fix(desktop): stop rate-limited reconnect backfill from tearing down the authenticated socket (#4990)

## Problem

Users on v0.5.5 report "Can't reach the relay" toggling with brief
"connected" flashes (field reports; also the macOS confirmation in
#4908). #4737 closed the stuck-reconnect gaps; this is the opposite
failure: the client redials fine, but then kills its own healthy socket.

Mechanism (all on `main`):

1. AUTH succeeds → session emits `connected`
(`relayClientSession.ts:583`), then awaits `replayLiveSubscriptions()`.
2. Paged channel backfill issues history REQs
(`relayReconnectReplay.ts`, page limit 500).
3. A `CLOSED rate-limited:` on a **history** REQ arms the rate-limit
gate but still rejects the history promise
(`relayClosedRecovery.ts:38-51`).
4. The rejection escapes `replayLiveSubscriptions()` →
`resetConnection()` tears down the authenticated socket.
5. Reconnect → AUTH OK → replay rate-limited again → loop. Each
iteration re-spends the rate-limit budget, so the loop is
self-sustaining.

## Fix

Contain backfill failures inside the replay. Each subscription's paged
backfill now retries behind the rate-limit gate up to
`PAGE_REPLAY_MAX_ATTEMPTS` (3), then degrades to live-only **for this
connection**. Socket health no longer depends on backfill success.
Nothing is lost: the replay cursor (`lastSeenCreatedAt`) only advances
on delivered events, so the next reconnect replays the same missed
window.

## Red/green proof

- Commit 1 (Pinky): e2e injecting `CLOSED rate-limited:` into the
mid-replay history REQ — **red on main** (expected 1 reconnect dial,
observed 2; connected-flash then teardown).
- Commit 2 (this fix): same test **green unchanged** — one dial, state
stays `connected` through the rate-limit hint plus the next backoff
window.

Why existing coverage missed it: the prior rate-limit e2e pre-armed the
gate *before* replay (replay politely waits), and the CLOSED-injection
test targeted a *live* subscription (which has its own retry path).
Nobody injected back-pressure from the history REQ itself.

## Verification

- `pnpm test`: 4374/4374 pass.
- `playwright test tests/e2e/relay-reconnect.spec.ts`: 14/14 pass,
including the new spec.
- `tsc --noEmit` clean; Biome clean on touched files (pre-existing
warnings on main in `personaCatalogRelay.test.mjs` / `terminal.css`
untouched).

## Not addressed here (follow-ups from the same field reports)

- AUTH terminal latch is too aggressive for relay-internal `error:`
rejections (3 strikes during a relay bad window → stuck until
click/relaunch; #4908).
- Server-side: `relay.drainJitterMs` (#4542) defaults to 0 — enabling it
on the hosted relay removes the deploy thundering herd that triggers
these rate-limit storms.

---------

Signed-off-by: Pinky <44b8e82baa6e0e254e0208d68f335c283c94e7b78dd1fa10d5a49d3f13dd0435@buzz.block.builderlab.xyz>
Signed-off-by: Brain <21994759fc7a6fa6b965551d35cfd7897d262f2495467f2d78694ddcfa6a5c7e@buzz.block.builderlab.xyz>
Co-authored-by: Pinky <44b8e82baa6e0e254e0208d68f335c283c94e7b78dd1fa10d5a49d3f13dd0435@buzz.block.builderlab.xyz>
Co-authored-by: Brain <21994759fc7a6fa6b965551d35cfd7897d262f2495467f2d78694ddcfa6a5c7e@buzz.block.builderlab.xyz>

desktop/src/shared/api/relayClientShared.ts

@@ -60,6 +60,17 @@ type LiveSubscription = {
   onEvent: (event: RelayEvent) => void;
   resolveReady?: () => void;
   lastSeenCreatedAt?: number;
+  /**
+   * Lower bound of a reconnect backfill window that has not yet completed.
+   *
+   * Events on the restored live REQ advance `lastSeenCreatedAt` regardless of
+   * backfill success, so after an exhausted backfill the cursor alone would
+   * make the next reconnect skip the unresolved older window — silent message
+   * loss. This floor is pinned when paging starts and cleared only when a
+   * backfill pass completes; the next replay starts from
+   * `min(pendingReplaySince, cursor window)`.
+   */
+  pendingReplaySince?: number;
   closedRetryAttempt?: number;
   closedRetryTimeout?: number;
 };

desktop/src/shared/api/relayReconnectReplay.test.mjs

@@ -3,11 +3,13 @@ import test from "node:test";
 
 import {
   buildReconnectReplayFilter,
+  PAGE_REPLAY_MAX_ATTEMPTS,
   replayLiveSubscriptions,
   REPLAY_BATCH_SIZE,
   shouldPageReconnectReplay,
 } from "./relayReconnectReplay.ts";
 import { buildChannelFilter } from "./relayChannelFilters.ts";
+import { prepareSubscriptionEvent } from "./relayClosedRecovery.ts";
 
 // ── Fake-timer + Date.now setup for gate tests ────────────────────────────────
 
@@ -600,6 +602,254 @@ test("batch-1 arms gate mid-replay: batch-2 is withheld until gate expires", asy
   );
 });
 
+// ── Backfill failure containment ─────────────────────────────────────────────
+
+test("history backfill rejection never escapes replayLiveSubscriptions", async () => {
+  resetGate(0);
+  const filter = buildChannelFilter("channel-1", 50);
+  const subscriptions = new Map([
+    [
+      "live-1",
+      {
+        mode: "live",
+        filter,
+        onEvent: () => {},
+        lastSeenCreatedAt: 1000,
+      },
+    ],
+  ]);
+
+  let historyCalls = 0;
+  // Must resolve — a rejection here is the socket-killing flap regression.
+  await replayLiveSubscriptions({
+    subscriptions,
+    now: 2000,
+    sendRaw: async () => {},
+    requestHistory: async () => {
+      historyCalls++;
+      throw new Error("rate-limited: quota exceeded; retry in 4s");
+    },
+  });
+
+  assert.equal(
+    historyCalls,
+    PAGE_REPLAY_MAX_ATTEMPTS,
+    "backfill must retry a bounded number of times, then degrade",
+  );
+});
+
+test("backfill retry waits out the armed gate, then succeeds", async () => {
+  resetGate(0);
+  const delivered = [];
+  const filter = buildChannelFilter("channel-1", 50);
+  const subscriptions = new Map([
+    [
+      "live-1",
+      {
+        mode: "live",
+        filter,
+        onEvent: (event) => delivered.push(event),
+        lastSeenCreatedAt: 1000,
+      },
+    ],
+  ]);
+
+  const attemptAtMs = [];
+  let armGate;
+  const gateArmed = new Promise((resolve) => {
+    armGate = resolve;
+  });
+  const replayPromise = replayLiveSubscriptions({
+    subscriptions,
+    now: 2000,
+    sendRaw: async () => {},
+    requestHistory: async () => {
+      attemptAtMs.push(fakeNow);
+      if (attemptAtMs.length === 1) {
+        // Mirror relayClosedRecovery: the CLOSED handler arms the gate
+        // before rejecting the history promise.
+        activateRateLimit(4);
+        armGate();
+        throw new Error("rate-limited: quota exceeded; retry in 4s");
+      }
+      return [event("recovered", 1500)];
+    },
+  });
+
+  // Wait until the gate is actually armed, then expire it. The retry loop is
+  // (or will be) suspended in waitForRateLimit; expiring the gate releases it.
+  await gateArmed;
+  tickTo(4_001);
+  await replayPromise;
+
+  assert.equal(attemptAtMs.length, 2, "one failure, one retry");
+  assert.ok(
+    attemptAtMs[1] >= 4_001,
+    "retry must not fire before the rate-limit gate expires",
+  );
+  assert.deepEqual(
+    delivered.map((e) => e.id),
+    ["recovered"],
+    "the retried backfill must deliver its events",
+  );
+});
+
+test("backfill retry aborts when the subscription was replaced", async () => {
+  resetGate(0);
+  const filter = buildChannelFilter("channel-1", 50);
+  const subscription = {
+    mode: "live",
+    filter,
+    onEvent: () => {},
+    lastSeenCreatedAt: 1000,
+  };
+  const subscriptions = new Map([["live-1", subscription]]);
+
+  let historyCalls = 0;
+  await replayLiveSubscriptions({
+    subscriptions,
+    now: 2000,
+    sendRaw: async () => {},
+    requestHistory: async () => {
+      historyCalls++;
+      // Simulate the subscription being torn down while the REQ is in flight.
+      subscriptions.delete("live-1");
+      throw new Error("rate-limited: quota exceeded; retry in 4s");
+    },
+  });
+
+  assert.equal(
+    historyCalls,
+    1,
+    "no retry may target a subscription that no longer exists",
+  );
+});
+
+test("exhausted backfill pins the floor: next replay still requests the original window after live events advance the cursor", async () => {
+  // The blocking review scenario on PR #4990: cursor=1000, all backfill
+  // attempts fail, a live event at 2100 then advances lastSeenCreatedAt via
+  // prepareSubscriptionEvent. Without the pinned floor, the next reconnect
+  // would start near 2095 and silently skip 1001..1999.
+  resetGate(0);
+  const filter = buildChannelFilter("channel-1", 50);
+  const subscription = {
+    mode: "live",
+    filter,
+    onEvent: () => {},
+    lastSeenCreatedAt: 1000,
+  };
+  const subscriptions = new Map([["live-1", subscription]]);
+
+  // Reconnect 1: every backfill attempt is rate-limited.
+  await replayLiveSubscriptions({
+    subscriptions,
+    now: 2000,
+    sendRaw: async () => {},
+    requestHistory: async () => {
+      throw new Error("rate-limited: quota exceeded; retry in 4s");
+    },
+  });
+  assert.equal(
+    subscription.pendingReplaySince,
+    995,
+    "exhausted backfill must pin the unresolved window's lower bound",
+  );
+
+  // A live event arrives through the normal cursor path.
+  prepareSubscriptionEvent(subscription, event("live-newer", 2100));
+  assert.equal(subscription.lastSeenCreatedAt, 2100);
+
+  // Reconnect 2: backfill now succeeds. It must request the ORIGINAL window.
+  const historyFilters = [];
+  await replayLiveSubscriptions({
+    subscriptions,
+    now: 2200,
+    sendRaw: async () => {},
+    requestHistory: async (filter) => {
+      historyFilters.push(filter);
+      return [];
+    },
+  });
+
+  assert.equal(historyFilters.length, 1);
+  assert.equal(
+    historyFilters[0].since,
+    995,
+    "replay must start from the pinned floor, not the advanced cursor",
+  );
+  assert.equal(
+    subscription.pendingReplaySince,
+    undefined,
+    "a completed backfill must clear the pinned floor",
+  );
+
+  // Reconnect 3: with the floor cleared, replay returns to the cursor.
+  const laterFilters = [];
+  await replayLiveSubscriptions({
+    subscriptions,
+    now: 2300,
+    sendRaw: async () => {},
+    requestHistory: async (filter) => {
+      laterFilters.push(filter);
+      return [];
+    },
+  });
+  assert.equal(
+    laterFilters[0].since,
+    2095,
+    "after recovery the cursor governs again",
+  );
+});
+
+test("in-flight stale abort keeps the pinned floor for the superseding connection", async () => {
+  // Race from re-review of b70a6716d/c493d378b: production supersession bumps
+  // the connection GENERATION while the same subscription key and object
+  // survive in the map. The identity guard alone stays true, so only the
+  // combined guard (outer isActive && identity) aborts the stale pass. That
+  // abort must NOT count as completion — the pinned floor belongs to the
+  // superseding connection's replay.
+  resetGate(0);
+  const filter = buildChannelFilter("channel-1", 50);
+  const subscription = {
+    mode: "live",
+    filter,
+    onEvent: () => {},
+    lastSeenCreatedAt: 1000,
+  };
+  const subscriptions = new Map([["live-1", subscription]]);
+
+  let generationActive = true;
+  let historyCalls = 0;
+  await replayLiveSubscriptions({
+    subscriptions,
+    now: 2000,
+    sendRaw: async () => {},
+    isActive: () => generationActive,
+    requestHistory: async () => {
+      historyCalls++;
+      // Connection A is superseded while the REQ is in flight: the generation
+      // advances, but the subscription keeps its key AND object identity —
+      // exactly what production supersession does.
+      generationActive = false;
+      // A full page would otherwise continue paging — the post-await
+      // combined guard must abort instead.
+      return eventRange("full", 1001, 500);
+    },
+  });
+
+  assert.equal(historyCalls, 1, "stale generation must stop paging");
+  assert.equal(
+    subscriptions.get("live-1"),
+    subscription,
+    "precondition: key and object survive supersession untouched",
+  );
+  assert.equal(
+    subscription.pendingReplaySince,
+    995,
+    "a stale-generation abort must not clear the floor the new connection needs",
+  );
+});
+
 // ── Teardown ──────────────────────────────────────────────────────────────────
 
 test("teardown — restore Date.now", () => {

desktop/src/shared/api/relayReconnectReplay.ts

@@ -13,6 +13,22 @@ const RECONNECT_REPLAY_SKEW_SECS = 5;
 export const RECONNECT_REPLAY_PAGE_LIMIT = 500;
 export const RECONNECT_REPLAY_PAGE_CONCURRENCY = 4;
 
+/**
+ * Maximum attempts for one subscription's paged history backfill.
+ *
+ * Backfill failures must never escape `replayLiveSubscriptions`: by the time
+ * paging starts, every live REQ has already been re-established on a healthy,
+ * authenticated socket. Letting a history rejection propagate makes the
+ * session tear that socket down (`resetConnection`) and reconnect straight
+ * into the same rate-limit window — the "briefly connected → can't reach the
+ * relay" flap loop. Instead each sub retries behind the rate-limit gate a
+ * bounded number of times, then degrades to live-only for this connection.
+ * The window's lower bound is pinned in `pendingReplaySince` while unresolved
+ * (live events advance `lastSeenCreatedAt` regardless of backfill success),
+ * so the next reconnect still requests the missed window.
+ */
+export const PAGE_REPLAY_MAX_ATTEMPTS = 3;
+
 /**
  * Maximum live subscriptions sent per relay REQ burst during reconnect.
  *
@@ -78,6 +94,16 @@ export function shouldPageReconnectReplay(filter: RelaySubscriptionFilter) {
   );
 }
 
+/**
+ * Page one subscription's missed-window history.
+ *
+ * Returns `true` only when the window was genuinely completed (short page or
+ * boundary reached). Returns `false` when the pass aborted because the
+ * connection went stale (`isActive()` false) — callers must NOT treat that as
+ * completion: the same subscription object is shared with the superseding
+ * connection, and clearing its pinned `pendingReplaySince` on a stale abort
+ * would erase the floor the new connection still needs.
+ */
 export async function replayReconnectHistoryPages({
   subscription,
   since,
@@ -90,11 +116,11 @@ export async function replayReconnectHistoryPages({
   until: number;
   isActive: () => boolean;
   requestHistory: (filter: RelaySubscriptionFilter) => Promise<RelayEvent[]>;
-}) {
+}): Promise<boolean> {
   let pageUntil = until;
 
   while (pageUntil >= since) {
-    if (!isActive()) return;
+    if (!isActive()) return false;
 
     const events = await requestHistory(
       buildReconnectReplayFilter(
@@ -105,17 +131,18 @@ export async function replayReconnectHistoryPages({
       ),
     );
 
-    if (!isActive()) return;
+    if (!isActive()) return false;
 
     for (const event of events) subscription.onEvent(event);
-    if (events.length < RECONNECT_REPLAY_PAGE_LIMIT) return;
+    if (events.length < RECONNECT_REPLAY_PAGE_LIMIT) return true;
 
     const oldestCreatedAt = events[0]?.created_at;
-    if (oldestCreatedAt === undefined || oldestCreatedAt <= since) return;
+    if (oldestCreatedAt === undefined || oldestCreatedAt <= since) return true;
 
     pageUntil =
       oldestCreatedAt < pageUntil ? oldestCreatedAt : oldestCreatedAt - 1;
   }
+  return true;
 }
 
 export async function replayLiveSubscriptions({
@@ -167,13 +194,21 @@ export async function replayLiveSubscriptions({
         entry[1].mode === "live",
     )
     .map(([subId, subscription]) => {
-      const replaySince =
+      const cursorSince =
         subscription.lastSeenCreatedAt === undefined
           ? undefined
           : Math.max(
               0,
               subscription.lastSeenCreatedAt - RECONNECT_REPLAY_SKEW_SECS,
             );
+      // A pinned floor from a previously failed backfill takes precedence
+      // over the cursor: live events kept advancing `lastSeenCreatedAt`
+      // while the older window stayed unresolved, and starting from the
+      // cursor would skip it permanently.
+      const replaySince =
+        cursorSince === undefined
+          ? subscription.pendingReplaySince
+          : Math.min(cursorSince, subscription.pendingReplaySince ?? Infinity);
       const shouldPageReplay =
         replaySince !== undefined &&
         shouldPageReconnectReplay(subscription.filter);
@@ -237,13 +272,53 @@ export async function replayLiveSubscriptions({
     ),
     pageReplayConcurrency,
     async ({ subId, subscription, replaySince }) => {
-      await replayReconnectHistoryPages({
-        subscription,
-        since: replaySince,
-        until: now,
-        isActive: () => subscriptions.get(subId) === subscription,
-        requestHistory,
-      });
+      // Backfill is best-effort: a failure here (typically a `rate-limited:`
+      // CLOSED on a history REQ) must never escape to the session and tear
+      // down the healthy, authenticated socket carrying the live REQs — that
+      // is the connect→drop flap loop. Retry behind the gate a bounded number
+      // of times, then degrade to live-only for this connection.
+      //
+      // Pin the window's lower bound before the first attempt: events on the
+      // already-restored live REQ advance `lastSeenCreatedAt` independently
+      // of backfill success, so without the pin an exhausted backfill
+      // followed by one live event would make the next reconnect skip the
+      // unresolved window permanently. Cleared only on a completed pass.
+      subscription.pendingReplaySince = replaySince;
+      for (let attempt = 1; attempt <= PAGE_REPLAY_MAX_ATTEMPTS; attempt++) {
+        try {
+          const completed = await replayReconnectHistoryPages({
+            subscription,
+            since: replaySince,
+            until: now,
+            // Both guards are required. The identity check catches the sub
+            // being torn down/replaced; the outer isActive() catches
+            // connection supersession, which bumps the generation while the
+            // SAME subscription key and object survive in the map — identity
+            // alone stays true and a stale pass could complete and clear the
+            // floor the superseding connection needs.
+            isActive: () =>
+              isActive() && subscriptions.get(subId) === subscription,
+            requestHistory,
+          });
+          // A stale-connection abort is NOT completion: the superseding
+          // connection shares this subscription object and still needs the
+          // pinned floor for its own replay. Only a genuinely completed
+          // window may release it.
+          if (completed) subscription.pendingReplaySince = undefined;
+          return;
+        } catch (error) {
+          console.warn(
+            `[reconnect replay] history backfill attempt ${attempt}/${PAGE_REPLAY_MAX_ATTEMPTS} failed for ${subId}:`,
+            error,
+          );
+          if (attempt === PAGE_REPLAY_MAX_ATTEMPTS) return;
+          // The failed REQ's CLOSED handler arms the rate-limit gate before
+          // rejecting; wait for it (no-op when the failure wasn't back-pressure)
+          // and re-check that this replay's connection is still current.
+          if (isRateLimited()) await waitForRateLimit();
+          if (subscriptions.get(subId) !== subscription || !isActive()) return;
+        }
+      }
     },
   );
 }

desktop/src/testing/e2eBridge.ts

@@ -1211,6 +1211,8 @@ declare global {
     ) => void;
     /** Inject CLOSED into every active mock live subscription. */
     __BUZZ_E2E_CLOSE_LIVE_SUBSCRIPTIONS__?: (reason: string) => number;
+    /** Queue CLOSED responses for channel history REQs. */
+    __BUZZ_E2E_QUEUE_CHANNEL_HISTORY_CLOSES__?: (reasons: string[]) => void;
     __BUZZ_E2E_SET_STALL_WEBSOCKET_SENDS__?: (stall: boolean) => void;
     __BUZZ_E2E_DISCONNECT_MOCK_WEBSOCKETS__?: () => number;
     __BUZZ_E2E_RESTART_MOCK_WEBSOCKETS__?: () => number;
@@ -2967,6 +2969,7 @@ const mockPersonaEvents: RelayEvent[] = [];
 let mockRelayMembers: RawRelayMember[] = [];
 const mockSockets = new Map<number, MockSocket>();
 const mockAuthResponses: Array<{ success: boolean; message: string }> = [];
+const mockChannelHistoryCloses: string[] = [];
 let mockWebsocketUnavailable = false;
 const relayWebsocketConnectAttemptStarts: number[] = [];
 let mockWebsocketSendMutexWedged = false;
@@ -9691,6 +9694,13 @@ function sendToMockSocket(args: {
     }
 
     const channelId = filter["#h"]?.[0];
+    if (channelId && subId.startsWith("history-")) {
+      const closeReason = mockChannelHistoryCloses.shift();
+      if (closeReason) {
+        sendWsText(socket.handler, ["CLOSED", subId, closeReason]);
+        return;
+      }
+    }
     if (!channelId) {
       // Aux-backfill filters (reactions/deletions) are `#e`-keyed with no
       // channel tag — serve them across all channel stores like the relay.
@@ -9945,6 +9955,7 @@ export function maybeInstallE2eTauriMocks() {
   mockClosedChannelLiveSubscription = false;
   mockWebsocketUnavailable = false;
   mockAuthResponses.length = 0;
+  mockChannelHistoryCloses.length = 0;
   relayWebsocketConnectAttemptStarts.length = 0;
   deferredSendMessageLiveEchoes.length = 0;
   mockGlobalAgentConfig = config.mock?.globalAgentConfig
@@ -10188,6 +10199,9 @@ export function maybeInstallE2eTauriMocks() {
   window.__BUZZ_E2E_QUEUE_AUTH_RESPONSES__ = (responses) => {
     mockAuthResponses.push(...responses);
   };
+  window.__BUZZ_E2E_QUEUE_CHANNEL_HISTORY_CLOSES__ = (reasons) => {
+    mockChannelHistoryCloses.push(...reasons);
+  };
   window.__BUZZ_E2E_CLOSE_LIVE_SUBSCRIPTIONS__ = (reason) => {
     let closed = 0;
     for (const socket of mockSockets.values()) {

desktop/tests/e2e/relay-reconnect.spec.ts

@@ -134,6 +134,19 @@ async function closeLiveSubscriptions(
   expect(closed).toBeGreaterThan(0);
 }
 
+async function queueChannelHistoryCloses(
+  page: import("@playwright/test").Page,
+  reasons: string[],
+) {
+  await page.evaluate((queued) => {
+    const queue = window.__BUZZ_E2E_QUEUE_CHANNEL_HISTORY_CLOSES__;
+    if (!queue) {
+      throw new Error("E2E channel history CLOSED seam is not installed.");
+    }
+    queue(queued);
+  }, reasons);
+}
+
 async function driveConnectionDegraded(
   page: import("@playwright/test").Page,
   state: "connected" | "reconnecting" | "stalled" | "disconnected",
@@ -253,6 +266,61 @@ test("authenticated reconnect reports connected while replay is rate-limited", a
   await expect(page.getByTestId("sidebar-relay-unreachable")).toHaveCount(0);
 });
 
+test("rate-limited reconnect backfill does not tear down the authenticated socket", async ({
+  page,
+}) => {
+  await page.goto("/");
+  await page.getByTestId("channel-general").click();
+  await expect(page.getByTestId("chat-title")).toHaveText("general");
+
+  // Give the channel live subscription a replay cursor. On reconnect this
+  // causes a paged channel-history REQ in addition to restoring the live REQ.
+  await emitMockMessages(page, [
+    {
+      content: `replay cursor ${Date.now()}`,
+      createdAt: Math.floor(Date.now() / 1_000),
+    },
+  ]);
+  const attemptsBeforeReconnect = (await getMockWebsocketConnectAttempts(page))
+    .length;
+
+  // Inject back-pressure from the history REQ itself. This differs from a
+  // pre-armed gate and from CLOSED on the live subscription: AUTH has already
+  // succeeded and the socket is healthy when replay backfill is rejected.
+  await queueChannelHistoryCloses(page, [
+    "rate-limited: quota exceeded; retry in 1s",
+  ]);
+  await disconnectMockWebsockets(page);
+
+  await expect
+    .poll(
+      async () =>
+        (await getMockWebsocketConnectAttempts(page)).length -
+        attemptsBeforeReconnect,
+      { timeout: 3_000 },
+    )
+    .toBe(1);
+  await expect
+    .poll(() =>
+      page.evaluate(() => window.__BUZZ_E2E_GET_RELAY_CONNECTION_STATE__?.()),
+    )
+    .toBe("connected");
+
+  // Hold through the rate-limit hint plus the next base-backoff window. The
+  // authenticated socket must remain the only reconnect attempt, rather than
+  // flashing connected and redialing after replay rejects.
+  await page.waitForTimeout(2_500);
+  expect(
+    (await getMockWebsocketConnectAttempts(page)).length -
+      attemptsBeforeReconnect,
+  ).toBe(1);
+  expect(
+    await page.evaluate(() =>
+      window.__BUZZ_E2E_GET_RELAY_CONNECTION_STATE__?.(),
+    ),
+  ).toBe("connected");
+});
+
 test("service restart close resets accumulated backoff", async ({ page }) => {
   await installMockBridge(page, {
     websocketConnectErrors: ["down 1", "down 2", "down 3"],