test(preview): pre-warm all heavy pages + 2 retries (kill the slow-window flake) (#2483)

* test(preview): pre-warm all heavy pages + 2 retries (kill the slow-window flake)

The dominant smoke flake is a cold/slow-window page.goto timeout on a heavy SSR
page — the offender rotates (/, /models, /user/membership, /generate,
/purchase/buzz, /moderator/*) depending on which page the slow window hits.

Two-pronged:
 - Pre-warm EVERY heavy page the suite navigates in preview-auth.setup.ts (was
   only /, /models, /images). Each route JIT-compiles on first hit, so warming
   them up front (gold cookie for member pages, mod cookie for /moderator/*)
   removes the cold-start cost that caused most flakes. Sequential + best-effort
   (a slow warm GET still triggers the server-side compile).
 - retries 1 -> 2: a 2nd retry covers the residual mid-run slow window so it
   flakes-and-recovers instead of surfacing as a failure.

Pairs with the 60s->90s timeout raise (#2480). Together these target a
consistently-clean run — a prerequisite for eventually flipping smoke to gating.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* test(preview): raise setup timeout for cold-pod sequential warm-up (pr-review)

Addresses the pr-reviewer finding on this PR: the 8+ sequential warm-up GETs
(each 60s-capped) can, on a genuinely cold pod, cumulatively exceed the suite's
90s per-test timeout → the setup test times out → ALL dependent smoke tests are
SKIPPED (worse than the flake this PR fixes). The passing run only saw 1.7s
because the pod was already warm from verify-preview.

Fix: setup.setTimeout(480_000) — a ceiling for worst-case cold sequential
warm-up. Kept sequential (parallel heavy SSR renders OOM the single-replica pod,
per the config comment — the OOM is about concurrent renders, not specifically
playwright workers, so concurrent HTTP GETs are NOT a safe alternative here). The
ceiling doesn't slow the normal path: the setup still finishes as fast as the
warm-ups actually take.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

playwright.preview.config.ts

@@ -25,7 +25,11 @@ export default defineConfig({
   testMatch: /(^|\/)preview-.*\.(setup|spec)\.ts$/,
   fullyParallel: false,
   forbidOnly: !!process.env.CI,
-  retries: process.env.CI ? 1 : 0,
+  // 2 retries (was 1): the preview pod can hit a transient CPU-throttle window on a
+  // contended node; with pre-warm (preview-auth.setup.ts) covering cold-start, the
+  // residual flake is a mid-run slow window. A 2nd retry gives it another recovery
+  // chance so a slow window flakes-and-recovers instead of surfacing as a failure.
+  retries: process.env.CI ? 2 : 0,
   // Run SERIALLY. The preview is a single-replica, cold, resource-modest pod.
   // Concurrent loads of the heavy pages (/models, /images, /purchase/buzz,
   // /pricing, model detail) across multiple workers segfaulted it (exit 139),

tests/preview-auth.setup.ts

@@ -84,6 +84,14 @@ async function mintStorageState(role: PreviewRole): Promise<string> {
 }
 
 setup('mint preview sessions', async ({ request }) => {
+  // This setup runs the cold-pod warm-up below: ~9 SEQUENTIAL heavy-SSR GETs (each
+  // capped at 60s) so one route compiles at a time (parallel heavy renders OOM the
+  // single-replica pod). On a genuinely cold pod the cumulative warm-up can exceed
+  // the suite's 90s per-test timeout — and a setup timeout SKIPS every dependent
+  // smoke test (worse than the flake we're fixing). So give just this setup a large
+  // ceiling. It's a CEILING, not the runtime: the setup still finishes as fast as
+  // the warm-ups actually take (~2s when the pod is already warm from verify-preview).
+  setup.setTimeout(480_000);
   if (!SECRET) throw new Error('NEXTAUTH_SECRET is required to mint preview sessions');
   if (!PREVIEW_URL) throw new Error('PREVIEW_URL is required for preview smoke tests');
   const jwts: Partial<Record<PreviewRole, string>> = {};
@@ -92,16 +100,38 @@ setup('mint preview sessions', async ({ request }) => {
   }
 
   // Warm the freshly-deployed preview before the suite so the first real test
-  // doesn't pay the full cold-SSR cost (Next warm-up + JIT + DB pools). Warm the
-  // HEAVY listing pages too, not just `/` — those are the slow ones. They must be
-  // warmed AUTHENTICATED: on a preview the gate 307s an UNauthenticated /models to
-  // /login, so an anon GET wouldn't touch the real /models render path. Use the
-  // gold (gate-passing) cookie. Sequential + non-fatal; the suite + retries cover
-  // any miss.
+  // doesn't pay the full cold-SSR cost (Next warm-up + JIT + DB pools). Each route
+  // JIT-compiles on its first hit, so we warm EVERY heavy SSR page the suite then
+  // navigates — cold-page timeouts were the dominant smoke flake (a slow-window
+  // page.goto exceeding the nav budget, then passing on retry once warm). They must
+  // be warmed AUTHENTICATED: on a preview the gate 307s an UNauthenticated request
+  // to /login, so an anon GET wouldn't touch the real render path. Sequential (one
+  // concurrent heavy SSR at a time — the single-replica pod OOM'd under parallel
+  // heavy loads) + non-fatal (.catch): a slow warm-up GET still triggers the
+  // server-side compile even if the client times out, and the suite + retries
+  // cover any miss.
   const gold = jwts.gold;
   if (gold) {
     const headers = { cookie: `${COOKIE_NAME}=${gold}` };
-    for (const path of ['/', '/models', '/images']) {
+    // gold (gate-passing paid member) reaches all non-mod heavy pages the suite hits.
+    for (const path of [
+      '/',
+      '/models',
+      '/images',
+      '/user/membership',
+      '/generate',
+      '/purchase/buzz',
+      '/pricing',
+    ]) {
+      await request.get(path, { timeout: 60_000, headers }).catch(() => {});
+    }
+  }
+  const mod = jwts.mod;
+  if (mod) {
+    // /moderator/* render only for a moderator (gold would be bounced), so warm the
+    // moderation-spec pages with the mod cookie.
+    const headers = { cookie: `${COOKIE_NAME}=${mod}` };
+    for (const path of ['/moderator/reports', '/moderator/images']) {
       await request.get(path, { timeout: 60_000, headers }).catch(() => {});
     }
   }