diff --git a/CHANGELOG.md b/CHANGELOG.md
index 2baaf7ab8..16228054f 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -47,6 +47,24 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
   The clear list is now covered by a test that derives the expected set from `docker-compose.yml`
   itself, so a forward added without its clear entry fails in CI rather than shipping inert.
 
+- **Restarting from Dashboard > Infrastructure no longer reloads the page into an error.** The restart
+  modal polls the server and reloads once it answers, but it polled `GET /api/infra/health` — a plain
+  ping with no knowledge of shutdown. That endpoint keeps answering `200` for the entire drain window
+  and the engine teardown that follows, so the very first poll, three seconds in, was answered by the
+  process that had just been asked to exit. The modal declared success and reloaded two seconds later,
+  by which time the old process was gone and the new one had not yet bound its port — leaving the
+  operator on a 502/503 from their reverse proxy until they refreshed by hand several times.
+
+  The poll now targets `GET /api/health/ready`, which reports `503` the moment draining starts and
+  keeps reporting it until both databases answer. The page therefore reloads only once the new process
+  is genuinely serving, and it waits for the databases too rather than only for the port to open.
+
+  With the poll now waiting for real readiness, its deadline started to matter for the first time.
+  It was a fixed 60 attempts at one second, while `POST /api/infra/restart` estimates up to 63 seconds
+  for a restart that brings up PostgreSQL, Redis and MinIO together — so a full-profile restart could
+  have reported failure while the stack was still coming up correctly. The deadline is now derived
+  from the estimate the server already returns, with the old 60 as a floor. (#1019)
+
 ## [0.12.2] - 2026-08-01
 
 ### Changed
diff --git a/dashboard/src/pages/Infrastructure.tsx b/dashboard/src/pages/Infrastructure.tsx
index d59a72847..1c19af7a6 100644
--- a/dashboard/src/pages/Infrastructure.tsx
+++ b/dashboard/src/pages/Infrastructure.tsx
@@ -15,6 +15,7 @@ import {
 } from 'lucide-react';
 import { infraApi, API_BASE_URL } from '../services/api';
 import { copyToClipboard } from '../utils/clipboard';
+import { restartPollAttempts } from '../utils/restartPoll';
 import { useDocumentTitle } from '../hooks/useDocumentTitle';
 import { useInfraStatusQuery, useInfraConfigQuery, useEnginesQuery, useCurrentEngineQuery } from '../hooks/queries';
 import { PageHeader } from '../components/PageHeader';
@@ -451,8 +452,12 @@ export function Infrastructure() {
 
     const profilesToRemove = previousProfiles.filter(p => !pendingProfiles.includes(p));
 
+    // Kept outside the try: the poll deadline is derived from it, and the restart call is expected to
+    // fail sometimes (the server may go down before it answers).
+    let estimatedTime: number | undefined;
     try {
       const response = await infraApi.restart(pendingProfiles, profilesToRemove);
+      estimatedTime = response.estimatedTime;
       if (response.estimatedTime) setRestartCountdown(response.estimatedTime);
     } catch {
       // Expected — server shutting down
@@ -477,12 +482,12 @@ export function Infrastructure() {
       });
     }, 1000);
 
-    checkServerHealth(stopCountdown);
+    checkServerHealth(stopCountdown, estimatedTime);
   };
 
-  const checkServerHealth = async (stopCountdown?: () => void) => {
+  const checkServerHealth = async (stopCountdown?: () => void, estimatedTime?: number) => {
     let attempts = 0;
-    const maxAttempts = 60;
+    const maxAttempts = restartPollAttempts(estimatedTime);
 
     const check = async () => {
       try {
diff --git a/dashboard/src/services/api.ts b/dashboard/src/services/api.ts
index eb86b3317..37ebf95a1 100644
--- a/dashboard/src/services/api.ts
+++ b/dashboard/src/services/api.ts
@@ -1010,7 +1010,11 @@ export const infraApi = {
       method: 'POST',
       body: JSON.stringify({ profiles: profiles || [], profilesToRemove: profilesToRemove || [] }),
     }),
-  healthCheck: () => request<{ status: string; timestamp: string }>('/infra/health'),
+  // Readiness, not the plain /infra/health ping. The restart poll must not be able to latch onto the
+  // process it just asked to shut down: /infra/health answers 200 for the whole drain and teardown,
+  // while /health/ready reports 503 as soon as draining starts and stays 503 until both databases
+  // answer. Public (no API key), like /infra/health.
+  healthCheck: () => request<{ status: 'ok' | 'error'; details: Record<string, { status: string }> }>('/health/ready'),
   // Data migration: export all Data-DB tables (call while still on the OLD database, before switching),
   // then import after the switch + restart. Used by the DB-switch migration guard so data isn't lost.
   exportData: () =>
diff --git a/dashboard/src/utils/restartPoll.test.ts b/dashboard/src/utils/restartPoll.test.ts
new file mode 100644
index 000000000..f6d45e872
--- /dev/null
+++ b/dashboard/src/utils/restartPoll.test.ts
@@ -0,0 +1,56 @@
+import { test } from 'node:test';
+import assert from 'node:assert/strict';
+import { readFileSync } from 'node:fs';
+import { restartPollAttempts } from './restartPoll.ts';
+
+// The poll runs once a second, so an attempt count IS a deadline in seconds. It used to be a fixed
+// 60, which never mattered: the poll hit the still-draining old process and "succeeded" within three
+// seconds. Once it waits for real readiness, that 60 becomes the thing that decides whether the
+// operator sees the dashboard or an error — and POST /infra/restart estimates up to 63s for a restart
+// that brings up postgres + redis + minio. The ceiling must not sit below the server's own estimate.
+
+test('no estimate: keeps the historical 60-attempt deadline', () => {
+  assert.equal(restartPollAttempts(undefined), 60);
+});
+
+test('an estimate the floor already covers does not shorten the deadline', () => {
+  assert.equal(restartPollAttempts(15), 60);
+});
+
+test('a full-profile restart gets more room than the server estimated', () => {
+  // 15 base + 20 postgres + 13 redis + 15 minio — the worst case infra.controller computes.
+  assert.equal(restartPollAttempts(63), 126);
+});
+
+test('a fractional estimate rounds up rather than truncating below it', () => {
+  assert.equal(restartPollAttempts(40.5), 81);
+});
+
+test('a bogus estimate cannot hang the modal or shorten the deadline', () => {
+  for (const bogus of [0, -1, NaN, Infinity]) {
+    assert.equal(restartPollAttempts(bogus), 60, `estimate ${bogus}`);
+  }
+  assert.equal(restartPollAttempts(99999), 300);
+});
+
+// The two assertions below guard the WIRING, which is where the bug actually lived. A correct helper
+// that nothing calls, or a correct deadline pointed at an endpoint that lies during shutdown, both
+// pass every assertion above while reproducing #1019 in full. The failure is invisible until an
+// operator restarts a production stack, so it has to be caught here.
+const read = (path: string): string => readFileSync(new URL(path, import.meta.url), 'utf8');
+
+test('the restart poll targets readiness, not a ping that survives the drain', () => {
+  const api = read('../services/api.ts');
+  const healthCheck = api.split('\n').find(line => line.trimStart().startsWith('healthCheck: () =>'));
+  assert.ok(healthCheck, 'infraApi.healthCheck not found — update this guard, do not delete it');
+  assert.match(
+    healthCheck,
+    /'\/health\/ready'/,
+    'the poll must hit /health/ready: /infra/health answers 200 throughout the drain, so the poll ' +
+      'latches onto the dying process and reloads the page into a gap (#1019)',
+  );
+});
+
+test('the poll deadline comes from the helper rather than a bare constant', () => {
+  assert.match(read('../pages/Infrastructure.tsx'), /const maxAttempts = restartPollAttempts\(/);
+});
diff --git a/dashboard/src/utils/restartPoll.ts b/dashboard/src/utils/restartPoll.ts
new file mode 100644
index 000000000..b99d6e455
--- /dev/null
+++ b/dashboard/src/utils/restartPoll.ts
@@ -0,0 +1,25 @@
+/**
+ * How many one-second readiness polls to allow before a restart is declared failed.
+ *
+ * The poll used to latch onto the OLD process — `/infra/health` answers 200 throughout the drain and
+ * the engine teardown that follows — so it "succeeded" in about three seconds and the fixed 60-attempt
+ * deadline was never approached. Polling readiness instead means the poll genuinely waits for the new
+ * process, and the deadline becomes the thing that decides whether the operator sees the dashboard or
+ * an error screen.
+ *
+ * `POST /infra/restart` returns its own `estimatedTime`, which reaches 63s when a restart brings up
+ * postgres, redis and minio together (and more with services to remove). A deadline below the server's
+ * own estimate would report failure while the stack is still coming up correctly, so the estimate sets
+ * the floor and is doubled for margin: the estimate is optimistic, and a container recreate plus
+ * migrations can overrun it.
+ *
+ * The value comes from a server response, so it is clamped at both ends — a missing or absurd estimate
+ * must not shorten the deadline below the historical 60, nor leave the modal waiting indefinitely.
+ */
+const MIN_ATTEMPTS = 60;
+const MAX_ATTEMPTS = 300;
+
+export function restartPollAttempts(estimatedSeconds?: number): number {
+  if (!Number.isFinite(estimatedSeconds) || (estimatedSeconds as number) <= 0) return MIN_ATTEMPTS;
+  return Math.min(MAX_ATTEMPTS, Math.max(MIN_ATTEMPTS, Math.ceil((estimatedSeconds as number) * 2)));
+}