feat(verify): the unique-scope derived proof family, runnable as os verify --unique-scopes (#15953 slice 3)

[objectstack-fleet]

Part of #15953
Clause-②: yes

What this is

Slice 3 of #15953: the D halves of the classification table's rows 6-7 (comment 6115219467). It adds the unique-scope derived proof family to @objectstack/verify, runs it from os verify --unique-scopes (opt-in), and documents it. Rows 6-7's K residues stay in hotcrm, and their R residues (driver-sql DDL, ADR-0120 D3) are not in this PR. #15953 stays open for its remaining rows.

Property, for any app. Every field declared unique collides as its declared scope says (UniqueScopeSchema, ADR-0120 D1 and D3):

• Within one organization, a second holder of the value is refused, whatever the scope. On a single-posture boot that organization is the install's only one: the untenanted install, with no wall, where every write lands.
• Across two organizations, a true / 'organization' value is admitted. A 'global' value, or any value of an object with no organization column (tenancy.enabled: false, which degrades the scope to the column alone), is refused.

The family (packages/verify/src/unique-scope.ts)

• Population. Every field of every declared object whose unique, after the spec's own FieldSchema.safeParse, is declared (isUniqueDeclared). This is the value a built artifact carries, so an autonumber that authors no unique states 'organization'. The family never re-derives that default by hand.
  • A unique indexes[] entry is NOT in the population. It is a different surface, with its own meaning of bare true (refused there since protocol 18), and the field-level-versus-indexes[] spelling is already an author-time rule (R10).
• Published sources each expectation is read from.
  • The scope: FieldSchema.unique and isUniqueDeclared / isGlobalUnique.
  • Its meaning: the UniqueScopeSchema docblock and ADR-0120 D1 / D3 / S11.
  • The organization column: isTenancyDisabled (ADR-0066).
  • A wall: postureEnforcesWall over the stack's tenancy() posture.
  • What the caller cannot set: RUNTIME_OWNED_FIELD_TYPES plus readonly / system.
  • The refusal's identity: DuplicateRecordError's documented split, DUPLICATE_RECORD in process and UNIQUE_VIOLATION on the wire.
• Doors. Each probe writes through one of two doors:
  • REST as a signed-in caller: POST /data/OBJECT, and PATCH /data/OBJECT/ID for the update probe.
  • The engine's own write as the system principal: hooks.run with { system: true }.
• Probes. Per door, within × insert and within × update; on a walled boot also across × insert.
  • A first holder is written and read back.
  • A second record is offered the value as the first one stores it, so a normalizing hook (hotcrm upper-cases SKUs) does not hide the collision.
• Boots and legs.
  • within runs on every boot.
  • across runs only on a walled boot (os verify --multi-tenant, bootStack({ multiTenant: true })). There the family provisions a second organization the way a user makes one: a sign-up that creates its own (POST /auth/organization/create).
  • The admin is the first organization's caller when the boot binds it to one; otherwise a second provisioned sign-up is.
  • The system door names the organization on the row (organization_id), which is how a system writer must address one under a wall: an org-less system write is refused there with ERR_SYSTEM_WRITE_ORGANIZATION_REQUIRED (measured).
  • On a single boot, across is reported not run on report.boot, with the reason, never silently skipped.
• Attribution. A refusal proves a collision only when two things hold:
  • it is the platform's unique refusal at that door, naming the probed object;
  • it is this field's: it names the field, or it names none and a twin (the same record with only the probed value changed) is admitted.
  • The twin exists because of a measured fact. A true / 'organization' constraint is the COALESCE expression index, and SQLite reports it by index name, so the envelope carries no field (uniqueViolationColumn's contract). A 'global' index, or an object with no organization column, names the field.
  • An update needs no twin: it changes nothing else.
  • Any other refusal (an app's own dedupe, a rule, a permission, a constraint on another field) is unattributed.
• Exclusions, with reasons, counted.
  • A field the caller cannot set (readonly, system, a runtime-issued autonumber) is probed through the system door only, which keeps an explicit value (measured: REST answers droppedFields: readonly for an autonumber; the system door stores it).
  • A derived value (formula, summary) is excluded.
  • These are not-derivable: a closed or tiny value space (boolean, select, …), a multi-valued reference, a target the app does not declare, an object the CRUD derivation cannot synthesize.
  • A one-to-one reference is probed within; its across leg is named, not run, because the value is one organization's record.
  • An object declaring its own tenancy.tenantField has its across leg named, not run.
• Fixtures. Reused deriveCrudCases / fillRelationalRefs. The family-local fixtureBody follows slice 2's cut-down shape (required fields, dates spaced, unique values distinct); slice 2's copy is module-private in delete-behavior.ts, outside this claim. Fixtures go through the probe's own door and principal, with per-writer base records. Run on a fresh stack of its own.

os verify --unique-scopes:

• boots its own fresh stack (with --multi-tenant, the walled one);
• adds uniqueScopes.summary.hardFailures (contradicted plus failed) to the exit status;
• emits uniqueScopes under --json.

Without the flag, nothing changes.

Evidence

All readings at this PR's head unless stated otherwise. The fixture and hotcrm runs used packages/verify/dist built from the merged tree 5fae9e5a13; the only later commit is the two NOT_A_STAND_IN rows.

Fixture app (packages/verify/src/unique-scope.test.ts, Tests 12 passed (12)):

• Single boot: 9 fields.
  • Probed: 6 per organization, 2 global.
  • 8 proven over 28 writes, all held.
  • 1 not derivable (a unique boolean).
  • across reported not run.
• Walled boot (multiTenant: 'posture-only', real organizations and per-caller stamping, see below): 8 proven over 39 writes, all held. Across, per-organization values are admitted, and a 'global' value or one with no organization column is refused.

Negative controls, one per leg, no engine change:

• within: a field declared unique on a table that already holds a duplicate when the declaration arrives.
  • driver-sql cannot create the index and logs that the constraint "is NOT enforced until the data is deduplicated" (ADR-0120 D4).
  • Every second holder is admitted, and the family reports contradicted on 4 of 4 writes (REST and system, insert and update), hardFailures 1.
• across: a field declared unique: true beside a declared indexes: [{ fields: ['name'], unique: 'global' }], the R10 double declaration.
  • The second organization is refused naming the field.
  • The family reports contradicted on 2 of 2 across writes, with within held.

Attribution control:

• An app hook's own dedupe (DUPLICATE_VALUE, the hotcrm shape) is unattributed on 4 of 4 writes.
• A field-less unique refusal that the twin meets too (a hook stamps one code on every record) is unattributed with twin: refused.
• Neither is ever credited.

Ablation of the family's own comparison (judge), through scripts/ablation-replace.mjs (wrap mode) inside a scratch script with trap restore EXIT INT TERM on absolute paths, at ed93727ce2. No build is involved: the test imports the module by a relative path.

• A. if (probe.expected === 'admitted') return { outcome: 'held' }; was replaced by an unconditional held plus the marker ABLATION_15953_S3_A.
  • On disk: anchor 1 to 0, marker 0 to 1, blob fee9f33f to e477ff09, marker 1 during the run.
  • RED: Tests 1 failed | 11 passed (12), exactly the within negative control.
  • Restored: blob fee9f33f equals HEAD and git diff HEAD is empty.
• B. The same on if (probe.expected === 'refused') … (marker ABLATION_15953_S3_B).
  • On disk: anchor 1 to 0, blob fee9f33f to 7b2dcfad, marker 1 during the run.
  • RED: Tests 1 failed | 11 passed (12), exactly the across negative control.
  • Restored to HEAD's blob, empty diff.
• Predicted direction for both: red. Observed: red.

hotcrm, the first consumer. Read-only clone at 1d7148bf, fetched first. main is at 0f84c9c7, and git diff 1d7148bf origin/main -- src adds or removes no unique / autonumber line (the same grep hits those files at HEAD). The clone's git status stayed empty throughout. os verify itself cannot boot hotcrm on this tree (protocol ^17.7.0 versus 18), so the family ran directly, as in slices 1-2.

• Artifact. Built by hotcrm's own CLI 17.7.0 into the scratchpad.
• Snapshot. This tree's os migrate meta --from 17 --out, with only engines.protocol moved to ^18.
• Population. 13 fields, all per-organization:
  • crm_account.name, crm_contact.email, crm_product.sku;
  • 10 autonumbers, nine of them 'organization' by the parse default, plus crm_case.case_number, which declares true by hand.

run	engine / boot	legs	proven	contradicted	inconclusive	writes held / total
A	hotcrm's pinned 17.7.0 bootStack over its own artifact, REST door only (17.7.0 has no system door)	within (across not run: hotcrm declares no @objectstack/organizations)	2	0	11	4 / 6
B	this tree, snapshot plus the artifact's 44 hook bodies grafted by name, single	within	9	0	4	23 / 32
B	the same, walled (multiTenant: true, the open @objectstack/organizations, OS_AUTH_MEMBERSHIP_POLICY=invite-only)	within + across	9	0	4	35 / 48
C	this tree, snapshot with hooks inert, single	within	11	0	2	28 / 32
C	the same, walled	within + across	11	0	2	42 / 48

Against what the retired pins asserted (D halves only):

• Row 6, test/account-name-tenant-scope.test.ts, crm_account.name: PROVEN on every run.
  • "Still rejects a duplicate account name WITHIN one organization": refused through REST and system, insert and update, on the walled boot and on the single boot.
  • "On an UNTENANTED install too": the single boot.
  • "Accepts the same account name in two different organizations": admitted through REST and system on the walled boot.
  • Run A shows the within half holding on hotcrm's own pinned engine.
• Row 7, test/case-number-tenant-scope.test.ts, crm_case.case_number: PROVEN on runs B and C.
  • Within: refused through the system door, insert and update, on both boots.
  • Across: admitted in a second organization on the walled boot.
  • It is an autonumber, so REST is excluded (that door drops the value), and Run A, which is REST only, names it inconclusive with that reason.
• 0 contradictions with what either file asserted.

Extra evidence, not a table row: test/contact-email-tenant-scope.test.ts. This file's subject is hotcrm's contact_integrity hook (an app dedupe, a K property), with the (organization, email) index behind it.

• Hooks live: within is unattributed, because the hook answers its own DUPLICATE_VALUE (the attribution control holding on a real app). On the single boot the system insert reaches the index and is credited: the hook scopes a system write by the row's own organization stamp, and that write carries none. Across, the value is admitted through both doors ("lets ANOTHER organization know the same person").
• Hooks inert: proven.

Inconclusive on hotcrm, with the app's own reasons:

• crm_campaign_member.member_number and crm_event_attendee.attendee_number: a generic fixture names no lead or contact party.
• crm_contract.contract_number: contract_validation ties the term to the dates (hooks live only).
• crm_contact.email: as above.

Verification

At head 9c89d3cc89: this branch with origin/main 0cbfb7c32c merged, so 9 files and 1637 changed lines (+1630 / -7) against merge base 0cbfb7c32. Readings below are at that head unless another commit is named.

Suites.

• @objectstack/verify: pnpm --filter @objectstack/verify exec vitest run --maxWorkers=2 reads Test Files 34 passed (34) / Tests 270 passed (270), lock verdict command-exit 0. The family file alone reads Tests 12 passed (12).
• Typechecks: pnpm --filter @objectstack/verify typecheck and pnpm --filter @objectstack/cli typecheck exit 0, and check:test-typecheck: OK holds for both (at 5fae9e5a13; no file of either program changed since). tsc --listFilesOnly -p tsconfig.test.json shows unique-scope.test.ts in verify's program (2 hits, with its module) and verify-unique-scopes.test.ts in the CLI's (1 hit; control verify-delete-behaviors.test.ts, 1 hit).
• CLI unit tier: pnpm --filter @objectstack/cli exec vitest run --project unit --maxWorkers=2 reads Test Files 283 passed (283) / Tests 4270 passed (4270) (at 5fae9e5a13; the second merge changed no packages/cli file).
• CLI integration tier: pnpm --filter @objectstack/cli exec vitest run --project integration --maxWorkers=2 over the six verify-* files, this PR's new one included, reads Test Files 6 passed (6) / Tests 18 passed (18). The rest of the integration tier is declared to CI.
• Core repo project: pnpm --filter @objectstack/core exec vitest run --project repo --maxWorkers=2 reads Test Files 6 passed (6) / Tests 83 passed (83). The #22661 second-object read pin is among them, green with no CALLERS row.
• pnpm check:verify-stand-in:
  • Before the rows: exit 1, 2 problem(s) (runUniqueScopeProofs and formatUniqueScopeReport unclassified).
  • After: exit 0, OK check-verify-stand-in-erasure: 2 stand-in check(s) guarded, 13 candidate(s) explicitly exempt, 10 call site(s) reached, 0 asserted driver arguments.
  • --self-test OK.

Gates.

• Derivation: node scripts/pm/dispatch-gates.mjs --commands --repo objectstack-ai/objectstack, with no paths, gives 115 commands (64 pnpm, 51 direct node).
• Results: every one exits 0, with its exit code captured before any pipe. --ran reads Run reconciliation — 115 derived, 115 run, 0 NOT-MEASURED, 0 UNRUN.
  • The derivation was current when the run began.
  • The reconcile pass printed STALE TREE, because origin/main moved 4 commits during the 40-minute run (to e448d3c27f). The stale family files are .github/workflows/lint.yml, scripts/check-system-context-census.mjs and scripts/sdui-manifest.record.json, none on this PR's paths.
  • An earlier battery at 5e776ba287 recorded 3 PREREQUISITE NOT MET (exit 3: check:dual-build-cjs-loads, spec check:skill-examples, check:i18n-coverage, each needing packages outside the built closures). After a full turbo run build --filter=!@objectstack/docs (73 tasks), this battery measured all three, exit 0.
  • Every command and its printed verdict line is in the report comment on verify: classify hotcrm's "platform-semantics pins" — each becomes a derived proof family in @objectstack/verify or a platform regression test in dogfood, never an app test (epic hotcrm#1579, step 5c) #15953.

Lint, by proven narrowing at 9c89d3cc89: eslint --no-inline-config --format json over the 6 touched source files.

1. Population, read from eslint's own config: --print-config resolves for 6 of 6.
2. Counts, read from the JSON: 6 files, 0 errors, 0 warnings.
3. Invariance: eslint.config.mjs never enables type-aware linting (no parserOptions.project, no typed @typescript-eslint rules, its own comment at :327-328), so this diff cannot move any untouched file's verdict.

The two .md / .mdx files are outside eslint's population.

Acceptance notes

Noted, not filed (none is a reproducible defect through a public door, a contract violation or an authoring trap; carrier: none):

• bootStack over an UNPARSED object literal registers the raw field. An autonumber that authors no unique therefore gets no index there, and a system-door duplicate is admitted (measured: raw literal 2 of 2 admitted; the same object through defineStack stores 'organization' and refuses the second).
  • Every config os verify / os validate / os build load is defineStack-built (STACK_PROVENANCE_MISSING otherwise), so no public door reaches it.
  • The family reads the parse's scope and would report such a programmatic boot contradicted.
  • This PR's positive fixture boots defineStack output for that reason.
• The ADR-0120 D3 NULL-organization bucket is unreachable through either door on today's runtime: a single install derives its one organization for a system write, and REST stamps the caller's. So "untenanted install" is the single-posture boot, plus objects with no organization column. The NULL-row DDL fact stays row 7's R residue (driver-sql).
• os verify --multi-tenant with the open @objectstack/organizations refuses to boot until a membership policy is declared (WalledMembershipPolicyError; OS_AUTH_MEMBERSHIP_POLICY). This is by that package's design, and it applies to --rls --multi-tenant alike. The README states it beside the flag.
• The CLI-level control in packages/cli/test/verify-unique-scopes.test.ts is a door that faults on a duplicate (exit 1), not a contradiction. Through os verify, every config is defineStack-built on a fresh database, so no metadata makes a same-organization duplicate land. The across leg needs --multi-tenant, whose runtime package no framework package may declare, so CI's test graph cannot boot it. The two contradiction controls are in packages/verify.
• The walled-boot unit suites use multiTenant: 'posture-only': a walled posture with real organizations (/auth/organization/create) and per-caller stamping, measured. The family's subject is which rows collide, which the index decides from each row's organization; read isolation is not under test. The real-package walled boot is exercised by the hotcrm runs above.

Deviations

• PM mechanism assumption 2 was measured and partly falsified. The per-organization refusal carries no field (SQLite names the expression index), so "proved only on that code naming the probed field" would prove no per-organization field at all. Attribution therefore also accepts a field-less unique refusal when the twin is admitted, and pins it with the attribution control above.
• The conditional NOT_A_STAND_IN surface was used (2 rows), because check:verify-stand-in asked. The #22661 pin needed no CALLERS row: the family reads a reference's target from the CRUD derivation and never calls referenceTargetOf (the repo project is green without one).
• origin/main (df7a137923, 4 commits, none on this surface) was merged before the PR, per AGENTS.md section 10. The build was refreshed, and every suite and gate here was run on the merged tree.
• Commits carry the model-free trailer pair per AGENTS.md, and this body ends with its session-URL footer rather than the harness's own attribution form.

---

Generated by Claude Code

[github-actions]

📓 Docs Drift Check

This PR changes 2 package(s): @objectstack/cli, @objectstack/verify, touching 49 documentable anchor(s). ⚠️ 2 changed file(s) yielded no anchor (packages/verify/README.md, packages/verify/src/index.ts), so the pages documenting them are NOT COVERED by this run — this is not a clean bill of health for those files.

28 hand-written doc(s) name something this change touched — list omitted above 15 rows. Re-derive on the tree named below: node scripts/docs-audit/affected-docs.mjs --json e448d3c27f4bd9031428347f6bdd865445486497.

⛔ 9 release-owned page(s) also affected — read-only, see AGENTS.md Documentation Guardrails.

What this run could not see

• 2 changed file(s) yielded no anchor (packages/verify/README.md, packages/verify/src/index.ts) — pages documenting those are invisible to this run
• 2 anchor(s) matched too much of the corpus to be a work list: organization_id (literal, 35 pages), /data/:object (route, 67 pages)
• 31 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 — 30 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 e448d3c27f4bd9031428347f6bdd865445486497 → packageMentionDocs.

Which tree this was computed on

This run read content/docs from 2e84e8ac290602763b4ebda2b61eee23df1e2bc1 — the merge of head 9c89d3cc89a967ef9eb61627c02855d831a158ab into base e448d3c27f4bd9031428347f6bdd865445486497, 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 2e84e8ac290602763b4ebda2b61eee23df1e2bc1 && git checkout 2e84e8ac290602763b4ebda2b61eee23df1e2bc1
# afterwards, rebuild it from the two parents, which stay fetchable
git fetch origin e448d3c27f4bd9031428347f6bdd865445486497 9c89d3cc89a967ef9eb61627c02855d831a158ab && git checkout -B drift-repro e448d3c27f4bd9031428347f6bdd865445486497 && git merge --no-ff 9c89d3cc89a967ef9eb61627c02855d831a158ab

node scripts/docs-audit/affected-docs.mjs --json e448d3c27f4bd9031428347f6bdd865445486497

⚠️ 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 e448d3c27f4bd9031428347f6bdd865445486497 → 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: 9c89d3cc89a967ef9eb61627c02855d831a158ab
Local-runs: none

Read at 2026-10-12T16:13Z, adversarially, against AGENTS.md Prime Directives #10 and #12 and Post-Task Checklist step 3. Inputs: card #15953 (body and all 15 comments: the classification table 6115219467, the slice 3 claim 6125131626, the dev report 6127784027, the seat's ACCEPT 6127885139 read as a claim, and slice 2's record 6118395322 on PR #22953 as the precedent); PR #23069 body, file list and the net diff of refs/pull/23069/head against merge base 0cbfb7c32c (9 files, +1630 / -7); the check-runs on the head. Nothing was built, run or re-run; the only local commands were git show / git diff against the fetched ref and the record template print.

① Derived judgments

The three functions and eleven types, reachable and coherent: right. packages/verify/package.json maps . to dist/index.js / dist/index.d.ts (and the require pair), and packages/verify/src/index.ts exports deriveUniqueScopeCases, runUniqueScopeProofs, formatUniqueScopeReport and exactly the eleven types the changeset lists from ./unique-scope.js, so every one is reachable through the published entry. Shapes mirror the family beside them: deriveXCases(config) / runXProofs(stack, token, config, { doors? }) / formatXReport(report) is deriveDeleteBehaviorCases / runDeleteBehaviorProofs / formatDeleteBehaviorReport to the parameter; app is computed by the same expression as delete-behavior.ts and reference-shape.ts; summary.hardFailures is contradicted + failed, the gate count slice 2 introduced; the six field statuses are slice 2's vocabulary with each docblock matching fieldStatus and judge line for line. Three places where an export says slightly more than the code measures, none a contract defect: (a) runUniqueScopeProofs promises it "never throws on a per-field failure", but the stack.rows read-backs in holder, offerInsert and offerUpdate are not caught, so a harness-level read fault escapes rather than landing as failed (the same gap slice 2's record noted on its family); (b) UniqueScopeCase.across ("what a second holder in ANOTHER organization meets") is still emitted as admitted on a case whose across leg is excluded (a one-to-one reference, a tenancy.tenantField object), so the report carries an expectation beside the reason it is not measured; no verdict rests on it, since the leg never runs; (c) boot.legs.within is the literal run even when options.doors is empty. Noted, not flagged.

os verify --unique-scopes: right. Flags.boolean with default: false, typed into runVerification. With the flag, bootStack(config, { multiTenant, hostRoot }) is a fresh stack stopped in finally, where multiTenant is the run's own resolveVerifyMultiTenant(flags) and hostRoot the same value the four sibling phases take; uniqueScopes.summary.hardFailures is added under ?? 0; uniqueScopes rides emitJson (a JSON.stringify, so an undefined value emits no key); the text branch is if (uniqueScopes). Without the flag nothing moves: hardFailures adds 0, the --json document carries no uniqueScopes key, stage 1 runs and exits exactly as before, and the exit status reads what it read. The only thing that moves for every caller is help text (the command description and one examples row), a widening the minor carries. packages/cli/test/verify-unique-scopes.test.ts spawns the real process and pins it ('uniqueScopes' in doc false and exit 0 without the flag; exit 0 on the proven app; exit 1 on the faulted door); packages/cli's test script is a bare vitest run, so both its projects, the integration one included, run under CI's Test Core.

The expectation's source: right, and every excluded shape is excluded, not modelled. deriveUniqueScopeCases reads the scope from FieldSchema.safeParse(def).data.unique through isUniqueDeclared / isGlobalUnique and from nowhere else, so an autonumber that authors nothing carries the 'organization' the spec's .overwrite() tail puts there, never a hand copy of that default. What each scope means is UniqueScopeSchema's docblock and ADR-0120 D1 (per-organization: refused in the same organization, admitted in another; 'global': installation-wide). The organization column is isTenancyDisabled (ADR-0066); the wall is postureEnforcesWall over stack.tenancy().posture (ADR-0105). The refusal's identity is DuplicateRecordError's documented split (DUPLICATE_RECORD thrown, UNIQUE_VIOLATION on every REST route per the 2026-09-03 ruling recorded in protocol.ts) and field only when uniqueViolationColumn named a column. The two shapes the brief asks about: for an object with no organization column, the cross-organization refused is read from the docblock's own sentence ("on an object with no tenant column at all (tenancy.enabled: false) both per-organization spellings degrade to the listed column alone") and ADR-0120 D1 / S11, a single-column index being installation-wide by construction, so published, not invented; for a one-to-one reference, the family states NO across expectation: acrossExcluded names the reason and the leg is never run, which is an exclusion, and the safe direction (the stale across: 'admitted' beside it is item (b) above). Also excluded or not-derivable with the reason, never given an expectation: a parse refusal, formula / summary, a closed value space, a multi-valued reference, an undeclared target, an unsynthesizable object, a tenancy.tenantField object, and the REST door on readonly / system / id / runtime-issued fields. A unique indexes[] entry is kept out of the population, and the README and changeset say so.

Attribution: right on every named path, with one residual the seat owes a carrier (see ③). judge credits held on a refusal only when the code is the door's unique code and the refusal's object is absent or the probed one; a refusal naming another field is unattributed; a field-less refusal on an insert is unattributed unless the twin was written; the twin is create(second, c.object, { [c.field]: freshValue }, n) with the SAME fixture counter n as the refused write, so only the probed value differs; an update changes only the probed field, so it needs no twin. A hook's own code (DUPLICATE_VALUE, the hotcrm shape), a rule, a permission, a constraint on another field (field named), and a field-less refusal the twin meets too (the uqa_tagged control) all stay out of proven, and the attribution control pins both halves. The residual: a declared unique COMPOSITE indexes[] entry that covers the probed column together with columns the fixtures hold constant (an optional field with a defaultValue, a hook-stamped value, a required type vary leaves alone such as a boolean) refuses the second holder field-less (a composite is undefined under uniqueViolationColumn's contract) and admits the twin, and the update probe credits the same way. Both paths then credit a refusal the field's own declaration did not cause. It matters only when the field's own index is absent: on os verify's fresh database that is an engine or driver fault, the class the family exists to find, masked by the author's composite; on a reused database it is the ADR-0120 D4 not-enforced state the driver logs at boot. R10 does not fence it: lintUniqueDeclarations is advisory (severity: 'warning', "never fails a build") and reads single-column declared indexes only, composites exempt by design. The family already reads indexes[] (uniqueFields), so the tightening is cheap: on a field-less refusal where a unique indexes[] entry covers the probed column, report unattributed, or write a second twin varying that entry's other columns. Not a FAIL: the shape needs the author's own composite over the same column plus an unenforced field index, the docs state the mechanism exactly ("it names the field, or it names none and the same record with only the probed value changed is admitted"), every other refusal errs toward inconclusive, and nothing in this PR's evidence (fixture, controls, hotcrm as reported) carries that shape.

The two NOT_A_STAND_IN rows: true of the code. The gate's candidate is an exported function whose parameter 0 is annotated with an interface declared in the package and re-exported as a type from its index. runUniqueScopeProofs(stack: VerifyStack, ...) and formatUniqueScopeReport(report: UniqueScopeReport) are exactly those two; deriveUniqueScopeCases(config: any) is not a candidate. VerifyStack is the concrete handle bootStack returns and UniqueScopeReport is a value this package produced, so neither is a surface an out-of-tree implementer provides; the rows' words are the sibling rows' words. Lint & Repo Gates, which carries check:verify-stand-in, is green on the head.

Docs and changeset sentences against the diff: right, two of them narrower than the code. Changeset: the export list is the full set (3 functions, 11 types), the signature line is the real one, the refusal identities, the twin sentence, the exclusions and hardFailures all match unique-scope.ts. Its provisioning sentence, "the way a user makes one (a sign-up that creates its own)", compresses two calls into one phrase: provisionOrganization is stack.signUp(email) THEN POST /auth/organization/create { name, slug } THEN a check that the new organization is the sign-up's active one; the PR body names the route, the changeset does not. Accurate in substance, imprecise in the parenthetical; and the leg is also reported not run on a walled boot whose provisioning fails, a second reason the sentence does not name but the code reports on boot.acrossNotRun. README: "Five app-agnostic proof families", the bullet, the opt-in paragraph, the exit-code sentence, the status table (each row matches fieldStatus / judge / deriveUniqueScopeCases) and the API line all match. Its membership-policy sentence is grounded in this repo: assertWalledMembershipPolicyDeclared in packages/plugins/organizations/src/membership-policy-gate.ts throws WalledMembershipPolicyError on any non-single posture when an auth service is present and no policy is declared. The declaration may come from the auth.membership_policy setting OR OS_AUTH_MEMBERSHIP_POLICY (plugin-auth.ts names both); the README names the env var only, which is the practical route on a fresh os verify boot with no seeded settings. Accurate, slightly narrow. content/docs/deployment/cli.mdx: the example row, the paragraph in the os verify section only (insert and update, both doors, the attribution sentence, the across provisioning, the not-run reason, the autonumber parse default, the system-door-only rule, what fails the run, "inconclusive, never proven") and the --json key list (uniqueScopes with its boot) match verify.ts and the family.

First-consumer evidence (hotcrm): consistent with the retired pins, as reported. Rows 6 and 7's D halves proven on every run that has the door (crm_account.name on all five runs; crm_case.case_number on B and C, inconclusive on the REST-only A with the reason), 0 contradicted over 13 fields, the inconclusive four carrying the app's own refusals. The evidence came from direct runs of the family, not from os verify; see ③.

② Semver level

.changeset/15953-verify-unique-scope-family.md: @objectstack/verify minor, @objectstack/cli minor. The diff publishes three new functions and eleven new types from @objectstack/verify and one new opt-in flag plus help text from @objectstack/cli; it removes and renames nothing, narrows no accept set and changes no default. Minor is the right level for both and meets the card's own acceptance ("@objectstack/verify minor per new exported family"). Not breaking, so no migration text and no ADR-0087 disposition marker is owed; Check Changeset on the head is green.

Clause-② carriers: PR body line 2 Clause-②: yes; the changeset body Clause-②: yes; the claim comment 6125131626 Clause-②: yes (the bare form claim lines take). The three agree. Each is legal: yes with no arm is a complete declaration (the (widening) arm would add nothing), and yes takes at least minor, which both entries are.

③ Boundary flags

Dev report 6127784027: open_questions is empty, so nothing to answer there. Its deviations and findings, each answered or escalated:

• The attribution twin (PM mechanism assumption 2 falsified by measurement): a per-organization refusal names no field because the constraint is the ADR-0120 D3 COALESCE expression index and uniqueViolationColumn never reports an index as a column, so "proved only on the code naming the probed field" would prove no per-organization field at all. The twin is the right differential and is pinned. Answered in ①, with the composite residual ESCALATED: the seat owes it a carrier, a follow-up card naming the shape and the one-line tightening above; it does not move this head.
• Second-organization provisioning (assumption 3): no sibling family provisions an organization; this one reuses signUp and the platform's own workspace door, and addresses the system door by organization_id on the row, which the engine requires under a wall. Answered; the changeset's phrasing of it is noted in ①.
• Walled unit suites booted multiTenant: 'posture-only': an existing BootOptions member on main (harness.ts documents it as the posture without the wall, with real organizations and per-caller stamping), used because no framework package may declare @objectstack/organizations. The family's subject is which rows collide, which the index decides from each row's organization, so the posture-only boot measures it; the real-package walled boot is exercised by the hotcrm B and C runs. Answered.
• The CLI control is a faulted door, not a contradiction: every config os verify loads is defineStack-built on a fresh database, so no metadata makes a same-organization duplicate land, and the across leg needs a runtime package CI's graph cannot boot. The two contradiction controls and the attribution control live beside the family. Answered.
• The positive fixture booted as defineStack output: the parse is what puts the autonumber's 'organization' on the registered field. Answered, and it is finding 1 below.
• One conditional surface used (2 NOT_A_STAND_IN rows, the gate red first); the #22661 CALLERS row not needed (the family reads a reference target from the CRUD derivation and never calls referenceTargetOf; the core repo project is in Test Core, green). Answered in ①.
• Two merges of origin/main (no conflicts, no force-push): the head is a merge commit whose net diff is the 9 files reviewed. Answered.
• hotcrm evidence by direct runs, not os verify: this tree is protocol 18 and hotcrm pins ^17.7.0, so the card's clause "runs from os verify on hotcrm's metadata" cannot be met on this tree for this slice either. ESCALATED, not a FAIL: the same open acceptance note slices 1 and 2 landed under (6116484972, 6124307791), the card stays open for its remaining rows, and whoever closes it owes that clause a run or a re-scope.
• Commit trailers and local leftovers (the read-only hotcrm clone left in place with a clean tree; the worktree removed): process notes, not contract matters. Answered.
• Out-of-scope finding 1, bootStack over an unparsed literal registers the raw field (an autonumber that authors no unique then gets no index and a system-door duplicate is admitted): ESCALATED as a choice for the seat, not a FAIL. bootStack is a public export this README documents (bootStack(config, opts?)), harness.ts checks no provenance, and the README's examples all pass defineStack output without saying it must be; a caller who hands a raw literal with an autonumber gets neither the spec's default nor a refusal. The family itself would report that stack contradicted, so it hides nothing, and every CLI door refuses an unparsed stack (STACK_PROVENANCE_MISSING). Under chore: version packages #10 this is a note that borders a trap: the carrier is either the provenance refusal on bootStack (one card, outside this claim) or one sentence on bootStack's docblock and README line saying the config is defineStack output. The seat picks; this PR's fixture is already correct.
• Finding 2 (the ADR-0120 D3 NULL-organization bucket is unreachable through either door on today's runtime): a statement about what the engine stamps, not a defect; row 7's R residue stays driver-sql's. A note is right. Answered.
• Finding 3 (os verify --multi-tenant with the open @objectstack/organizations refuses to boot until a membership policy is declared): by that package's design (membership-policy-gate.ts), it applies to --rls --multi-tenant alike, and the README now states it beside the flag. A note is right. Answered.
• Finding 4 (a generic fixture cannot satisfy every app rule, so some hotcrm autonumbers read inconclusive with the app's own refusal): an app fact, reported inconclusive, never proven. A note is right. Answered.

Check-runs on the head, read four times in this act, the last just before posting (latest run per name). At the final read every run had concluded: 34 success, 7 skipped, 0 failure, 0 in progress. Concluded success: Build Core, Build Docs, Check Changeset, Check Documentation Links, Dogfood Regression Gate, Flag docs affected by code changes, Governed Surface Queue Guard (not governed: no governed path in the file list), the four PR-shape guards, Spec property liveness, all six Test Core shards, the four Type Check jobs, TypeScript Type Check, Lint & Repo Gates, the eight Lint · jobs and filter. Skipped: Auto Label, Check PR Size, Console Pin Gate, the Dogfood Regression Gate matrix row, Dogfood Verify CLI, Packed-tarball smoke (opt-in) and Temporal Conformance (live PG + MySQL); the last three and Console Pin Gate carry github.event_name != 'pull_request' in their own if:, so they are roster skips on a PR run, not gates that should have run. No gate was re-run by this act. What the seat owes before the queue: the two carriers named above (the composite-attribution residual, and the bootStack provenance choice), each of which is a follow-up outside this diff.

Implemented-by: claude/issue-15953-s3-unique-scope-family
Reviewed-by: session_01VoSxBQujKLZKPwK2u5ehQ6

VERDICT: PASS

---

Generated by Claude Code