We started grading the engine on real websites, and five things broke immediately

Until now the engine was measured against a suite of test pages we wrote ourselves. It does well on those -> the typical one is within four percent of the reference browser -> and that had stopped being useful, because real websites still looked wrong. So this release changes what gets measured: twelve real sites are now loaded twice, once by the reference browser and once by Floati, and compared pixel by pixel. Five faults turned up in the first look, and each was found by looking at one pair of pictures rather than by reading a specification. a news site rendered as a completely blank white page. Its whole front page sits inside one very tall container, and the engine had a size limit that, when reached, gave up and drew nothing at all rather than drawing the part you can actually see. a link-sharing site rendered as though it had no styling: blue underlined links in the wrong typeface. A page can tell the browser "resolve my links against this address", and the engine ignored that instruction entirely, so the page never found its own stylesheet. Any site served from a different path than it was written for hit this. the encyclopedia's Appearance panel was pushed off the right edge of the window on every article, and the article column ran the full width underneath it. Three separate faults in how the engine divides a page into columns: a column could claim more room than the page had, a column meant to be as narrow as its contents was treated as a flexible one instead, and the engine was measuring the wrong item's width for each column. And on a tech news site the menu read "Sections" where every browser shows "SECTIONS": text placed directly inside a modern layout container was losing the styling it should inherit -> capitalisation, letter spacing, all of it.