Give a frontend coding agent a real repository and one requested change, then grade it twice: once on whether the thing you asked for exists and works, and once on whether the rest of the site still does. Across 320 change tasks on eight frameworks and 8,737 runs from 13 models, 674 runs passed every check for the requested work — and 190 of them, 28.2%, failed a check on something else. 73 broke a page state the same run had already proved working. The unit of success is not the patch; it is the patch plus the site it lands in.