feat(DPoP): UI: bcast events to support multi tab

add also playwright test with the multi tab

tests/e2e/playwright.config.ts

@@ -66,13 +66,14 @@ export default defineConfig({
       // Verbose startup so a CI webServer-readiness timeout shows where the
       // server stalls (DB connect, migrations, bind) instead of nothing.
       RUST_LOG: 'info,oxicloud=debug,sqlx=warn,tower_http=info',
-      // OPAQUE substrate is off for the E2E suite — Hurl exercises it via
-      // `tests/common/server.env`; the SPA-facing coverage suite doesn't
-      // need the boot-time init nor the ~200 KiB WASM client. Blanking
-      // the inherited commonEnv values takes the DI factory's
-      // `effective_mode == Off` short-circuit.
-      OXICLOUD_AUTH_OPAQUE_MODE: 'off',
-      OXICLOUD_AUTH_OPAQUE_SERVER_SETUP: '',
+      // OPAQUE + DPoP are inherited from `../common/server.env`:
+      //   OXICLOUD_AUTH_OPAQUE_MODE=migrate  (Phase 2 silent-migration
+      //     on first legacy login, Phase 4 refusal thereafter)
+      //   OXICLOUD_DPOP_MODE=required        (verify every proof; unbound
+      //     sessions still exempt per Gate 5 design)
+      // Testing under the production shape catches breakage where the
+      // SPA's fetch interceptor or the migration hook regresses in
+      // ways that only surface in a real browser + real crypto.
     },
   },
 });

tests/e2e/scenarios/helpers.ts

@@ -98,22 +98,61 @@ export async function seedAdmin(baseURL: string, admin = TEST_ADMIN): Promise<vo
 }
 
 /**
- * Authenticate the page's browser context via the API, so a subsequent
- * `page.goto()` loads already signed in — no UI clicks. Selector-independent,
- * which keeps it robust while the SPA login markup is still in flux.
+ * Authenticate the page's browser context, ready for subsequent
+ * `page.goto()` calls to load already-signed-in.
  *
- * `POST /api/auth/login` is CSRF-exempt and sets the auth cookies on the
- * context's (shared) cookie jar, so `page.request` here authenticates the
- * page too. Use this at the top of any spec that needs an authenticated app
- * (and in the codegen recorder, so you record post-login flows).
+ * Uses the SPA's real login flow (`page.goto('/login')` → fill form
+ * → submit) rather than a bare `POST /api/auth/login`, so this works
+ * correctly under both auth modes the test env supports:
+ *
+ *   * `OXICLOUD_AUTH_OPAQUE_MODE=off` — SPA does legacy login,
+ *     server accepts.
+ *   * `OXICLOUD_AUTH_OPAQUE_MODE=migrate` — first login legacy-
+ *     succeeds + silently mints an OPAQUE envelope (Phase 2 hook);
+ *     every subsequent login the SPA detects the envelope via
+ *     `/api/auth/opaque/login/lookup` and does the full KE1/KE3
+ *     OPAQUE handshake. Legacy `POST /api/auth/login` would 403
+ *     with `opaque_migrated_use_opaque` (Phase 4) from the second
+ *     login on — that's what the old bare-POST apiLogin used to
+ *     hit as soon as OPAQUE went from `off` to `migrate`.
+ *   * `OXICLOUD_DPOP_MODE=required` — the SPA computes and sends
+ *     `dpop_jkt` in the login body; the session is created bound.
+ *     A bare-POST wouldn't include it, so subsequent requests
+ *     wouldn't get DPoP-signed. Going through the SPA keeps the
+ *     end-to-end flow honest.
+ *
+ * Overhead vs the old direct POST: ~200-500 ms per test to load
+ * `/login`, submit, and wait for the post-login redirect. Runs
+ * once per test (from `beforeEach`), so the total suite tax is
+ * modest and the coverage payoff is real.
  */
 export async function apiLogin(page: Page, admin = TEST_ADMIN): Promise<void> {
-  const res = await page.request.post('/api/auth/login', {
-    data: { username: admin.username, password: admin.password },
-  });
-  if (!res.ok()) {
-    throw new Error(`apiLogin failed: ${res.status()} ${await res.text()}`);
+  // Idempotence check — many specs' beforeEach + test body both call
+  // apiLogin; the old bare-POST version was a no-op on a live
+  // session, and callers depend on that. Under UI-driven login,
+  // navigating to /login when already authenticated triggers the
+  // SPA's layout guard to redirect away → the login form never
+  // renders → the fill() below times out. Probe /api/auth/me FIRST:
+  // 2xx means we're already signed in as SOMEONE. If that's the
+  // right admin, no-op; otherwise fall through to a fresh login.
+  const probe = await page.request.get('/api/auth/me').catch(() => null);
+  if (probe?.ok()) {
+    const body = (await probe.json().catch(() => ({}))) as { username?: string };
+    if (body.username === admin.username) return;
   }
+
+  await page.goto('/login');
+  await page.locator('[data-testid="login-username-input"]').fill(admin.username);
+  await page.locator('[data-testid="login-password-input"]').fill(admin.password);
+  await page.locator('[data-testid="login-submit-btn"]').click();
+  // Post-login the SPA's `goto(redirectTarget)` sends the user
+  // to `/files` (default) or a `?redirect=` target. Match the
+  // default with a glob — the same shape `uiLogin` uses in
+  // `spa/coverage-helpers.ts` and that Playwright handles well
+  // under SvelteKit's client-side navigation. The 15s ceiling
+  // covers the OPAQUE-post-migration path: WASM load + KE1 +
+  // KE3 + Argon2id.
+  await page.waitForURL('**/files**', { timeout: 15_000 });
 }
 
 /**

tests/e2e/spa/dpop-multi-tab.spec.ts

@@ -0,0 +1,99 @@
+import { test, expect, uiLogin } from './coverage-helpers';
+
+/**
+ * SPA · DPoP multi-tab coverage — Gate 8 follow-up.
+ *
+ * IndexedDB, cookies, and `BroadcastChannel` are shared across every
+ * tab of a single Playwright `BrowserContext`. That's the correct
+ * shape for testing the multi-tab DPoP invariants:
+ *
+ *   * shared keypair — a second tab opened after login already sees
+ *     the first tab's persisted keypair via IndexedDB, so both tabs
+ *     sign requests with the same JWK thumbprint (`dpop_jkt`) →
+ *     server accepts both under a single bound session.
+ *   * `BroadcastChannel('oxicloud-session-cleared')` — logout on
+ *     one tab must cause the other tab's root layout to reset the
+ *     session store and redirect to `/login` synchronously, without
+ *     waiting for a network round trip to 401. See
+ *     `frontend/src/lib/auth/session-broadcast.ts`.
+ *
+ * Runs under `OXICLOUD_AUTH_OPAQUE_MODE=migrate` +
+ * `OXICLOUD_DPOP_MODE=required` inherited from
+ * `tests/common/server.env` — so the actual OPAQUE login handshake
+ * fires (WASM client → KE1 → KE3) and every subsequent request
+ * carries a DPoP proof the middleware verifies.
+ */
+test.describe('SPA · DPoP multi-tab', () => {
+  test('a second tab shares the first tab\'s DPoP keypair (IndexedDB)', async ({ context }) => {
+    const tabA = await context.newPage();
+    await uiLogin(tabA);
+    // Sanity: tab A landed on an authenticated view.
+    await expect(tabA.getByTestId('appshell-logo-link')).toBeVisible();
+
+    // Second tab in the same context — cookies + IndexedDB shared.
+    const tabB = await context.newPage();
+    // Deep-link straight into an authenticated route. If the session
+    // cookie is shared (it is — cookies are per-context) AND the
+    // DPoP keypair is shared (it is — IndexedDB is per-origin per-
+    // context), tab B loads without redirecting to /login.
+    await tabB.goto('/files');
+    await expect(tabB.getByTestId('appshell-logo-link')).toBeVisible({ timeout: 15_000 });
+
+    // Both tabs' auth store agrees on the same user id — proves the
+    // shared cookie + shared keypair combination actually authorised
+    // an API call under DPoP=required against a bound session.
+    const [uidA, uidB] = await Promise.all([
+      tabA.evaluate(async () => {
+        const res = await fetch('/api/auth/me', { credentials: 'same-origin' });
+        return res.ok ? ((await res.json()) as { id: string }).id : null;
+      }),
+      tabB.evaluate(async () => {
+        const res = await fetch('/api/auth/me', { credentials: 'same-origin' });
+        return res.ok ? ((await res.json()) as { id: string }).id : null;
+      })
+    ]);
+    expect(uidA).not.toBeNull();
+    expect(uidB).toBe(uidA);
+  });
+
+  test('logging out on one tab redirects the other via BroadcastChannel', async ({ context }) => {
+    const tabA = await context.newPage();
+    await uiLogin(tabA);
+
+    const tabB = await context.newPage();
+    await tabB.goto('/files');
+    await expect(tabB.getByTestId('appshell-logo-link')).toBeVisible({ timeout: 15_000 });
+
+    // Log out from tab A. Bypass the user-menu UI (which drifts as
+    // the shell markup evolves) — call `/api/auth/logout` directly
+    // then post to the BroadcastChannel by hand. Same shape as
+    // `endpoints/auth.ts::logout()` — the two side-effects the SPA
+    // does after a successful server logout are (a) wipe DPoP
+    // state (moot here since tab A is about to close/redirect) and
+    // (b) broadcast, which is exactly what we simulate.
+    await tabA.evaluate(async () => {
+      const csrf =
+        document.cookie
+          .split(';')
+          .map((c) => c.trim())
+          .find((c) => c.startsWith('oxicloud_csrf='))
+          ?.slice('oxicloud_csrf='.length) ?? '';
+      const res = await fetch('/api/auth/logout', {
+        method: 'POST',
+        credentials: 'same-origin',
+        headers: { 'Content-Type': 'application/json', 'x-csrf-token': csrf },
+        body: '{}'
+      });
+      if (!res.ok) throw new Error(`logout returned ${res.status}`);
+      new BroadcastChannel('oxicloud-session-cleared').postMessage({
+        kind: 'session_cleared',
+        at: Date.now()
+      });
+    });
+
+    // Tab B should navigate to /login on its own. No API call
+    // needed — the BroadcastChannel handler in the root layout
+    // does session.reset() + goto('/login').
+    await tabB.waitForURL('**/login**', { timeout: 5_000 });
+  });
+});

tests/e2e/spa/favorites.spec.ts

@@ -21,8 +21,19 @@ test('favorite, view, and unfavorite a folder', async ({ page }) => {
   await page.goto('/files');
   await expect(page.getByTestId(name)).toBeVisible({ timeout: 15_000 });
   await page.getByTestId(name).click({ button: 'right' });
+  // The context-menu `favorite` click is fire-and-forget in the SPA
+  // (closeContext() runs before the POST) — the test's next
+  // navigation can race the write. Wait for the actual POST to
+  // land before going to /favorites so the list-fetch there sees
+  // the new row committed. The batch test doesn't need this because
+  // it queues 2 POSTs sequentially, which naturally gives the first
+  // one time to commit.
+  const favorited = page.waitForResponse(
+    (r) => r.url().includes('/api/favorites') && r.request().method() === 'POST' && r.ok()
+  );
   await page.getByTestId('files-ctx-favorite-item').click();
   await expect(page.getByTestId('files-context-menu')).toHaveCount(0);
+  await favorited;
 
   await page.goto('/favorites');
   const row = page.getByTestId(name);