Most of the 135 million users registered with Scratch by July 2024 were children, and the tools that grade, test and rank their games replay them in a browser, assuming that a program under the same inputs ends the same way in every browser. For one popular game in fourteen it does not: a game that works at home can end differently when the teacher grades it on another computer. The main cause is how Scratch tests touching: it compares collision outlines that the browser rasterizes and that Scratch keeps at the largest scale drawn so far, so the browser, the screen and the computer are hidden inputs of every game that tests touching with a vector costume. Our replay, built on Scratch's own virtual machine and renderer, fixes the keys, the random numbers and the clock, can pin the browsers' arithmetic, and names the first pixel test whose answer differs. We prove that browsers that draw every edge within a pixel can disagree about touching only near contact, and we give a fix: one outline per costume, drawn at a fixed scale by one portable rasterizer. We replayed 500 popular Scratch games, chosen before any evaluation run, for ten seconds of scripted play. Of the 473 that use no camera or web service and finish in all three browsers, 33 (7.0%) end differently across Chrome, Safari and Firefox, and the Scratch editor itself, in release Chrome and Firefox, splits 22 of the 33. Within one browser, a sharper screen or a larger stage changes 7.0% of the games, and moving release Chrome from a Mac to Windows changes 2.5%. With the arithmetic pinned, 24 of the 28 browser splits start at a pixel test, and feeding one browser the other's renderer answers reproduces the other's ending in all 24 games tested. The fix cuts browser splits to 1.3%, and those of every other setting on one computer to at most one game.