Most of our tools have the same shape: a form, then a card headed Result, then the findings. Each finding has a short bold title. On 281 pages, the script that draws the findings wrote that title as an <h4>, straight under the <h2>Result</h2>.
To a sighted reader the page looks fine. A screen-reader user who jumps from heading to heading hears "heading level 2, Result", then "heading level 4". Level 3 is missing, as if a section had been cut out. Lighthouse's accessibility audit counts this too (axe's heading-order rule).
Why nothing caught it
We already had a heading-order check. It reads the built HTML and strips <script> first, because code inside a script is not a heading. That is right, but it means the check never sees a heading that a script writes into the page later. The result card is empty until you press Run, so the check saw h1, h2, h2 and passed.
A check that presses the button
The new rendered-headings.mjs opens each tool in a headless Chromium, clicks the sample or example button if there is one, then the run button (Check, Calculate and so on), and reads the headings that are actually on screen. On the 300 pages in the first group that drew a finding, 270 skipped from h2 to h4. The other 30 put their findings under an <h3>, where <h4> is correct, so they were left alone.
The fix in each of the 270 pages is two words: <h4> becomes <h3> in the script, and the page's own .finding h4 { rule becomes .finding h3 {. Changing a heading's level can change how it looks, because browsers give h3 a bigger default size. So the same script recorded each finding title's computed font size, weight, line height, margins, colour and rendered height before and after. None of them moved on any page.
A second sweep over all 493 tool pages whose scripts draw headings found 11 more. Those needed a hand fix: one page had no .finding rule at all and needed one to keep the old size. One page (the VSME report builder) drew its form's section titles straight under the page title, so its form card now has a visible heading of its own. After that, 493 pages, 0 skips, 0 page errors.
Keeping it fixed
A cheap test would be better than a browser sweep. But guessing statically where a script writes its findings matched the browser on only 87% of pages, which is too weak to fail a build on. So finding-heading.test.mjs is a ratchet instead. It knows the 40 pages that still draw finding titles as <h4>, mostly correctly, and fails if a new page joins them. A new tool draws <h3>. If its findings really do sit under an <h3>, the test's comment says how to check that and add it.
Notes for the next run, including the 8 pages the clicker could not get a result from, are in the polish report.