diff --git a/backend/CLAUDE.md b/backend/CLAUDE.md
index d081ca1f..71e8a248 100644
--- a/backend/CLAUDE.md
+++ b/backend/CLAUDE.md
@@ -53,8 +53,17 @@ thin: it proxies OpenAI requests with the server-held API key and writes study l
   static root doesn't exist. Cache rules: `no-store` for `.html`/`manifest.xml`
   (they reference content-hashed bundles by name), `immutable` for hashed assets.
 - **Telemetry** (`src/posthog.ts`): optional PostHog error capture; a no-op when
-  `POSTHOG_PROJECT_TOKEN` is unset. `deletePosthogPerson` (used by both erasure
-  paths) hits the PostHog *management* API and needs `POSTHOG_PERSONAL_API_KEY` +
+  `POSTHOG_PROJECT_TOKEN` is unset. **Nothing here may be awaited on the request
+  path.** `posthogMiddleware` used to `await posthog.flush()` after every request;
+  `flush()` rejects on any non-2xx, so when the ingestion host started returning 526
+  every request 500'd (and, because flushes serialize, hung) — health probe included,
+  which failed a deploy. Ingestion is now left to posthog-node's background batching,
+  with retries/timeouts capped so a dead host costs seconds rather than ~49s; the tail
+  is drained by `shutdownPosthog()` on SIGTERM. `/api/ping` is deliberately not
+  captured. `POSTHOG_HOST` defaults to PostHog US **directly** — the
+  `e.thoughtful-ai.com` reverse proxy is for the browser SDK (ad blockers) and only
+  adds a failure mode server-side. `deletePosthogPerson` (used by both erasure paths)
+  hits the PostHog *management* API and needs `POSTHOG_PERSONAL_API_KEY` +
   `POSTHOG_PROJECT_ID`; without them it warns and no-ops.
 - **Tool launcher** (`src/toolGrants.ts`): the handoff grant flow for launching
   external writing tools from the sidebar (see `docs/tool-launcher-plan.md`, Phase 1).
diff --git a/backend/README.md b/backend/README.md
index aa1af957..f35aa15a 100644
--- a/backend/README.md
+++ b/backend/README.md
@@ -42,6 +42,7 @@ Listens on `5000` in Docker (`PORT=5000`).
 `OPENAI_API_KEY`, `OPENAI_DEMO_API_KEY` (pays for sessionless/demo requests),
 `LOG_SECRET` (gates the log-viewer endpoints + `/api/usage_summary`), `DATA_DIR` (root
 for `app.db` + `logs/`), `PORT` (default 8000), `DEBUG`, `POSTHOG_PROJECT_TOKEN`,
-`POSTHOG_HOST`, `LOG_DIR` (overrides just the logs subdir). Auth (Better Auth) adds
+`POSTHOG_HOST` (defaults to `https://us.i.posthog.com`; the server does not go through
+the browser-SDK reverse proxy), `LOG_DIR` (overrides just the logs subdir). Auth (Better Auth) adds
 `BETTER_AUTH_ENABLED`, `BETTER_AUTH_SECRET`, `GOOGLE_CLIENT_ID`, `GOOGLE_CLIENT_SECRET`,
 and related vars — see [CLAUDE.md](CLAUDE.md) for the full list.
diff --git a/backend/src/__tests__/posthog.test.ts b/backend/src/__tests__/posthog.test.ts
new file mode 100644
index 00000000..4613d0ac
--- /dev/null
+++ b/backend/src/__tests__/posthog.test.ts
@@ -0,0 +1,135 @@
+import { createServer, type Server } from 'node:http';
+import type { AddressInfo } from 'node:net';
+import { afterAll, beforeAll, beforeEach, describe, expect, it, vi } from 'vitest';
+
+/**
+ * Regression test for the outage where a failing PostHog ingestion host took the
+ * whole backend down: posthogMiddleware used to `await posthog.flush()` after every
+ * request, and flush() rejects on any non-2xx from the host. A transient 526 from
+ * the analytics reverse proxy therefore turned every request — including the k8s
+ * health probe — into a 500, and the deploy never went ready.
+ *
+ * Telemetry must stay strictly off the request's critical path: it may not decide a
+ * response's status, and it may not add to its latency.
+ *
+ * The bad host here is a real local HTTP server rather than a mocked `fetch`:
+ * posthog-node resolves the global fetch once when *it* is imported, which is
+ * outside Vitest's module registry, so a `vi.stubGlobal('fetch', …)` in this file
+ * may never be the one it calls.
+ */
+
+let server: Server;
+let hostUrl: string;
+/** Swapped per test to make the ingestion host fail in a particular way. */
+let handle: (respond: (status: number) => void) => void;
+/**
+ * Bodies of the batches the fake PostHog received. Assertions match on the event
+ * names inside them rather than counting requests: a client from an earlier test
+ * can still have a chained flush in flight, so only the contents identify who sent
+ * what.
+ */
+let batches: string[] = [];
+
+beforeAll(async () => {
+	server = createServer((req, res) => {
+		let body = '';
+		req.on('data', (chunk) => {
+			body += chunk;
+		});
+		req.on('end', () => {
+			batches.push(body);
+			handle((status) => res.writeHead(status).end('error'));
+		});
+	});
+	await new Promise<void>((resolve) => server.listen(0, '127.0.0.1', resolve));
+	hostUrl = `http://127.0.0.1:${(server.address() as AddressInfo).port}`;
+});
+
+afterAll(async () => {
+	await new Promise((resolve) => server.close(resolve));
+});
+
+// posthog.ts builds its client from the environment at module load, so each test
+// stubs the env, resets the module registry, and re-imports.
+async function loadApp(route: string) {
+	vi.stubEnv('POSTHOG_PROJECT_TOKEN', 'phc_test_token');
+	vi.stubEnv('DISABLE_POSTHOG', '');
+	vi.stubEnv('POSTHOG_HOST', hostUrl);
+	vi.resetModules();
+
+	const { posthogMiddleware } = await import('../posthog.js');
+	const { Hono } = await import('hono');
+	const app = new Hono();
+	app.use('*', posthogMiddleware);
+	app.get(route, (c) => c.json({ ok: true }));
+	return app;
+}
+
+// posthog-node batches: it only sends once flushAt (20) events are queued, so a run
+// of requests is what provokes an ingestion attempt at all.
+const OVER_FLUSH_AT = 25;
+
+beforeEach(() => {
+	batches = [];
+	handle = (respond) => respond(526); // what the reverse proxy was returning
+});
+
+describe('posthogMiddleware with a failing PostHog host', () => {
+	it('serves requests normally while ingestion 526s', async () => {
+		const app = await loadApp('/api/thing');
+
+		for (let i = 0; i < OVER_FLUSH_AT; i++) {
+			const res = await app.request('/api/thing');
+			expect(res.status).toBe(200);
+			expect(await res.json()).toEqual({ ok: true });
+		}
+
+		// The failing ingestion really was attempted — otherwise this test would
+		// pass just as well with telemetry silently doing nothing.
+		await vi.waitFor(() =>
+			expect(batches.some((b) => b.includes('GET /api/thing'))).toBe(true),
+		);
+	});
+
+	it('does not make requests wait on a hung ingestion host', async () => {
+		handle = () => {
+			/* accept the connection, never respond */
+		};
+		const app = await loadApp('/api/thing');
+
+		const started = Date.now();
+		for (let i = 0; i < OVER_FLUSH_AT; i++) {
+			expect((await app.request('/api/thing')).status).toBe(200);
+		}
+		// The client's own requestTimeout is 3s; the old await-flush middleware
+		// would have blocked for at least that long before failing the request.
+		expect(Date.now() - started).toBeLessThan(1000);
+	});
+
+	it('does not capture health-probe requests', async () => {
+		const app = await loadApp('/api/ping');
+
+		for (let i = 0; i < OVER_FLUSH_AT; i++) {
+			expect((await app.request('/api/ping')).status).toBe(200);
+		}
+		// Well past flushAt with nothing queued, so no probe ever reaches PostHog.
+		await new Promise((r) => setTimeout(r, 100));
+		expect(batches.filter((b) => b.includes('/api/ping'))).toEqual([]);
+	});
+});
+
+describe('captureException with a failing PostHog host', () => {
+	it('resolves promptly instead of throwing or hanging', async () => {
+		handle = () => {
+			/* accept the connection, never respond */
+		};
+		vi.stubEnv('POSTHOG_PROJECT_TOKEN', 'phc_test_token');
+		vi.stubEnv('POSTHOG_HOST', hostUrl);
+		vi.resetModules();
+		const { captureException } = await import('../posthog.js');
+
+		const started = Date.now();
+		await captureException(new Error('boom'));
+		expect(Date.now() - started).toBeLessThan(1000);
+	});
+});
diff --git a/backend/src/posthog.ts b/backend/src/posthog.ts
index 72818dfd..00dbcd4e 100644
--- a/backend/src/posthog.ts
+++ b/backend/src/posthog.ts
@@ -2,8 +2,11 @@ import { createMiddleware } from 'hono/factory';
 import { PostHog } from 'posthog-node';
 
 const token = (process.env.POSTHOG_PROJECT_TOKEN ?? '').trim() || "placeholder-token";
+// Defaults to PostHog US directly, NOT the e.thoughtful-ai.com reverse proxy: the
+// server has no ad-blocker to evade (that's only worth it for the browser SDK), so
+// the proxy added a hop that could — and did — fail on its own.
 // Strip any trailing slash so building management-API URLs (deletePosthogPerson)
-// can't produce a double slash — the shipped POSTHOG_HOST default ends in '/', and
+// can't produce a double slash — an explicitly-set POSTHOG_HOST may end in '/', and
 // many reverse proxies 404 on `//api/...`.
 const host =
 	((process.env.POSTHOG_HOST ?? '').trim() || 'https://us.i.posthog.com').replace(
@@ -13,15 +16,40 @@ const host =
 
 const shouldDisablePosthog = token === "placeholder-token" || process.env.DISABLE_POSTHOG === '1';
 
-const posthog = new PostHog(token, { host, disabled: shouldDisablePosthog });
+const posthog = new PostHog(token, {
+	host,
+	disabled: shouldDisablePosthog,
+	// Telemetry must never be able to hold a request (or the whole process) open.
+	// posthog-node's defaults are 10s per attempt x 4 attempts with 3s between them
+	// — ~49s of hanging per flush when the ingestion host is down. Ingestion is
+	// fire-and-forget analytics, so cap it at a few seconds and let the batch drop.
+	requestTimeout: 3000,
+	fetchRetryCount: 1,
+	fetchRetryDelay: 1000,
+});
 
 export const posthogMiddleware = createMiddleware(async (c, next) => {
-	posthog.capture({
-		distinctId: 'server',
-		event: `${c.req.method} ${c.req.path}`,
-	});
+	// Health probes are not product analytics: they'd add a queue entry (and a
+	// distinct event name) every few seconds forever.
+	if (c.req.path !== '/api/ping') {
+		try {
+			posthog.capture({
+				distinctId: 'server',
+				event: `${c.req.method} ${c.req.path}`,
+			});
+		} catch {
+			// Never let telemetry break the request path.
+		}
+	}
 	await next();
-	await posthog.flush();
+	// Deliberately NOT `await posthog.flush()`. That made every request wait on —
+	// and 500 on — a PostHog ingestion failure: flush() rejects on any non-2xx,
+	// the rejection propagated out of this middleware into app.onError, and
+	// because flush() serializes behind the previous flush the latency compounded
+	// across concurrent requests. A transient 526 from the analytics host thereby
+	// took the whole backend down, health probe included. posthog-node already
+	// flushes in the background (batches of 20, or every 10s) and swallows its own
+	// errors there; shutdownPosthog() drains the tail on SIGTERM.
 });
 
 export async function captureException(
@@ -31,7 +59,9 @@ export async function captureException(
 	try {
 		const err = error instanceof Error ? error : new Error(String(error));
 		posthog.captureException(err, undefined, properties);
-		await posthog.flush();
+		// No flush() here either — see posthogMiddleware. The background flush
+		// timer sends this within ~10s without putting the ingestion host on the
+		// critical path of an error response.
 	} catch {
 		// Never let error tracking break the request path.
 	}
diff --git a/docker-compose-prod.yml b/docker-compose-prod.yml
index 4532ceb3..ead5ddd5 100644
--- a/docker-compose-prod.yml
+++ b/docker-compose-prod.yml
@@ -11,7 +11,9 @@ services:
       # on OpenAI's side. Unset => sessionless requests are refused. See openaiProxy.ts.
       - OPENAI_DEMO_API_KEY=${OPENAI_DEMO_API_KEY:-}
       - POSTHOG_PROJECT_TOKEN=${POSTHOG_PROJECT_TOKEN:-}
-      - POSTHOG_HOST=${POSTHOG_HOST:-https://e.thoughtful-ai.com/}
+      # Straight to PostHog US — see docker-compose.yml for why the backend does
+      # not go through the e.thoughtful-ai.com reverse proxy.
+      - POSTHOG_HOST=${POSTHOG_HOST:-https://us.i.posthog.com}
       # Management API (personal key + project id) — required for "delete my data"
       # to purge a user's PostHog person/events. Unset => deletion no-ops with a
       # warning (see deletePosthogPerson). Verified post-deploy in issue #517.
diff --git a/docker-compose.yml b/docker-compose.yml
index 0c7a248e..7c6a4896 100644
--- a/docker-compose.yml
+++ b/docker-compose.yml
@@ -19,7 +19,11 @@ services:
       - OPENAI_DEMO_API_KEY=${OPENAI_DEMO_API_KEY:-}
       - LOG_SECRET=${LOG_SECRET}
       - POSTHOG_PROJECT_TOKEN=${POSTHOG_PROJECT_TOKEN:-}
-      - POSTHOG_HOST=${POSTHOG_HOST:-https://e.thoughtful-ai.com/}
+      # Server-side ingestion goes straight to PostHog US, not the
+      # e.thoughtful-ai.com reverse proxy. The proxy exists to get past browser ad
+      # blockers, which is irrelevant here, and it added a hop whose outage (a
+      # Cloudflare 526) used to take the backend down with it.
+      - POSTHOG_HOST=${POSTHOG_HOST:-https://us.i.posthog.com}
       # Management API (personal key + project id) — required for "delete my data"
       # to purge a user's PostHog person/events. Unset => deletion no-ops with a
       # warning (see deletePosthogPerson).