The repo-scale angle is the useful part here. Small synthetic tasks miss a lot of the integration and context retrieval failures you only see in a codebase this large.