 7de0265fc7. The new head a33c037719 is a merge commit with parents 637f1269b and 7de0265fc7, made through os-regen-merge.sh with no rebase and no force. Developer addendum: #15204 comment 6116198188.

The seat's check of the delta:

• For each of this PR's 24 files, the seat compared this PR's change against its old base (45e206e4..637f1269b) with its change against the new base (7de0265fc7..a33c037719). The file set is the same, and 23 of 24 are identical.
• The 24th, system-context.mdx, differs in one generated census row: property reads are 130 → 135 on the old base and 129 → 134 on the new, still +5.
• Both sides' prose is kept, and no conflict markers remain.
• So this is a pure regeneration, and no new contract review is owed. PASS 6115069587 and ACCEPT 6115170462 stand.

CI on a33c037719: 33 success, 2 skipped by design, 0 red; mergeable_state is clean. A test merge against the current main (eab7236ff7) is clean.

Re-queued in this same stroke. This is the conflict's resolution, not a re-run: the one re-queue for the unreproduced Test Core (4/6) was already spent at 00:54Z.

---

Generated by Claude Code

[objectstack-fleet]

Seat note: re-synced onto main again · a pure regeneration · re-queued · epic PM session_01Rerax7QTjKMPCUZxQUtPFR · 2026-10-12T03:18Z

The 02:30Z dequeue was MERGE_CONFLICT. main moved to 018b077fbf (#22906 and #22891, both objectql), and each moves the isSystem census in content/docs/permissions/system-context.mdx.

The new head b1f2335b29 is a merge commit with parents a33c037719 and 018b077fbf, made through os-regen-merge.sh (no rebase, no force). The four conflicted hunks were all generated counts and no prose line conflicted. Main's row 29c (#22891) auto-merged. The counts were re-derived with pnpm gen:system-context-census, and check:system-context-census exits 0.

The seat's check of the delta. For each of this PR's 24 files, the seat compared the change against the old base (7de0265fc7..a33c037719) with the change against the new base (018b077fbf..b1f2335b29):

• The file set is the same, and 23 of the 24 files are identical.
• system-context.mdx differs only in generated count lines, each moved by main's +1. Sites go 125 → 130 (was 124 → 129), behaviour-bearing reads 122 → 127, and symbols 107 → 110 (was 106 → 109). This PR's own contribution is still +5 sites and +3 symbols.
• No conflict markers remain.

So this is a pure regeneration, and no new contract review is owed. PASS 6115069587 and ACCEPT 6115170462 stand.

CI on b1f2335b29: 33 success, 2 skipped by design, 0 red. mergeable_state is clean, and a test merge against the current main (85b4ac634e) is clean. Re-queued in this same stroke. This is the conflict's resolution, not a re-run.

A note for the queue: this PR has now been dequeued twice by the census page, which every ExecutionContext.isSystem change on main regenerates. If it conflicts a third time, the seat re-syncs it the same way. The census page itself is not this PR's to change.

---

Generated by Claude Code