diff --git a/e2e/lib/ui.js b/e2e/lib/ui.js
index 81c2659..42bdf89 100644
--- a/e2e/lib/ui.js
+++ b/e2e/lib/ui.js
@@ -4,7 +4,15 @@
 export async function dismissCookies(page) {
   const banner = page.locator('.cookie-consent-banner')
   if (await banner.count()) {
+    // Click to persist the consent choice so it doesn't reappear on
+    // later navigation...
     await page.getByRole('button', { name: 'Save choice' }).click().catch(() => {})
-    await banner.waitFor({ state: 'hidden' }).catch(() => {})
+    await banner.waitFor({ state: 'hidden', timeout: 5000 }).catch(() => {})
+    // ...then forcibly remove any node still in the DOM. The dismiss
+    // click occasionally doesn't land before a test clicks a catalog
+    // footer checkbox (flaky over the deployed edge), and a lingering
+    // fixed banner intercepts those pointer events. Removing the node
+    // makes the dismissal deterministic.
+    await banner.evaluate((el) => el.remove()).catch(() => {})
   }
 }
diff --git a/e2e/playwright.config.js b/e2e/playwright.config.js
index fb5024a..db83667 100644
--- a/e2e/playwright.config.js
+++ b/e2e/playwright.config.js
@@ -1,9 +1,23 @@
 import { defineConfig } from '@playwright/test'
 
+// The metadata specs sign in, navigate, and (SLICE-4/5) write real commits,
+// so a handful of steps are timing-sensitive: the cookie-consent banner's
+// dismiss animation, first-render of the detail panel, and the round trip
+// after a write. These flake intermittently on a busy local box and more so
+// against a deployed host (network latency). `retries` makes the suite robust
+// to that (and finally makes `trace: 'on-first-retry'` meaningful); the
+// timeouts are bumped a notch for deployed runs over the public edge.
+const DEPLOYED = !!process.env.E2E_TEST_AUTH_SECRET
+
 export default defineConfig({
   testDir: '.',
-  timeout: 30_000,
-  expect: { timeout: 10_000 },
+  timeout: DEPLOYED ? 60_000 : 45_000,
+  expect: { timeout: DEPLOYED ? 20_000 : 12_000 },
+  retries: 2,
+  // The metadata specs run in order against one seeded collection and write
+  // real commits (SLICE-4/5); parallel workers would race on shared state —
+  // and on a deployed host, on concurrent git pushes through the bot. Serialize.
+  workers: 1,
   use: {
     baseURL: process.env.BASE_URL || 'http://localhost:8080',
     trace: 'on-first-retry',