diff --git a/.github/actions/test/action.yml b/.github/actions/test/action.yml
index aa85a05..206bbf3 100644
--- a/.github/actions/test/action.yml
+++ b/.github/actions/test/action.yml
@@ -21,7 +21,11 @@ runs:
       run: test -s dist/index.html
     # The point of this site is that it ships no JavaScript. Nothing else
     # would notice an integration being added back, and by then the page has
-    # a runtime it did not need.
+    # a runtime it did not need. Two gates, because each sees what the other
+    # cannot: the find catches a script file nothing references, and
+    # check_scripts.py catches markup -- any <script src>, and any inline
+    # <script> whose type is not application/ld+json (the structured metadata
+    # in the layout is data, and stays allowed).
     - name: No JavaScript was shipped
       shell: bash
       run: |
@@ -30,6 +34,19 @@ runs:
           find dist -name '*.js'
           exit 1
         fi
+    - name: No executable scripts in the markup
+      shell: bash
+      run: python3 tools/check_scripts.py dist
+    # A page that builds can still be broken in a browser. This loads the
+    # built site in headless Chrome -- preinstalled on the runner, so there is
+    # no browser download -- and fails on console errors, uncaught exceptions,
+    # failed or >=400 requests, missing landmarks, unlabelled links or SVGs,
+    # horizontal overflow, and broken links. The exact parameters are written
+    # in tools/browser-check.mjs: viewports 1280x800 and 390x844, external
+    # link timeout 10000ms, zero tolerance for errors.
+    - name: The page works in a browser
+      shell: bash
+      run: node tools/browser-check.mjs dist
     # The frames are the page. They are pasted-in captures, and a build that
     # dropped them would still look like a working site until someone read it.
     - name: The captures made it onto the page
diff --git a/.mise.toml b/.mise.toml
index ac33547..ab482cb 100644
--- a/.mise.toml
+++ b/.mise.toml
@@ -1,2 +1,5 @@
 [tools]
 bun = "1.3"
+# The browser smoke check (tools/browser-check.mjs) runs on node, not bun --
+# playwright-core does not support the bun runtime.
+node = "22"
diff --git a/README.md b/README.md
index 9b4bef3..67f4188 100644
--- a/README.md
+++ b/README.md
@@ -7,8 +7,11 @@ the four pixel styles look like, and how to write a game for it.
 There is no authentication, no application state, no backend and no runtime
 environment configuration — the arcade and its registry are the application,
 and this is only the front door. The page ships no JavaScript, which is why
-there are no integrations in `astro.config.mjs` and why CI fails on a
-`<script src>` in the build output.
+there are no integrations in `astro.config.mjs` and why CI rejects any
+executable `<script>` in the build output — external or inline, with only the
+`application/ld+json` metadata allowed — and then loads the built page in
+headless Chrome (1280×800 and 390×844, 10s link timeout, zero tolerated
+console or resource errors; the full list is in `tools/browser-check.mjs`).
 
 ## Commands
 
diff --git a/bun.lock b/bun.lock
index dbf608c..acc611f 100644
--- a/bun.lock
+++ b/bun.lock
@@ -8,6 +8,9 @@
         "@primer/octicons": "^19.29.2",
         "astro": "^5.18.1",
       },
+      "devDependencies": {
+        "playwright-core": "^1.62.1",
+      },
     },
   },
   "packages": {
@@ -541,6 +544,8 @@
 
     "picomatch": ["picomatch@4.0.5", "", {}, "sha512-RvwwcruNjI1ncT5xRakeyS9Lf8lcItv34KD+aif+VH9kduAyfYBipGh12274xtenIPZ119/R9BdTBa8gAwSh0A=="],
 
+    "playwright-core": ["playwright-core@1.62.1", "", { "bin": { "playwright-core": "cli.js" } }, "sha512-wPYSwEBJY9GHraISXqyqtx0na0LpO3XEX7jNDhntbex7tzUS7kLnZsOlFruFJB4Hi/rhDMjXGqHewDZ68nYZVw=="],
+
     "postcss": ["postcss@8.5.25", "", { "dependencies": { "nanoid": "^3.3.16", "picocolors": "^1.1.1", "source-map-js": "^1.2.1" } }, "sha512-DTPx3RWSSnWyzLxQnlH0rJP+EW5ekl16ZU4/psbIhA0e53kJfdgaN5vKM+xP7yJtXVu+nfdVFmlgFDEKAe4Pyw=="],
 
     "prismjs": ["prismjs@1.30.0", "", {}, "sha512-DEvV2ZF2r2/63V+tK8hQvrR2ZGn10srHbXviTlcv7Kpzw8jWiNTqbVgjO3IY8RxrrOUF8VPMQQFysYYYv0YZxw=="],
diff --git a/package.json b/package.json
index af2e8d1..a5a2785 100644
--- a/package.json
+++ b/package.json
@@ -15,5 +15,8 @@
   "dependencies": {
     "@primer/octicons": "^19.29.2",
     "astro": "^5.18.1"
+  },
+  "devDependencies": {
+    "playwright-core": "^1.62.1"
   }
 }
diff --git a/tools/browser-check.mjs b/tools/browser-check.mjs
new file mode 100644
index 0000000..1084ded
--- /dev/null
+++ b/tools/browser-check.mjs
@@ -0,0 +1,222 @@
+// Browser smoke check for the built site. Serves dist/ over a throwaway
+// local HTTP server, loads the page in headless Chrome at two viewports, and
+// fails on anything a visitor would notice:
+//
+//   - console errors, uncaught exceptions, failed requests, HTTP >= 400
+//     responses (tolerance: zero, at either viewport)
+//   - missing document landmarks: exactly one <h1>, a <header>, <main> and
+//     <footer>, html[lang], a title and a meta description
+//   - inaccessible content: every link and every unhidden <svg>/<img> has a
+//     non-empty COMPUTED accessible name, read from Chrome's accessibility
+//     tree (CDP Accessibility.getFullAXTree) rather than from raw markup --
+//     an empty aria-label, an aria-labelledby pointing nowhere, or text
+//     hidden with display:none all compute to an empty name, and raw
+//     attribute checks wave them through. Decorative SVGs carry aria-hidden
+//     and never reach the tree. There are no raster images on the page, so
+//     no alt rule binds. The terminal frames are text by design -- selectable
+//     and readable as text, which is the point of them -- so no label
+//     assertion binds to them either; a previous version asserted aria-labels
+//     on their plain <div> wrappers, which assistive technology ignores on
+//     role-less elements, and the assertion was removed rather than kept as
+//     theatre.
+//   - horizontal overflow: no page-level horizontal scrollbar, and no
+//     top-level section extending past the viewport edge
+//   - broken links: every internal href returns 200 from the built site, and
+//     every external href answers 200 within LINK_TIMEOUT_MS
+//
+// Parameters, deliberately written down rather than implied:
+//
+//   viewports      1280x800 (desktop) and 390x844 (mobile, touch + isMobile)
+//   link timeout   10000ms per external link, one attempt, redirects followed
+//   accessibility  the landmark assertions above, plus a non-empty computed
+//                  accessible name for every link and unhidden SVG/image,
+//                  taken from the browser accessibility tree at each viewport
+//
+// Usage: node tools/browser-check.mjs dist
+import { chromium } from 'playwright-core';
+import http from 'node:http';
+import { readFile } from 'node:fs/promises';
+import path from 'node:path';
+
+const VIEWPORTS = [
+	{ name: 'desktop', width: 1280, height: 800 },
+	{ name: 'mobile', width: 390, height: 844, isMobile: true, hasTouch: true },
+];
+const LINK_TIMEOUT_MS = 10_000;
+// The top-level layout containers: if any of these is wider than the
+// viewport, the layout is broken no matter what clips it.
+const SECTIONS = ['.site-header', '.hero', '.games', '.pixels', '.principles', '.arcade', '.sdk', '.closing', 'footer'];
+
+const TYPES = {
+	'.html': 'text/html; charset=utf-8',
+	'.css': 'text/css; charset=utf-8',
+	'.svg': 'image/svg+xml',
+	'.txt': 'text/plain; charset=utf-8',
+	'.xml': 'application/xml; charset=utf-8',
+};
+
+function serve(root) {
+	const server = http.createServer(async (req, res) => {
+		const url = new URL(req.url, 'http://localhost');
+		const file = path.join(root, url.pathname === '/' ? 'index.html' : url.pathname);
+		try {
+			const body = await readFile(file);
+			res.writeHead(200, { 'content-type': TYPES[path.extname(file)] ?? 'application/octet-stream' });
+			res.end(body);
+		} catch {
+			res.writeHead(404);
+			res.end('not found');
+		}
+	});
+	return new Promise((resolve) => {
+		server.listen(0, '127.0.0.1', () => resolve({ server, port: server.address().port }));
+	});
+}
+
+async function launch() {
+	try {
+		// CI runners and most dev machines already have Chrome; using it keeps
+		// the check free of a browser download.
+		return await chromium.launch({ channel: 'chrome' });
+	} catch {
+		// Fall back to a playwright-managed chromium if one is installed.
+		return await chromium.launch();
+	}
+}
+
+const failures = [];
+const fail = (what) => failures.push(what);
+
+async function checkAccessibleNames(context, page, tag) {
+	// The browser's own accessibility tree, not the markup: names here are
+	// what assistive technology computes, after aria-label/aria-labelledby
+	// resolution, display:none, and every other rule of the accName spec.
+	const session = await context.newCDPSession(page);
+	await session.send('Accessibility.enable');
+	const { nodes } = await session.send('Accessibility.getFullAXTree');
+	const live = nodes.filter((n) => !n.ignored);
+	const describe = async (n) => {
+		try {
+			const { node } = await session.send('DOM.describeNode', { backendNodeId: n.backendDOMNodeId });
+			const attrs = [];
+			for (let i = 0; i < node.attributes.length; i += 2) {
+				attrs.push(`${node.attributes[i]}="${node.attributes[i + 1]}"`);
+			}
+			return `<${node.nodeName.toLowerCase()} ${attrs.join(' ')}>`;
+		} catch {
+			return `<${n.role?.value ?? '?'}>`;
+		}
+	};
+	for (const [role, what] of [
+		['link', 'link with an empty computed accessible name'],
+		['image', 'unhidden svg or image with an empty computed accessible name'],
+	]) {
+		const bad = live.filter((n) => n.role?.value === role && !(n.name?.value ?? '').trim());
+		for (const n of bad.slice(0, 3)) fail(`${tag} ${what}: ${await describe(n)}`);
+		if (bad.length > 3) fail(`${tag} ${bad.length - 3} more ${role} node(s) with empty computed names`);
+	}
+}
+
+async function checkPage(page, base, viewport) {
+	const tag = `[${viewport.name} ${viewport.width}x${viewport.height}]`;
+	for (const error of page.consoleErrors) fail(`${tag} console error: ${error}`);
+	for (const error of page.pageErrors) fail(`${tag} uncaught exception: ${error}`);
+	for (const failure of page.requestFailures) fail(`${tag} request failed: ${failure}`);
+	for (const bad of page.badResponses) fail(`${tag} HTTP ${bad}`);
+
+	const doc = await page.evaluate((sections) => {
+		const overflowing = sections
+			.map((sel) => document.querySelector(sel))
+			.filter(Boolean)
+			.map((el) => el.getBoundingClientRect())
+			.filter((r) => r.left < -1 || r.right > window.innerWidth + 1).length;
+		return {
+			lang: document.documentElement.lang,
+			title: document.title,
+			description: document.querySelector('meta[name="description"]')?.content ?? '',
+			h1: document.querySelectorAll('h1').length,
+			header: document.querySelectorAll('header').length,
+			main: document.querySelectorAll('main').length,
+			footer: document.querySelectorAll('footer').length,
+			horizontalScrollbar: document.documentElement.scrollWidth > document.documentElement.clientWidth + 1,
+			overflowingSections: overflowing,
+			hrefs: [...new Set([...document.querySelectorAll('a[href]')].map((a) => a.getAttribute('href')))],
+		};
+	}, SECTIONS);
+
+	if (doc.lang !== 'en') fail(`${tag} html lang is "${doc.lang}", expected "en"`);
+	if (!doc.title) fail(`${tag} no <title>`);
+	if (!doc.description) fail(`${tag} no meta description`);
+	if (doc.h1 !== 1) fail(`${tag} expected exactly one <h1>, found ${doc.h1}`);
+	for (const landmark of ['header', 'main', 'footer']) {
+		if (doc[landmark] < 1) fail(`${tag} no <${landmark}> landmark`);
+	}
+	if (doc.horizontalScrollbar) fail(`${tag} page has a horizontal scrollbar`);
+	if (doc.overflowingSections) fail(`${tag} ${doc.overflowingSections} section(s) extend past the viewport`);
+
+	return doc.hrefs;
+}
+
+async function checkLinks(hrefs, base) {
+	for (const href of hrefs) {
+		const url = href.startsWith('http') ? href : new URL(href, base).href;
+		try {
+			const res = await fetch(url, {
+				redirect: 'follow',
+				signal: AbortSignal.timeout(LINK_TIMEOUT_MS),
+				headers: { 'user-agent': 'termcade-web link check' },
+			});
+			if (res.status !== 200) fail(`link ${href}: HTTP ${res.status}`);
+		} catch (error) {
+			fail(`link ${href}: ${error.message} (timeout ${LINK_TIMEOUT_MS}ms)`);
+		}
+	}
+}
+
+async function main() {
+	const root = process.argv[2];
+	if (!root) {
+		console.error('usage: node tools/browser-check.mjs dist');
+		process.exit(1);
+	}
+	const { server, port } = await serve(root);
+	const base = `http://127.0.0.1:${port}`;
+	const browser = await launch();
+	try {
+		const allHrefs = new Set();
+		for (const viewport of VIEWPORTS) {
+			const context = await browser.newContext({
+				viewport: { width: viewport.width, height: viewport.height },
+				isMobile: viewport.isMobile ?? false,
+				hasTouch: viewport.hasTouch ?? false,
+			});
+			const page = await context.newPage();
+			page.consoleErrors = [];
+			page.pageErrors = [];
+			page.requestFailures = [];
+			page.badResponses = [];
+			page.on('console', (msg) => msg.type() === 'error' && page.consoleErrors.push(msg.text()));
+			page.on('pageerror', (error) => page.pageErrors.push(String(error)));
+			page.on('requestfailed', (req) => page.requestFailures.push(`${req.url()} (${req.failure()?.errorText})`));
+			page.on('response', (res) => res.status() >= 400 && page.badResponses.push(`${res.status()} ${res.url()}`));
+			await page.goto(base + '/', { waitUntil: 'load' });
+			const hrefs = await checkPage(page, base, viewport);
+			await checkAccessibleNames(context, page, `[${viewport.name} ${viewport.width}x${viewport.height}]`);
+			hrefs.forEach((h) => allHrefs.add(h));
+			await context.close();
+		}
+		await checkLinks([...allHrefs], base);
+	} finally {
+		await browser.close();
+		server.close();
+	}
+	if (failures.length) {
+		console.error('browser check failed:');
+		for (const f of failures) console.error(`  ${f}`);
+		process.exit(1);
+	}
+	console.log(`browser check passed at ${VIEWPORTS.map((v) => `${v.name} ${v.width}x${v.height}`).join(' and ')}: ` +
+		'no console or resource errors, landmarks present, links healthy, no horizontal overflow');
+}
+
+main();
diff --git a/tools/check_scripts.py b/tools/check_scripts.py
new file mode 100644
index 0000000..468d25c
--- /dev/null
+++ b/tools/check_scripts.py
@@ -0,0 +1,65 @@
+#!/usr/bin/env python3
+"""Fail if dist/ ships any executable JavaScript.
+
+The page promises no client JavaScript. The old gate only looked for *.js
+files, which an inline <script> sails straight past. This walks every HTML
+file in dist/ and rejects:
+
+- any <script src=...>  — external executable script
+- any inline <script> whose type is not application/ld+json — that includes a
+  missing type, text/javascript, and module. All of them execute.
+
+application/ld+json is data, not code: it is the one script type this site
+ships (structured metadata in the layout) and the one the gate allows.
+
+    python3 tools/check_scripts.py dist
+"""
+import pathlib
+import sys
+from html.parser import HTMLParser
+
+ALLOWED_TYPES = {'application/ld+json'}
+
+
+class Scripts(HTMLParser):
+    def __init__(self, path):
+        super().__init__()
+        self.path = path
+        self.bad = []
+
+    def handle_starttag(self, tag, attrs):
+        if tag != 'script':
+            return
+        attrs = dict(attrs)
+        line, _ = self.getpos()
+        if 'src' in attrs:
+            self.bad.append(f'{self.path}:{line}: external script '
+                            f'<script src={attrs["src"]!r}>')
+        elif attrs.get('type', '').lower() not in ALLOWED_TYPES:
+            self.bad.append(f'{self.path}:{line}: inline executable script '
+                            f'<script type={attrs.get("type")!r}>')
+
+
+def main(root):
+    bad = []
+    pages = sorted(pathlib.Path(root).rglob('*.html'))
+    if not pages:
+        sys.exit(f'{root}: no HTML files — was the site built?')
+    for page in pages:
+        parser = Scripts(page)
+        parser.feed(page.read_text())
+        bad.extend(parser.bad)
+    if bad:
+        print('dist ships executable JavaScript — this site is meant to '
+              'ship none:', file=sys.stderr)
+        for line in bad:
+            print(f'  {line}', file=sys.stderr)
+        sys.exit(1)
+    print(f'{root}: no executable scripts in {len(pages)} page(s) '
+          '(application/ld+json metadata is allowed)', file=sys.stderr)
+
+
+if __name__ == '__main__':
+    if len(sys.argv) != 2:
+        sys.exit('usage: check_scripts.py dist')
+    main(sys.argv[1])