Rollout

What this stage is, and why its account came last

The screens of this stage were built before the account of it was written, and across four stages rather than one. Stage 07 coloured a sample of seven screens to prove the language, stage 09 built Alerts out of the system alone to prove the system, the public page arrived on its own in August, and the remaining ten screens were coloured in one pass afterwards, because a screen is built once, when everything it needs already exists. What was missing was never the pixels. It was the thing a stage is judged on: a page of its own, a row per screen, and a number that a script can reproduce.

So this page reports rather than plans. Every figure on it comes from an instrument in design/kit/screens/, each re-runnable and each deriving its corpus from a registry rather than from a list somebody typed. rollout12.cjs reads the source; width12.cjs walks all 55 pages across 58 widths in a browser; shape12.cjs reads what each screen does between a real 360 and a real 1440; fp12.cjs fingerprints every box on both sides of a change; inert12.cjs re-asks the nine "checked rather than assumed" sweeps on the corpus that exists now. And the counter it inherited was extended rather than trusted: counts.cjs had read the CSS header alone since stage 11, while the same sentence lives twice, in the header and on the component's page. It reads both copies now, and the first run of the second column found 18 pages disagreeing on a corpus the old run called clean.

The corpus: 55 and 55, and nothing on either side alone

Both counts exclude overview.html on each side, because a hub is a stand and not a product screen: the same basis the component counter and the inventory already use. The pairing is the whole of what "the rollout is built" means, and it is now a pairing rather than a claim: the two folders are read, matched by filename, and any file alone on its side is named. The count was 28 coloured pages when stage 09 closed and 32 when stage 10 did.

The audit row per screen, which stage 10 asked this stage for

Stage 10 wrote one row per screen saying what a wider window should give the person standing there, in three categories and a fourth that stayed empty: SAME, WIDER by grid, WIDER by air, and new behaviour nowhere. Ten of those seventeen rows had no coloured screen to check against at the time. Here is the built product measured against them, at a real 360 and a real 1440, with the scrollbar suppressed so the container is the number it says.

A screen carries no style of its own, re-proved on all 55

The rule in CLAUDE.md is four bans and one of them is the width rule stage 10 asked this stage to re-check: no style block, no style attribute, no @media, no class the system does not define, and no width literal anywhere in a screen file. Stage 10 proved it on 32 pages. It is re-proved here on all 55, by script, over the source rather than over the render.

The width column, recounted on the whole system

Every row of docs/inventory.md carries a "behaviour at width" cell, and an empty one would mean "behaves unknown on a wide screen". Stage 10 filled it on 60 entries and left a note saying the count would be re-taken when the corpus stopped moving, which is this stage. The corpus is 72 now: 69 component files and 3 patterns.

Nine cells were wrong and are corrected. Eight of the twelve new entries had their width cell filled with what the organism is rather than what it does with width, which is the column answering a different question; the ninth, the FAQ list, said "does not adapt" while its file carries a desktop point and a reading measure. Both are in docs/inventory.md now, and the prose that was displaced is under the organisms table rather than deleted.

The width sweep: 55 pages by 58 widths, and what it caught

Stage 10 swept 32 pages and its argument for sweeping rather than sampling was exact: a defect lives between the points, and the worst width is the one where the content has stopped fitting and the query has not fired. Twenty-three pages had never been walked. The run is 3190 measurements, and it drives the browser itself rather than asking a person to press Run.

Two of those four are not the clean answers they look like, and both are worth the sentence. The pane check is the one that found the 28px band at stage 10, where a layout scrolled sideways inside .screen while the document stayed clean; it is asked of every scroll container on the page here, and it is silent on all 55. And the landing runs 128 elements past the window on purpose: a marquee track is wider than its box and hidden by it. The first run of the sweep reported all of them at all 58 widths, which is a fault in the instrument rather than in the page, so a clipped element is now counted separately instead of being reported or silently dropped.

What it did find is prose, and all six are fixed. Six blocks read past the system's 52ch measure, four of them on screens that had never been swept. Each fix is a measure on the element that holds the words, in the component that owns it, and the re-run reads zero prose leaves past 52ch anywhere in the product, worst 52.0:

Every row this stage opened, and where it went

The stage began with eight rows addressed to it, four inherited from stages 09 and 10 and four found by its own instruments. All eight are closed, and three of them closed as a decision rather than as a fix, which is the distinction worth keeping: a row that is answered "no, and here is the measurement" is not the same as one quietly dropped.