@@ -0,0 +1,34 @@
+---
+"@objectstack/plugin-security": patch
+---
+
+fix(plugin-security): #2936 — RLS field-existence / tenancy-disabled safety nets now recognize canonical `==`
+
+`extractTargetField` (the lightweight left-hand-field parser feeding the Layer 1
+**field-existence** net and the **tenancy-disabled** skip net in
+`computeLayeredRlsFilter`) only matched the legacy single-`=` / `IN` shape. It
+returned `null` for canonical CEL `==` (`field == …`), which is how real seeds
+and business policies author equality. A `null` target field means "keep the
+policy", so both safety nets were **inert** for every `==` policy: a wildcard
+policy targeting a column the object lacks was NOT failed closed, and a
+`==`-form `organization_id` policy on a `tenancy.enabled:false` object was NOT
+skipped. The regex now recognizes `==` (listed before `=` so the ordered
+alternation does not mis-match the first `=`), alongside the existing `=`/`IN`.
+Recognition is only **extended** — the net semantics are unchanged, and
+`!=`/`>`/`<`/`>=`/`<=` still return `null` (conservative keep), matching prior
+behavior for any unmatched shape.
+
+Behavior delta (fail-closed strengthening, same effective visibility): the
+wildcard `owner_only_writes` / `owner_only_deletes` seed policies
+(`created_by == current_user.id`) now correctly fail closed on an object that
+lacks a `created_by` column (platform-global / system tables). Previously they
+slipped the net and compiled to a phantom `{ created_by: … }` filter against a
+missing column — a driver-dependent, effectively-deny result; now the net drops
+the sole applicable write policy and yields the deny sentinel. A member could not
+by-id write such a column-less object either way, so the visible/writable row set
+is unchanged; only the mechanism is now an explicit fail-closed deny. All ordinary
+tenant/business objects carry `created_by`, so they are unaffected (proven green by
+the dogfood authz-conformance + RLS matrices). The tenancy-disabled skip net has no
+effect on any current seed (no `==` seed policy targets `organization_id` on a
+tenancy-disabled object). The tenant wall itself is Layer 0 (`tenant-layer.ts`),
+which never used this parser, so tenant isolation is unaffected (ADR-0095 D1).
@@ -203,11 +203,22 @@ const EXPECTED_MATRIX: Record<string, Record<string, { read: unknown; write: unk
     org_admin: { read: null, write: 'BYPASS(no-write-filter)' },
     // [D1 accepted delta (c)] A member reading a `tenancy.enabled:false` global
     // object: Layer 0 treats it as a NON-tenant object (no phantom org filter) →
-    // the global catalog is VISIBLE (read: null). The write drops the dead
-    // org-disjunct to plain owner scope (same visibility, no org column exists).
-    member: { read: null, write: [{ created_by: 'u1' }] },
-    // [D1 accepted delta (c)] no-org member likewise sees the global catalog (was deny-sentinel).
-    no_org_member: { read: null, write: [{ created_by: 'u2' }] },
+    // the global catalog is VISIBLE (read: null).
+    // [#2936] WRITE now fail-closes to the DENY sentinel. `sys_package` has no
+    // `created_by` column, and the wildcard `owner_only_writes` policy authors
+    // its predicate as `created_by == current_user.id` (canonical `==`).
+    // Pre-#2936 `extractTargetField` did not recognize `==`, so the
+    // field-existence net let the policy through and it compiled to a phantom
+    // `{ created_by: … }` filter against a column the object lacks — a
+    // driver-dependent, effectively-deny result. Now the net recognizes `==`,
+    // sees `created_by` is absent, and drops the only applicable write policy →
+    // deny sentinel (fail-closed). SAME effective visibility (a member cannot
+    // by-id write a column-less global object either way); the mechanism is now
+    // an explicit fail-closed deny instead of a phantom-column filter.
+    member: { read: null, write: [{ id: DENY }] },
+    // [#2936] no-org member: same fail-closed deny on the write (see above); the
+    // global catalog stays VISIBLE on read (Layer 0 non-tenant, delta (c)).
+    no_org_member: { read: null, write: [{ id: DENY }] },
   },
   better_auth: {
     // [W2] better-auth-managed posture → superuser read bypass → null.
 
@@ -702,6 +702,34 @@ describe('SecurityPlugin', () => {
       expect(filter).toEqual(RLS_DENY_FILTER);
     });
 
+    it('[#2936] fail-closed on a CANONICAL `==` wildcard policy targeting a missing column → deny sentinel', async () => {
+      // The `=`-form twin above is the L692 test. This is its canonical-CEL
+      // sibling: pre-#2936 `extractTargetField` did not recognize `==`, so the
+      // field-existence net was INERT for it — the policy slipped through and
+      // compiled to a phantom `{ organization_id: … }` filter against a column
+      // the object lacks. The net now recognizes `==` and fails closed here,
+      // exactly as it always did for the legacy `=` form. Real seeds/business
+      // policies author equality as `==`, so this is the path that mattered.
+      const eqPolicySet: PermissionSet = {
+        name: 'member_default',
+        label: 'Member',
+        objects: { '*': { allowRead: true, allowCreate: true, allowEdit: true, allowDelete: true } },
+        rowLevelSecurity: [
+          { name: 'tenant_isolation', object: '*', operation: 'all', using: 'organization_id == current_user.organization_id' },
+        ],
+      } as any;
+      const plugin = new SecurityPlugin({ fallbackPermissionSet: 'member_default' });
+      const harness = makeMiddlewareCtx({
+        permissionSets: [eqPolicySet],
+        objectFields: ['id', 'name'], // no organization_id, tenancy not opted out
+        orgScoping: true,
+      });
+      await plugin.init(harness.ctx);
+      await plugin.start(harness.ctx);
+      const filter = await plugin.getReadFilter('task', { userId: 'u1', tenantId: 'org-1', positions: [], permissions: [] });
+      expect(filter).toEqual(RLS_DENY_FILTER);
+    });
+
     it('tenancy opt-out → undefined (not denied), matching the find-path', async () => {
       const plugin = new SecurityPlugin({ fallbackPermissionSet: 'member_default' });
       const harness = makeMiddlewareCtx({
 
@@ -2487,12 +2487,21 @@ export class SecurityPlugin implements Plugin {
    */
   private extractTargetField(using?: string): string | null {
     if (!using) return null;
-    // Match `field =` or `field IN`/`in`. Note: `\b` is omitted after `=`
-    // because `=` is non-word and the next char (space) is non-word too —
-    // a word boundary cannot exist between two non-word chars, so `=\b`
-    // would never match. We instead require the alternation token to be
-    // followed by whitespace or `(`.
-    const m = using.match(/^\s*([a-z_][a-z0-9_]*)\s*(?:=|IN|in)(?=\s|\()/);
+    // Match `field ==` (canonical CEL), `field =` (legacy SQL-ish), or
+    // `field IN`/`in`. [#2936] `==` MUST be listed before `=` in the
+    // alternation: alternation is ordered, so a bare `=` branch would match
+    // the first `=` of `==` and then the `(?=\s|\()` lookahead — seeing the
+    // SECOND `=`, not whitespace — would fail, leaving `==` unrecognized (the
+    // original bug: real seeds/business policies author equality as `==`, so
+    // the field-existence and tenancy-disabled safety nets that consume this
+    // were inert for them). `\b` is omitted after the operator because `=` is
+    // non-word and the next char (space) is non-word too — a word boundary
+    // cannot exist between two non-word chars — so we require the operator to
+    // be followed by whitespace or `(` instead. `!=`/`>`/`<`/`>=`/`<=` are
+    // deliberately NOT recognized: returning `null` keeps such a policy (the
+    // conservative default), matching the pre-#2936 behavior for any shape the
+    // regex did not match.
+    const m = using.match(/^\s*([a-z_][a-z0-9_]*)\s*(?:==|=|IN|in)(?=\s|\()/);
     return m ? m[1] : null;
   }
 }