diff --git a/packages/cli/test/metadata-type-schema-gate.test.ts b/packages/cli/test/metadata-type-schema-gate.test.ts
new file mode 100644
index 0000000000..72f150428f
--- /dev/null
+++ b/packages/cli/test/metadata-type-schema-gate.test.ts
@@ -0,0 +1,374 @@
+// Copyright (c) 2026 ObjectStack. Licensed under the Apache-2.0 license.
+
+/**
+ * Does the CLI hold metadata to the SAME schema the write path does? (#5000)
+ *
+ * #5000 measured that `os build` / `os validate` "never parse page metadata by
+ * `PageSchema`": an undeclared key on a page component was said to pass both
+ * commands and land in `dist/objectstack.json`, which would make the #4001
+ * acceptance line — "all three example apps `validate` clean" — empty evidence
+ * on the page surface, because nothing on that path ever parsed a page.
+ *
+ * Re-measured against `origin/main`, the claim does not hold. Both commands
+ * parse the WHOLE stack through `ObjectStackDefinitionSchema`, and its `pages`
+ * element is the very object `getMetadataTypeSchema('page')` returns — the same
+ * gate `MetadataManager.validate` and `GET /api/v1/meta` use. The issue's own
+ * repro and its own negative control both exit non-zero today.
+ *
+ * So this file is not the gate the issue asked for; it is the evidence the
+ * issue found missing. Nothing pinned either half of the claim, which is why a
+ * stale `packages/spec/dist` (AGENTS.md §9) or a refactor onto a lenient
+ * publish shape could reopen it without a single test turning red. Two claims,
+ * because they fail independently:
+ *
+ *   A. the CLI parses through the registry's schemas — one undeclared key, the
+ *      same verdict from both gates, for every registered metadata type. Three
+ *      carriers are structurally different and are asserted at their real
+ *      positions; one type (`api`) has a schema #4001 has not closed yet, so
+ *      there the claim is AGREEMENT plus "the author is still told", with the
+ *      gap filed rather than papered over (#5384);
+ *   B. the commands GATE on that parse — the issue's undeclared-key repro and
+ *      the #4001 batch-13 `responsiveStyles.large` → `.lg` negative control,
+ *      run through the real binary: non-zero exit, prescription in the output,
+ *      and `os build` writes no artifact.
+ *
+ * (A) without (B) is a strict schema whose verdict a command swallows — the
+ * #3782 shape. (B) without (A) is a command that gates on a schema nobody
+ * checked is the canonical one. Both have happened here before: #3782 wired
+ * four lints into `os build` alone, and #4409 found 23 of 26 rules running on
+ * a strict subset of the three authoring commands.
+ */
+
+import { describe, expect, it } from 'vitest';
+import { execFileSync } from 'node:child_process';
+import { existsSync, mkdtempSync, rmSync, writeFileSync } from 'node:fs';
+import { tmpdir } from 'node:os';
+import { join } from 'node:path';
+import { fileURLToPath } from 'node:url';
+import { ObjectStackDefinitionSchema, lintUnknownAuthoringKeys, formatUnknownAuthoringKey } from '@objectstack/spec';
+import { getMetadataTypeSchema, listMetadataTypeSchemaTypes } from '@objectstack/spec/kernel';
+
+const cliBin = join(fileURLToPath(new URL('.', import.meta.url)), '..', 'bin', 'run-dev.js');
+
+/** The key #5000 injected. Kept verbatim so the repro reads as the issue wrote it. */
+const INJECTED_KEY = 'aKeyPageComponentHasRejectedSinceADR0089';
+
+/**
+ * Registered metadata type → the stack-root collection the CLI parses it in.
+ *
+ * Asserted behaviourally (same undeclared key, same verdict on both gates)
+ * rather than by schema-instance identity: `@objectstack/spec` ships one bundle
+ * per entry point, so `@objectstack/spec/kernel`'s `getMetadataTypeSchema('page')`
+ * and the `PageSchema` embedded in the root entry's `ObjectStackDefinitionSchema`
+ * are equal-by-source copies that are never `===`. Identity is unobservable
+ * across that boundary; the verdict is what the author actually meets.
+ */
+const GATED_AT: Readonly<Record<string, string>> = {
+  object: 'objects',
+  hook: 'hooks',
+  seed: 'data',
+  mapping: 'mappings',
+  page: 'pages',
+  dashboard: 'dashboards',
+  app: 'apps',
+  action: 'actions',
+  report: 'reports',
+  dataset: 'datasets',
+  flow: 'flows',
+  job: 'jobs',
+  datasource: 'datasources',
+  email_template: 'emailTemplates',
+  doc: 'docs',
+  book: 'books',
+  permission: 'permissions',
+  position: 'positions',
+  agent: 'agents',
+  tool: 'tools',
+  skill: 'skills',
+};
+
+/**
+ * The three types the stack does NOT carry as a flat collection of the
+ * registry's own shape. Each is a structural difference between "how an app
+ * authors it" and "what a stored row looks like" — not a hole — so each gets
+ * its own placement in the assertion below rather than the generic one.
+ */
+const STRUCTURAL_EXCEPTIONS: Readonly<Record<string, string>> = {
+  field: 'authored INSIDE its object (`objects[].fields`), never as a stack-root collection; '
+    + 'ObjectSchema carries FieldSchema for it.',
+  translation: 'the stack authors locale → data BUNDLES (`TranslationBundleSchema`, a record); '
+    + 'the registry carries `TranslationItemSchema`, the per-row shape the runtime metadata API stores.',
+  view: 'the registry schema is the #3095 union over all three persisted view shapes (wire ViewItem, '
+    + 'defineView container, flattened personalization overlay); the stack authors the container '
+    + 'member, `ViewSchema`.',
+};
+
+/**
+ * Registered types whose SCHEMA is not closed yet, so "both gates reject" is
+ * not the claim to make — "both gates agree, and the author is still told" is.
+ *
+ * `api` is the live one. #5312 registered the type and the stack authors it at
+ * `apis:` (ADR-0121; note the neighbouring singular `api:` block, which is
+ * server-facing REST config, not metadata), but `ApiEndpointSchema` is still a
+ * plain `z.object` — the #4001 campaign has not reached it, and the strictness
+ * ledger still files all of `api/` as wire. An undeclared key on an endpoint
+ * is therefore DROPPED on both the write path and here, identically. Filed as
+ * #5384 (sub-issue of #4001) rather than papered over.
+ *
+ * What is asserted instead: the CLI is no looser than the write path, and the
+ * #3786 pre-parse layer still names the key, so the author is not left with
+ * silence. When #5384 closes the shape, the agreement assertions below go red
+ * — that is the ratchet working; move the row into `GATED_AT` then.
+ */
+const NOT_YET_CLOSED: Readonly<Record<string, { collection: string; tracking: string }>> = {
+  api: { collection: 'apis', tracking: '#5384 (sub-issue of #4001)' },
+};
+
+/** Every `unrecognized_keys` issue naming `INJECTED_KEY`, with its path. */
+function undeclaredKeyRejections(result: { success: boolean; error?: any }): string[] {
+  if (result.success) return [];
+  return (result.error.issues as any[])
+    .filter((i) => i.code === 'unrecognized_keys' && Array.isArray(i.keys) && i.keys.includes(INJECTED_KEY))
+    .map((i) => i.path.join('.'));
+}
+
+const MANIFEST = {
+  id: 'gate_probe',
+  name: 'Gate Probe',
+  namespace: 'gate_probe',
+  version: '1.0.0',
+  type: 'app',
+} as const;
+
+/** The issue's page, with `injected` deciding whether the defect is planted. */
+const pageWith = (component: Record<string, unknown>) => ({
+  name: 'gate_probe_page',
+  label: 'Gate Probe',
+  type: 'app',
+  template: 'default',
+  kind: 'full',
+  regions: [{ name: 'main', components: [component] }],
+});
+
+const CLEAN_COMPONENT = {
+  id: 'styling_root',
+  type: 'flex',
+  responsiveStyles: { large: { display: 'flex' } },
+  properties: { children: [] },
+};
+
+/** Run a CLI command in `dir`; returns its exit code and combined output. */
+function runCli(command: string, dir: string, args: string[] = []): { exitCode: number; output: string } {
+  try {
+    const output = execFileSync(process.execPath, [cliBin, command, ...args], {
+      cwd: dir,
+      encoding: 'utf8',
+      stdio: 'pipe',
+    });
+    return { exitCode: 0, output };
+  } catch (error: any) {
+    return { exitCode: error.status ?? 1, output: `${error.stdout ?? ''}${error.stderr ?? ''}` };
+  }
+}
+
+/**
+ * A config written as a plain literal — no `defineStack` / `definePage`.
+ *
+ * Load-bearing: those factories parse eagerly, so a config authored through
+ * them is rejected before the command's own gate is ever consulted. #5000's
+ * repro edited `examples/app-showcase`, where every page goes through
+ * `definePage`, so its exit code could not distinguish "the CLI gates" from
+ * "the factory threw". A literal isolates the command's own parse.
+ */
+function withConfig<T>(stack: Record<string, unknown>, body: (dir: string) => T): T {
+  const dir = mkdtempSync(join(tmpdir(), 'os-metadata-gate-'));
+  try {
+    writeFileSync(join(dir, 'objectstack.config.mjs'), `export default ${JSON.stringify(stack, null, 2)};\n`);
+    return body(dir);
+  } finally {
+    rmSync(dir, { recursive: true, force: true });
+  }
+}
+
+describe('the CLI parses metadata through the registry schemas (#5000)', () => {
+  it('classifies every registered metadata type', () => {
+    // A newly registered type with no classification fails here rather than
+    // quietly acquiring no CLI-side gate — the generalized form of #5000's
+    // worry, which was about exactly one type nobody had checked.
+    const classified = new Set([
+      ...Object.keys(GATED_AT),
+      ...Object.keys(STRUCTURAL_EXCEPTIONS),
+      ...Object.keys(NOT_YET_CLOSED),
+    ]);
+    const registered = listMetadataTypeSchemaTypes();
+    const unclassified = registered.filter((t) => !classified.has(t));
+    expect(
+      unclassified,
+      'a registered metadata type is in none of the three tables — decide which it is (gated, structurally '
+        + 'different, or a schema #4001 has not closed yet), so `os validate` cannot silently stop gating it. '
+        + 'This is the row `api` needed when #5312 registered it mid-flight.',
+    ).toEqual([]);
+    // And the reverse: a table row for a type nobody registers any more is a
+    // guard describing a surface that no longer exists.
+    const stale = [...classified].filter((t) => !registered.includes(t));
+    expect(stale, 'the table names metadata types that are no longer registered').toEqual([]);
+  });
+
+  it('reaches the same verdict as the write path on an undeclared key, per type', () => {
+    // Left: the gate `MetadataManager.validate` / `GET /api/v1/meta` / the
+    // Studio form use. Right: the schema `os validate` and `os build` parse
+    // the whole stack through. #5000's claim was that the right-hand column
+    // is blank for `page`; it is blank for nothing.
+    const writePathAccepts: string[] = [];
+    const cliAccepts: string[] = [];
+
+    for (const [type, collectionKey] of Object.entries(GATED_AT)) {
+      const registry = getMetadataTypeSchema(type);
+      expect(registry, `no registered schema for '${type}'`).toBeDefined();
+      if (undeclaredKeyRejections(registry!.safeParse({ [INJECTED_KEY]: 1 })).length === 0) {
+        writePathAccepts.push(type);
+      }
+      const viaCli = undeclaredKeyRejections(
+        ObjectStackDefinitionSchema.safeParse({ manifest: MANIFEST, [collectionKey]: [{ [INJECTED_KEY]: 1 }] }),
+      );
+      if (!viaCli.includes(`${collectionKey}.0`)) cliAccepts.push(`${type} (stack '${collectionKey}')`);
+    }
+
+    // Guard the guard: the detector must be able to say NO. A collection the
+    // stack does not declare is silently DROPPED (the root object is not
+    // strict — that is what the #3786 warning layer exists for), so an
+    // ungated type produces zero rejections here. If this ever came back
+    // non-empty, every row above would be passing on a detector that always
+    // says yes.
+    expect(
+      undeclaredKeyRejections(
+        ObjectStackDefinitionSchema.safeParse({ manifest: MANIFEST, notAStackCollection: [{ [INJECTED_KEY]: 1 }] }),
+      ),
+      'the undeclared-key detector reported a rejection for a collection the stack never declares',
+    ).toEqual([]);
+
+    expect(writePathAccepts, 'the metadata-type registry stopped rejecting undeclared keys for these types').toEqual([]);
+    expect(
+      cliAccepts,
+      'the stack schema `os validate` / `os build` parse through no longer rejects an undeclared key for these '
+        + 'types — either the collection is gone from the stack root (silently dropped, since the root is not '
+        + 'strict) or it now parses through a looser shape than the write path. That divergence is #5000.',
+    ).toEqual([]);
+  });
+
+  it('rejects an undeclared key on each structurally-different carrier', () => {
+    // `field` — inside its object.
+    expect(
+      undeclaredKeyRejections(
+        ObjectStackDefinitionSchema.safeParse({
+          manifest: MANIFEST,
+          objects: [{ name: 'gate_obj', label: 'Gate', fields: { title: { type: 'text', label: 'T', [INJECTED_KEY]: 1 } } }],
+        }),
+      ),
+    ).toContain('objects.0.fields.title');
+
+    // `translation` — inside a locale of the bundle.
+    expect(
+      undeclaredKeyRejections(
+        ObjectStackDefinitionSchema.safeParse({
+          manifest: MANIFEST,
+          translations: [{ 'en-US': { [INJECTED_KEY]: 1 } }],
+        }),
+      ),
+    ).toContain('translations.0.en-US');
+
+    // `view` — the container member of the registry union.
+    expect(
+      undeclaredKeyRejections(
+        ObjectStackDefinitionSchema.safeParse({
+          manifest: MANIFEST,
+          views: [{ name: 'gate_view', label: 'Gate', object: 'gate_obj', [INJECTED_KEY]: 1 }],
+        }),
+      ),
+    ).toContain('views.0');
+  });
+
+  it('is no looser than the write path on a type #4001 has not closed, and still names the key', () => {
+    for (const [type, { collection, tracking }] of Object.entries(NOT_YET_CLOSED)) {
+      const registry = getMetadataTypeSchema(type);
+      expect(registry, `no registered schema for '${type}'`).toBeDefined();
+
+      // Agreement, both directions. If the registry schema closes (that is
+      // what `tracking` is for), the first expectation flips and this row
+      // moves into GATED_AT — a deliberate step, not a surprise.
+      expect(
+        undeclaredKeyRejections(registry!.safeParse({ [INJECTED_KEY]: 1 })),
+        `${type}'s schema now rejects undeclared keys (${tracking} closed it?) — move it into GATED_AT`,
+      ).toEqual([]);
+      expect(
+        undeclaredKeyRejections(
+          ObjectStackDefinitionSchema.safeParse({ manifest: MANIFEST, [collection]: [{ [INJECTED_KEY]: 1 }] }),
+        ),
+        `the CLI rejects on '${collection}' while the write path accepts — a divergence in the other direction`,
+      ).toEqual([]);
+
+      // Not rejected is not the same as not reported: the #3786 pre-parse diff
+      // is what stands between the author and silence while the shape is open.
+      const reported = lintUnknownAuthoringKeys({ manifest: MANIFEST, [collection]: [
+        { name: 'gate_endpoint', path: '/api/v1/apps/gate_probe/things', method: 'GET', type: 'proxy', target: 'https://example.test', [INJECTED_KEY]: 1 },
+      ] } as Record<string, unknown>).map(formatUnknownAuthoringKey);
+      expect(
+        reported.join('\n'),
+        `an undeclared key on a '${type}' item is neither rejected nor reported — that is silent metadata loss`,
+      ).toContain(INJECTED_KEY);
+    }
+  });
+});
+
+describe('the authoring commands gate on that parse (#5000)', () => {
+  // Reverse verification, direction declared up front: the SAME stack without
+  // the planted key must exit 0. Without this control the three cases below
+  // would also pass if the stack failed for some unrelated reason — a green
+  // that proves nothing, which is the failure mode #5000 itself ran into.
+  it('accepts the control stack (no planted key)', () => {
+    const { exitCode, output } = withConfig({ manifest: MANIFEST, pages: [pageWith(CLEAN_COMPONENT)] }, (dir) =>
+      runCli('validate', dir),
+    );
+    expect(exitCode, `os validate rejected the CONTROL stack:\n${output}`).toBe(0);
+  }, 120_000);
+
+  it('os validate rejects an undeclared key on a page component, with the prescription', () => {
+    const { exitCode, output } = withConfig(
+      { manifest: MANIFEST, pages: [pageWith({ ...CLEAN_COMPONENT, [INJECTED_KEY]: 1 })] },
+      (dir) => runCli('validate', dir),
+    );
+    expect(exitCode, `os validate exited 0 on #5000's repro:\n${output}`).not.toBe(0);
+    expect(output).toContain(INJECTED_KEY);
+    // A rejection that does not say WHICH schema refused, and what changed,
+    // sends the author to the wrong file. ADR-0089 D3a is the decision.
+    expect(output).toContain('ADR-0089 D3a');
+  }, 120_000);
+
+  it('os build rejects the same stack and writes no artifact', () => {
+    const { exitCode, output, wroteArtifact } = withConfig(
+      { manifest: MANIFEST, pages: [pageWith({ ...CLEAN_COMPONENT, [INJECTED_KEY]: 1 })] },
+      (dir) => ({ ...runCli('build', dir), wroteArtifact: existsSync(join(dir, 'dist', 'objectstack.json')) }),
+    );
+    expect(exitCode, `os build exited 0 on #5000's repro:\n${output}`).not.toBe(0);
+    expect(output).toContain(INJECTED_KEY);
+    // #5000's second complaint: the artifact carried the bad value onward.
+    // The build emits from `result.data`, so a rejected parse emits nothing.
+    expect(wroteArtifact, 'os build wrote an artifact for a stack it rejected').toBe(false);
+  }, 120_000);
+
+  it("os validate turns red on #4001 batch 13's own negative control (`large` → `lg`)", () => {
+    // The control that did NOT turn red when #5000 was filed, which is what
+    // made the batch reach for a slot-tracing probe instead. It is red now:
+    // `ResponsiveStylesSchema` closed, and the CLI parses through it.
+    const { exitCode, output } = withConfig(
+      {
+        manifest: MANIFEST,
+        pages: [pageWith({ ...CLEAN_COMPONENT, responsiveStyles: { lg: { display: 'flex' } } })],
+      },
+      (dir) => runCli('validate', dir),
+    );
+    expect(exitCode, `os validate exited 0 on a page styled under the wrong breakpoint vocabulary:\n${output}`).not.toBe(0);
+    expect(output).toContain('responsiveStyles');
+    expect(output).toContain('large');
+  }, 120_000);
+});