The site has a check, platform/pages/sideways.mjs, that loads pages in a 360 px window and flags any page wider than the screen. That is WCAG 1.4.10, Reflow. A full run over every page takes a long time, so most runs use SAMPLE=300, which picks 300 pages evenly spaced through the built site.
Earlier today a hand screenshot found the research index 428 px wide on a phone. The sample had never complained, and the reason is arithmetic. The build makes 8,223 HTML pages, mostly tool pages and blog posts. Evenly spaced samples of 300 and of 400 pages contained none of the eight pages a visitor is most likely to land on: the hub, /research/, /writing/, /compare/, /guides/, /about/, /engineering/ and /stories/. A sample of 1,000 contained one.
The fix is small. A sample now always starts with every top-level index page the site itself renders (16 today), then spreads the rest of its budget evenly. To check it, I put an unbreakable file path back into /research/ and ran the 300-page sample. It reported "+345px research/index.html, widest: code". On the clean build it found 0 of 300 pages scrolling sideways at 360 px and also at 320 px.
One trap along the way: the first version decided whether a page was "site-owned" by looking for the site stylesheet in the first 20 KB of the file. /guides/ links that stylesheet 328 KB in, after its inline styles, so the first version still skipped it. The check now reads the whole file.
The code is MIT licensed: platform/pages/sideways-sample.mjs and its test.