Files
AndreaSlopFotografie/aidlc-docs/features/prerendering/construction/start-migration/code/verification-log.md
T
SluijsensandClaude Opus 5.5 8867c7833c Runs the browser tests in CI on the client output
The build artifact is now dist/client, so only that goes to the deploy jobs.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-31 19:06:41 +01:00

27 KiB

Verification log (V1 to V18)

Unit 1 start-migration. Rows are filled in the step named in the plan (section 3.2). Status: OK, BLOCKED, DEVIATION, open.

Resolved versions (Step 4, 2026-10-31)

Package Version Note
@tanstack/react-start 1.169.4 added exact (pnpm wrote no caret)
@tanstack/react-router 1.170.46 was ^1.170.39 (resolved 1.170.39); raised to Start's own version, see note 1
@tanstack/router-core 1.171.39 one copy for router and Start
@tanstack/start-plugin-core 1.171.55 transitive
@tanstack/router-plugin 1.168.48 transitive only (inside Start); the direct devDependency was removed
playwright-core 1.55.0 devDependency, exact
zod (4.x, transitive) 4.6.7 via override, see note 2
vite 8.3.1 unchanged

Notes:

  1. With only @tanstack/react-start added, pnpm kept two router copies (react-router 1.170.39 for the app and 1.170.46 for Start). The plan's fallback (raise @tanstack/react-router to the spike's 1.170.46) removes the second copy. No peer conflict was reported.
  2. pnpm audit:deps failed after adding Start: high advisory GHSA-mrhx-c7xc-m3w5 (zod < 4.6.6) through @tanstack/start-plugin-core and router-generator. Fixed by a patch bump only, with a scoped override in web/pnpm-workspace.yaml (zod@4: ^4.6.6, same pattern as the existing seroval overrides); zod 3 (Chakra CLI, Lighthouse) is not touched. Afterwards pnpm audit:deps: no known vulnerabilities. The main session should decide whether it accepts this extra override (BR-BLD-11 only names the seroval overrides as "stay").
  3. seroval and seroval-plugins overrides still apply (pnpm why seroval: 1.7.3, both router-core copies).

Table

V# Point Step Outcome Evidence Status
V1 Custom src/client.tsx hydrates with Start; startSentry() before hydration 4 Start picks src/client.tsx by name (default entry client in srcDirectory); it replaces the default entry, whose body is startTransition(() => hydrateRoot(document, <StrictMode><StartClient /></StrictMode>)). startSentry() runs at module level before that call. Build output dist/client/assets/index-*.js contains hydrateRoot and the Sentry loader (sentryClient-*.js dynamic import). Chrome (playwright-core, system Chrome, static server, no CSP yet): /, /over-mij, /contact, /contact?pakket=gezin hydrate with no console error or warning. Not proven here: that Sentry initialises (only on the test and production hostnames, so a no-op on 127.0.0.1); unit tests of monitoring/ cover that. OK
V2 RouterProvider in jsdom renders shellComponent 5 Yes, Router calls the root shellComponent (Match.js: ShellComponent = route.isRoot && route.options.shellComponent). React 19 treats html/head/body as singletons, so no <html> lands in the RTL container (container has 0 html/head/body), but it sets lang="nl" on document.documentElement and wraps ChakraProvider a second time. Variant used: renderApp replaces the root route's shellComponent by a test shell (HeadContent + children) on its router instance; production code untouched. Probe test (removed) printed shellComponent: function, html=0 head=0 body=0, document.documentElement lang=nl. All 888 tests pass with the test shell (with the real shell, 45 of 46 in app.test.tsx passed; the one failure was the /$ register test). OK
V3 Inline scripts per real page 4 (first look), 13 (final) Final list from the real runner on a scratch copy of the Step 4 output (12 HTML files): every page has 3 moved scripts, /contact has 5, 16 distinct files in total. (1) router scroll restoration, 613 chars, no attributes, same file on all pages; (2) Start bootstrap data-tsr-stream-part="", 916 to 1349 chars, one file per page; (3) boundary document.currentScript.remove();/*$tsr-stream-boundary*/, 56 chars, same file everywhere; only /contact: (4) <script id="_R_">requestAnimationFrame(...) 57 chars and (5) React completion runtime $RB=[];$RV=...;$RC("B:0","S:0") 845 chars. No <!--, no async/defer on an inline script, no inline module, so no UNSUPPORTED_SCRIPT or UNTERMINATED_SCRIPT. Positions (see Script positions below): scroll restoration is the last child of <main> (before the footer); bootstrap, the external module entry and boundary sit after the footer at the end of body; on /contact _R_ is followed by the streamed form (<div hidden id="S:0">) and the completion runtime is the last element before </body>. node config/externalizeInlineScripts.ts <scratch copy> printed only EMAIL_IN_OUTPUT (12 files) and wrote nothing; a throwaway script in the scratch folder (b2-scratch/tools/run.mjs) ran planExternalization with address: null: no violations, 16 script files plus 12 changed HTML files planned, applied, and a second plan had 0 writes and 0 violations. OK
V4 Minimum moved scripts 13 Keep the minimum of one (BR-SCR-17 unchanged). Every real page has at least 3 moved scripts (5 on /contact), so the minimum of one is far below what Start emits and cannot fail on today's output; its job is to catch a step that silently moves nothing, which one reference also catches. Demanding more than one adds nothing on current output and would break on a Start version that bundles scripts differently. A stricter check (for example requiring the bootstrap with data-tsr-stream-part) would be a design change; not made. Counts per page in V3. OK (recommendation: keep one)
V5 index.html written, crawling off 4 / is written as index.html with autoSubfolderIndex: false; /x as x.html, /soorten-fotoshoots/x as soorten-fotoshoots/x.html. Option names (from the start-plugin-core schema): prerender.crawlLinks: false and prerender.autoStaticPathsDiscovery: false (the latter stops Start adding static routes by itself), page list from Object.keys(sitePages) plus { path: '/404', prerender: { outputPath: '/404.html' } }. Prerender log: "Prerendered 12 pages". find dist/client -name '*.html' gives exactly 404, index, over-mij, portfolio, privacyverklaring, contact, soorten-fotoshoots and five soorten-fotoshoots/<shoot>.html (12 files, no others). Step 24: pnpm build (tsc, vite build, runner) exits 0 with the same 12 pages; vite.config.ts now takes NOT_FOUND_PATH and NOT_FOUND_FILE from config/pageFiles.ts. OK
V6 Code splitting; router plugin unused 4 (30 later) Start's plugin includes tanstackRouterGenerator and tanStackRouterCodeSplitter itself (client and server environments), so route code splitting is on without an option and @tanstack/router-plugin is not needed as a direct dependency. Removed with pnpm remove @tanstack/router-plugin; only Start's own copy (1.168.48) remains. Client build has one chunk per route (over-mij-*.js, contact-*.js, portfolio-*.js, newborn-*.js, ..., _-*.js for the catch-all), plus ContactForm, Lightbox, TestimonialCarousel chunks. Step 30 re-checks. OK
V7 Sitemap in the client environment only 4 this.environment.name === 'client' identifies the client build in generateBundle (Vite 8.3.1, Start environments client and ssr). dist/client/sitemap.xml exists once (10 <loc>, no /privacyverklaring, no /404); find dist -name sitemap.xml gives only that file; public/ has no sitemap.xml. Step 24: the same build gives the runner line Inline scripts verplaatst: 12 pagina's gecontroleerd, 16 scripts verplaatst, 28 bestanden geschreven (no violations), and dist keeps the sitemap in the client build only. OK
V8 Lazy ContactForm in HTML and hydration 27 What is true (differs from the plan text and BR-LZY-01 "the form fields are in the HTML inside the gate"): the prerender waits for the lazy chunk, so the form and its fields (data-testid="contact-form", contact-form-shoot, ...) are in the bytes of contact.html, but as a streamed Suspense segment, not in place. In <main> the gate holds <!--$?--><template id="B:0"></template> plus the empty FORM_PLACEHOLDER_HEIGHT box; the form sits in <div hidden id="S:0"> at the end of the body, and React's completion runtime ($RC("B:0","S:0"), moved to an external script by the runner) moves it into place. Without JavaScript the form is therefore in the HTML but not visible (isVisible false); FR-PR-13 allows that. With JavaScript, under the production CSP and the moved scripts, the form appears inside main, there is no securitypolicyviolation and no console message or hydration error. A late chunk is safe too (BR-LZY-05): with the ContactForm-*.js request held back 1500 ms, the page hydrates without any message and the form appears. e2e/hydration.e2e.ts "contactformulier op /contact (V8)", 3 tests, pass; also the 390 and 1280 px page tests of /contact. OK (BR-LZY-01 text to be corrected: fields are in the HTML only as the hidden streamed segment)
V9 Testimonials in HTML, hidden slides 21, 27 Code part (step 18, 21): renderToString of Testimonials holds the quote of every visible review; the copies besides the longest sit in a VisuallyHidden wrapper with aria-hidden and inert; axe has no violations in jsdom with the placeholder showing. Browser part (step 27): the raw HTML of / contains the quote and author of every visible review. Measured at 390 px with and without JavaScript: the longest review is 350 x 533 px in the flow, the other copy is 1 x 1 px, aria-hidden and inert; the testimonials section is 840 px high in both cases (no layout change at hydration while the carousel has not loaded). Reading order: the hidden copies are out of the accessibility tree, so each review is read once. Testimonials.test.tsx (4 new tests); e2e/rawHtml.e2e.ts "reviews op de homepage"; e2e/interaction.e2e.ts "reviews voor het laden van de carrousel (V9)" OK
V10 Placeholder sizes, CLS below 0.01 17, 34 First estimates (step 17), not measured: footer item minH 44px below lg, 1lh from lg (one line of the list item); contact card item minH 44px; gate notice inline block 38ch (lead " Mail me voor nu op " 20 + address 18 characters, from the parts without assembling) plus a hidden button with the recipe and label. To be measured in 27 and 34. Footer.tsx, ContactDetails.tsx, ContactFormGate.tsx open
V11 One title, one noindex meta 4 (normal pages), 27 (404.html) Normal pages in the prerendered HTML: one <title>, one description, one canonical, no data-fallback, a robots meta only on /privacyverklaring. 404.html: one <title> ("Pagina niet gevonden | Andrea Slop Fotografie") and one <meta name="robots" content="noindex">, no description or canonical. React writes <meta charSet="utf-8"> with a capital S (valid, attribute names are case-insensitive). scripts.mjs counts per file. After hydration in Chrome: /, /over-mij, /contact still 1 title. Caution for Step 27: /soorten-fotoshoots/onbekend showed 2 titles and 2 robots metas after React error #418 (the two extra 404 buttons render before hydration; Step 16 fixes that); /onbekend is clean. OK (final check in Step 27, see the V11 paragraph below the table)
V12 404 hydration with real shell 27 Found and fixed. With the real shell (navigation, footer, e-mail placeholders) the 404 HTML for /404 hydrated at another URL gave React #418 in a race, reproducible below 1024 px width and under CPU or network delay (Lighthouse). Root cause and fix: "404 hydration error #418" below the table (pendingComponent of the catch-all route equals its component). After the fix /404.html, /404, /onbekend, /soorten-fotoshoots/onbekend give no console message, no CSP violation, one title and one robots meta at 390 and 1280 px, and also with the route chunk delayed by 800 ms; /soorten-fotoshoots/onbekend shows one button without JavaScript and two extra after hydration. e2e/hydration.e2e.ts (48 tests); src/routes/hydration.test.tsx (4 new tests with hydrateAppAsStart). Without the fix 5 e2e tests and 4 unit tests fail. OK
V13 Timestamps and build paths 4, 13 Repeated on the final output of the runner (HTML plus the 16 moved script files plus all JS/CSS/XML/TXT/JSON): no build timestamp (YYYY-MM-DDTHH:MM) in any text file, no .map file, no repository root path (forward slashes, backslashes or doubled backslashes) in any text file. File names are content hashes, so they cannot carry a timestamp. Scan in the scratch run (repoPathFiles: 0, timestampFiles: 0, maps: 0); the runner's own BUILD_PATH_LEAK check reported nothing. OK
V14 Focus on first load, tidy navigation 27 Focus: on a first load of / (after hydration plus 500 ms) document.activeElement is not main; after a client navigation to /over-mij it is main; on /contact?pakket=gezin the tidy navigation (search removed) does not move focus either. Tidy navigation: /contact?pakket=gezin ends on /contact with the same history.length as before (replace, no extra entry) and scrollY 0, one render later than before as BR-HYD-11 expects. /?pakket=gezin#contact scrolls the contact section into view after hydration. e2e/interaction.e2e.ts ("zet de focus op main...", "verandert de focus niet...", "tidy navigatie op de contactpagina", "/?pakket=...#contact"), pass. OK
V15 Emotion/Chakra flicker 4 (styles in HTML), 27 (screenshots) Styles are in the HTML: <style data-emotion x69 in /over-mij, x120 on /, plus the bundled stylesheet <link rel="stylesheet" href="/assets/index-*.css"> (fonts and Chakra base). After hydration in Chrome /over-mij has 70 emotion style tags (one added by the client); no console warning. Step 27 screenshots (/over-mij, JavaScript off versus on, 390 and 1280 px; reviewed by eye): without JavaScript the header, navigation (desktop links or hamburger), eyebrow "Andrea Slop Fotografie" and the heading "Over mij" are fully styled, with the Jost and Marcellus fonts, in the same positions as with JavaScript; below the heading the page is empty (the Reveal content is hidden in the HTML, by design until unit 2) and with JavaScript the photo and text appear in that space (390 px: photo 237 to 499, then text; 1280 px: photo and two-column text). No unstyled flash, no moved header or heading; measured: position and width of h1 and the top of main are equal within 0.5 px with and without JavaScript at both widths. No console warning about styles. scripts.mjs; Chrome probe; e2e/interaction.e2e.ts "stijlen zonder en met JavaScript (V15)"; screenshots in the scratch folder (shots/over-mij-{390,1280}-{zonder,met}-js.png). OK
V16 Intl price text 27 The text nodes containing a euro sign are identical in the server HTML (JavaScript off) and after hydration in Chrome (production build) for / and the five shoot pages, and non-empty; no hydration error. Node and Chrome ICU give the same nl-NL text including the non-breaking space. e2e/hydration.e2e.ts "prijstekst (V16)", 6 tests, pass. OK
V17 No on*= or javascript: 13, 27 Step 13 on the runner output: 0 HTML files with an on*= attribute and 0 HTML files with javascript:. The only javascript: in dist/client is two strings in assets/index-*.js: React's own guard text (javascript:throw new Error('React has blocked a javascript: URL...), which is not markup and not a URL in a page. Step 27: no attribute starting with on and no javascript: attribute value in any of the 11 pages or 404.html (parsed with DOMParser), and no securitypolicyviolation on any of the 48 hydration page loads, the interaction flows and /contact with the reveal script, under securityHeaders('production'). Scratch run counts (onAttrFiles: 0); grep -l javascript: over the HTML and the bundle; e2e/rawHtml.e2e.ts and e2e/hydration.e2e.ts. OK
V18 CI artifact layout 28 The upload of web/dist/client/ puts index.html at the artifact root; the download to web/dist/client (lighthouse job) and the deploy steps' copy of the artifact contents (dist/* with dotglob, web/dist in the deploy jobs) put index.html at the destination root. dist/server is a sibling of dist/client, not inside it, so it never ends up in the artifact. Build output (pnpm build in this step's CP4 run): web/dist/client/index.html and web/dist/server/ exist side by side. Scratch emulation (scratchpad/v18): copy of dist/client/. as artifact root, dotglob copy of ./* into a destination; index.html at destination root, no server and no dist in it. OK

V11 step 27: in e2e/rawHtml.e2e.ts all 11 pages have exactly one <title> equal to pageHead, one description, one canonical; 404.html has one <title> ("Pagina niet gevonden | Andrea Slop Fotografie") and one <meta name="robots" content="noindex">, no data-fallback. In e2e/hydration.e2e.ts the same counts hold after hydration, at 390 and 1280 px, for the 11 pages and /404, /404.html, /onbekend, /soorten-fotoshoots/onbekend (one title; robots 1 on the 404 pages and /privacyverklaring, 0 elsewhere). The two titles and two robots metas seen on /soorten-fotoshoots/onbekend before Step 16 are gone. Status: OK. (Not covered: a URL the router matches but the server does not, see BR-404-07 below, ends with two titles.)

404 hydration error #418 on 404.html (root cause, Step 27)

Symptom: Lighthouse (Step 24) reported React error #418 on /404.html. In Chrome it reproduced deterministically below 1024 px viewport width (/404.html, /404, /onbekend, /soorten-fotoshoots/onbekend; two titles and two robots metas afterwards) and was clean at 1280 px.

Evidence for the mechanism:

  1. A build with development React and no minification (temporary config, deleted) gave the full message: the client rendered <Suspense fallback={null}> inside MatchImpl routeId="/$" where the server HTML has the <section> of the 404 page.
  2. The HTML is made for /404. Start dehydrates match ids with # for /: the bootstrap script has __root__# and #$#404. The browser at /404.html or /onbekend computes #$#404.html or #$#onbekend, so hydrate() in @tanstack/router-core (load-client.ts: dehydrated.i !== dehydrateSsrMatchId(candidate.id)) stops at index 1, sets pendingBoundary = 1 and presents that match as { status: 'pending', ssr: false }. A match with ssr: false renders <ClientOnly fallback={pendingElement}>; pendingElement comes from the route's pendingComponent (none) or defaultPendingComponent (none), so the first browser render of the page body was empty, against server HTML that shows the page.
  3. React hydrates that Suspense boundary on its own, later than the root. If the router has not yet finished loading the match (a status update replaces the pending match with the loaded one, which then renders the same page as the HTML), the boundary is hydrated with the empty client render: mismatch, #418, client re-render, and the hoisted <title> and robots meta of the server HTML stay next to the new ones. If the match loads first, nothing is reported. So it is a race, which is why the outcome depends on viewport width and on CPU and network speed (Lighthouse throttles both). Why width matters is only a timing effect (below 1024 px the page reproduced every time on this machine, from 1024 px not); it was not traced further. With the route chunk _-*.js held back 400 ms, 1280 px fails too; with it held back 0 ms, 1280 px passes.
  4. A plain pending component is also not an option left to chance: pendingComponent is not in the default split groups, so it stays in the route file, and the code splitter drops the NotFound import there when NotFound is also used by the split component (build error NotFound is not defined).

Why hydration.test.tsx missed it: hydrateApp creates the client router and awaits router.load() before hydrating, so every match is already loaded and the dehydrated-id comparison and the ssr: false pending presentation never occur. The browser has the router in the state Start's hydrate() leaves it in. The new hydrateAppAsStart (in src/test/hydrate.tsx) sets up window.$_TSR.router.matches with the server's ids, lets the router's own hydrate() run against a client router that has not loaded, and then hydrates. Four new tests in src/routes/hydration.test.tsx (/onbekend, /404.html, /over-mij/x, /soorten-fotoshoots/onbekend) fail without the fix with the same mismatch and pass with it. The e2e tests also fail without the fix (5 of 48 hydration tests: /404.html, /onbekend, /soorten-fotoshoots/onbekend at 390 px, plus the two delayed-chunk tests at both widths).

Fix (inside BR-HYD, no design change; BR-404-01 still holds): src/routes/$.tsx sets pendingComponent to the same page as component, through one shared CatchAll = () => <NotFound /> (needed because of the splitter issue in point 4). The first browser render of the unknown-URL match is then identical to the prerendered page, whether React hydrates the boundary before or after the match has loaded. After the fix, /404.html, /404, /onbekend, /soorten-fotoshoots/onbekend are clean in the production build at 412 and 1280 px, also with the route chunk held back 0, 400, 600 and (development build) 2500 ms. Side effect to know: when the chunk takes longer than pendingMs (1 s), Router shows the pending component next to the hydrated HTML (the server copy is hidden by React with display: none), so the page briefly holds two copies of the 404 content until the chunk arrives; invisible to the visitor, but a test that counts headings in that window would see two.

/over-mij/x-style URLs: the same mechanism applies to any match whose id differs from the dehydrated one. For the catch-all it is fixed. For a real page served as 404.html (BR-404-07) it is not, see below.

BR-404-07 observation (for unit 3)

/Over-Mij: the preview server on this machine (Windows, case-insensitive file system) maps it to over-mij.html with status 200 and the page hydrates cleanly, URL stays /Over-Mij. On a case-sensitive host the server would send 404.html with 404; replayed through a route handler in Chrome (production build, 1280 px): the router matches /over-mij, the page shows "Over mij", the document ends with two titles ("Over mij ..." and "Pagina niet gevonden ...") and no #418 in this run (the match loaded before React hydrated the boundary, so the outcome is the race of point 3), only the console line for the 404 status. /over-mij/: resolveRequest gives 404.html with 404 (BR-SRV-07 interim); in Chrome the router renders the real "Over mij" page at /over-mij/ (no redirect), again with two titles and the 404 console line, no #418 in this run. Both are observations, not assertions (e2e/hydration.e2e.ts "URL die de server niet kent en de router wel"); unit 3 must decide the URL rules so that server and router agree, and the stale 404 title and noindex meta in the head are the SEO consequence to weigh.

Other findings in Steps 3 to 5

  • Email address: info@andreaslop.nl is in all 12 HTML files (footer link, contact details, gate notice) and nowhere in dist/client/assets/*.js, robots.txt or sitemap.xml. Expected until Step 17.
  • The route tree gets a /$ route, so RoutePath = keyof FileRoutesByTo would force a /$ entry in sitePages. src/navigation.ts now uses Exclude<keyof FileRoutesByTo, '/$'>; the test "bevat precies de routes van de route tree" ignores /$. Not in the design text; needed for tsc (see proof-report).
  • Start's client.base default is /_build, but it is not used in this build: assets stay under /assets/ (needed for /assets/inline-<hash>.js of the runner).

Script positions per page (Step 13)

From the scratch run; positions are offsets in the rewritten HTML. A classic script that was inline now loads from /assets/inline-<hash>.js and blocks the parser at that point until fetched and run; that is a finding for the measurement (BR-MEA-03), not a design change.

Script Where it sits Pages
Scroll restoration (613 chars) Last child of <main>, after the last section and before </main> and the footer; about 78 to 92% into the document, so it blocks the parsing of the footer and everything after all 12
Start bootstrap (data-tsr-stream-part) After the footer, end of body, before the external <script type="module" async src="/assets/index-*.js"> all 12
Boundary (56 chars) Directly after the module entry script, end of body all 12
_R_ requestAnimationFrame (57 chars) After the boundary script, directly before the streamed <div hidden id="S:0"> with the contact form /contact
React completion runtime (845 chars, ends in $RC("B:0","S:0")) Last element of the body, after the hidden form segment (about 14 KB after the bootstrap), just before </body> /contact

On /contact the lazy ContactForm is a streamed Suspense boundary: its place in <main> is <!--$?--><template id="B:0"> and the form HTML sits in a hidden S:0 div at the end of the body. It becomes visible only when the completion script runs. That script was inline before the move and is now an external classic script, so the form appears one request later than in the raw Start output, and not at all without JavaScript. Relevant for V8 and for the measurement (CLS and LCP of /contact).

Lighthouse on the new preview server (Step 24)

Command: pnpm exec lhci collect --config=lighthouserc.cjs --numberOfRuns=1 --url=<URL> on a fresh dist/client; lhci starts pnpm preview --port 4173 --strictPort itself and stops it again. Port 4173 was free before the run. Only collect runs, so assertions and targets are not checked here.

  • /over-mij: exit 0, no runtime error. Performance 0.77, accessibility 1, best practices 1, SEO 1. LCP 6.2 s (score 0.11). Target for a photo page is 0.85 (see lighthouserc.cjs, to be checked in Step 34 under the measurement conditions).
  • /bestaat-niet (before the change): exit 1, Runtime error encountered: Lighthouse was unable to reliably load the page you requested. Make sure you are testing the correct URL and that the server is properly responding to all requests. (Status code: 404). Cause: the preview server answers 404 for the path. The fix is in the step: notFoundPath is now /404.html.
  • /404.html (after the change): exit 0, no runtime error. Performance 0.80, accessibility 1, best practices 0.96, SEO 0.58. Failing audits: errors-in-console (score 0): React error #418 (hydration mismatch) in index-*.js on this page; is-crawlable (score 0, expected, the page is noindex); meta-description (score 0, expected, 404.html has no description by design); LCP score 0.22 (5.3 s). The #418 belongs to V12 (404 hydration); Step 27 must check it again. The assertMatrix entry for notFoundPath requires best practices 1, so this URL fails that assertion until V12 is fixed; this step only checks that collect works.
  • notFoundPath change: lighthouserc.cjs line 22 is now '/404.html'. The guard test loopt gelijk met de URLs van Lighthouse CI in src/app.test.tsx now also checks it.