The worst score on the board turned out to be a cookie banner

Last release started grading the engine against twelve real websites instead of test pages we wrote ourselves. The single worst number on that board was a news site, at ninety-eight percent of pixels different from the reference browser -> more than half the total error of every site combined. It had never been looked at, because a number that large reads as hopeless rather than suspicious. It was a cookie banner. The reference browser was showing the consent dialog and Floati was showing the article underneath it, and the two pictures were being compared as though that were a rendering fault. It had also invented a second fault that did not exist: the dialog locks the page from scrolling, which made the reference browser report the page as a tenth of its real height, so the first minutes of the investigation went into chasing a layout collapse that was never there. Both browsers are now given the same list of hosts to refuse, so the comparison is of the page. Underneath it were seven real faults, and every one was found by shrinking a broken page down to the smallest version that still broke. An image that fails to load was drawing a blank box a hundred and fifty pixels wide instead of the small marker every other browser draws, so a single blocked tracking pixel pushed a whole page down. An empty container was a full line tall, which sounds harmless until you count the empty containers a modern site puts at the top of its page. The parser behaved as though scripting were switched off, so the fallback content sites write for people without JavaScript was being drawn on top of the real page. An icon coloured through a theme variable -> which is how nearly every design system colours icons -> was drawn as nothing at all. A typeface named only in an element style attribute was silently replaced with a default one, making text measurably narrower than it should be. Then a set of faults in how the engine reads the rules that say when a style applies. A rule guarded by "this OR that" was not understood, and rather than being ignored it was treated as always true, so styles meant only for a phone were applied on a desktop. Rules meant for a screen reader were applied to the screen. And a container measuring how wide it needed to be forgot to count the gaps between its own items, so rows of icons wrapped onto three lines when they should have sat on one. Across the twelve sites the total difference from the reference browser fell by thirty-six percent. The engine conformance gate rose to 6370 checks.