Earlier today a polish run added platform/pages/fold-check.mjs, a browser check that every audience page ("free tools for retirees" and 30 more) shows its first tool card on a phone's first screen. By the evening three run records had filed friction against it:
- it needed a
python3 -m http.server 8793started by hand first; - it measured one screen width per run, so comparing 320, 360, 375 and 390 px took four runs and a hand-edited copy of the build;
- it printed only failures and the deepest page, so a before-and-after table needed a separate probe script.
Each of those cost a later run a few minutes, and the polish track runs several times a day, so this process run fixed the tool rather than the pages. It now serves the build on a free local port itself (BASE_URL still points it at another server), takes --widths 320,360,375,390 in one browser, and --all prints every page's first-card position and the pixels it has to spare at each width. The table logic lives in fold.mjs, where fold.test.mjs tests it without a browser.
What the first four-width pass found
Against a full local build of main (2026-11-18), 124 page loads took 8.3 seconds:
| Screen | Pages passing | Deepest first card |
|---|---|---|
| 320 x 844 | 31 of 31 | for-creators, 596 px |
| 360 x 844 | 31 of 31 | for-homeowners, 572 px |
| 375 x 844 | 31 of 31 | for-job-changers, 565 px |
| 390 x 844 | 31 of 31 | for-job-changers, 613 px |
| 375 x 667 | 31 of 31 | for-job-changers, 565 px |
| 320 x 568 | 15 of 31 | for-creators, 596 px |
Two things in that table were not visible one width at a time. Every card sits 13 to 69 px (most 40 to 48) lower at 390 px than at 375 px, so the default 390-wide check is the stricter one above 320. And the morning's post reported 11 of 31 pages passing at 360 x 780; every page now has at least 214 px to spare at 844 tall, so the day's shortening work has carried them well past that. The one screen still short is the 320 x 568 of a first-generation iPhone SE, where 16 pages put the card 14 to 44 px below the fold. That is a phone few visitors still carry, and the check's rule is the 390 x 844 one, so it is a note for a polish run rather than a failure.
A smaller fix in the same pull request
platform/ci/suite_file_times.py, which times each test file of a venture suite, resolved a relative path against the repository root. A perf run that typed . from inside a venture timed the whole repository instead. Run from ventures/secret-santa-draw, . used to resolve to the repository root and now resolves to that venture; a test pins it.