Document 09 of 14 · docs/QA.md

Verifying by driving, the sweep

Nothing on this site is verified by reading it. Every claim the site makes to a visitor has an assertion that drives the page and checks. studio-src/sweep.cjs runs them all; studio-src/pdfcheck.mjs checks the files. Both are hand-run and run in the weekly Action. When an assertion is added, it is proved non-vacuous first: break the thing on purpose and watch it fail. Three false alarms on measureandbuy came from assertions that had gone stale, so check the assertion before believing a regression.

1. Every page, at 1440 and at a real 390 viewport

• No horizontal overflow (scrollWidth === clientWidth, and no element past the right edge outside an intended horizontal scroller).
• Zero console errors.
• Every image has width, height and alt.
• The shell renders: header, nav with the same links on every page type, footer with the Synergy mark and the legal links.
• Contrast: every pair in DESIGN.md section 2 measured from computed styles, 4.5:1 or better for text, 3:1 for focus rings. Measured whole-surface, not one pair.
• Focus ring visible on every interactive element by keyboard.
• Fraunces or Newsreader actually loaded (document.fonts.check), and zero requests to any host but our own (and Stripe, on the customizer's buy click only).

2. Template page

• H1 equals meta.title. Canonical, og:*, product:price:*, twitter tags present and pinned to www.
• Three JSON-LD blocks parse and validate: Product with an Offer whose price equals the tier, BreadcrumbList, FAQPage with the same five questions as the page.
• The live sheet iframe renders and its theme chips repaint it (assert the --t-accent computed value changes).
• The paid CTA carries the price; the free CTA resolves to a 200 with Content-Type: application/pdf.
• Related grid: at least three cards, every card a link hit-tested at five points.
• Inbound links to this page from elsewhere in the site: at least three.

3. Customizer

• Every field in fields.json is rendered, with its label, counter and limit.
• Typing in every text field changes the sheet (assert the bound element's text).
• Overflow guard: set every text field to its longest allowed value at once; the sheet must not overflow anywhere (overflow() returns empty). Then set the recipient (or the widest-set field) to max + 1 characters: the warning appears, the button disables and reads "Fix the fit to continue". Restore: the button returns. This test runs on every template, every build.
• Every optional field empty: the sheet still renders with no blank labels showing (the measureandbuy lesson: non-default states need their own check).
• Theme chips: every theme applies; the light-ink chip is present and labelled printer-friendly.
• Paper size control switches the a4 class and the sheet's aspect ratio.
• Logo dropzone accepts a PNG under 400 KB, rejects an SVG with a plain message, downsizes a 3000 px image to 1200.
• List editor: add to the cap, then the add button disables; remove; reorder; totals recompute (invoice: 3 lines at the defaults total 1,905.00 before tax).
• Autosave: reload the page, the values persist.
• Mobile: the stage is 40vh, expandable; the fixed bottom bar shows until the real button scrolls into view, then hides.
• Print stylesheet: under emulated print media the stage is hidden.
• noindex present in the head and in the response headers.

4. The files (pdfcheck.mjs)

For every template's free file and for the two files of a test-mode order:

• Page count equals meta.pages.
• Page size, read from the PDF's MediaBox: Letter 612 by 792 points, A4 595.28 by 841.89 (within 1 point), rotated for landscape templates.
• Every font referenced is embedded (a FontFile2 or FontFile3 stream for each font resource, no base-14 substitutes) and no Type3 font dictionary exists, which is what a variable font turns into; see spikes/render/README.md. Chromium compresses these dictionaries, so the checker inflates every stream before searching.
• The credit line is present in the free file's text and absent from the paid file's.
• The recipient's name appears in the text of the paid file (so the values really reached the render).
• File size under 2 MB.

Locally without poppler this is a small Node script reading the PDF objects; with poppler, pdfinfo and pdffonts do the same and the numbers must agree.

5. The purchase, in test mode, by hand, on a phone

1. Customize the invoice with a logo and twelve lines.
2. Pay with Stripe's test card. The success page shows the waiting state, then READY inside fifteen seconds, with two buttons.
3. Both PDFs open in Preview and in Acrobat. Fonts embedded, both sizes correct, the logo present.
4. The email arrives inside a minute with the same two buttons and the same order id.
5. The next day the link in the email still works. A refund from the Stripe dashboard flips the order to refunded and the link stops.
6. Repeat once with a delayed-payment test method to prove async_payment_succeeded is handled.

This is the Phase 1 acceptance test and it is not done until all six pass.

6. Launch checklist (cutover day)

• The free file of every live template downloads from www.
• sitemap.xml lists every template, month page, category and help page, and nothing noindexed.
• robots.txt serves. The Rich Pin validator passes on one template page.
• GA4 fires on www and not on the Vercel preview URL (the host guard, checked with the blackhole method from measureandbuy's ga-check.cjs).
• The old paths redirect: /auth, /dashboard, /admin, /payment-success to /; /buy/x and /customize/x to /templates/.
• One real purchase with live keys, refunded immediately.
• HANDOFF.md updated with the date and the numbers.

7. The weekly Action

Runs the sweep against www, pdfcheck on the free files, one end-to-end render through Browser Run into a scratch path (then deleted), the Supabase keep-warm query, and a link check. Opens one issue (commented on, not duplicated) when anything fails. It never changes anything; it tells a person. Same posture as measureandbuy's monthly source check, and PIPESTATUS[0] not $? after any pipe, for the same reason.