Both groups have issues, but they look different. Heritage repos show accumulated coordination burden: protocol evolution, config hubs, large controllers, years of backfilled responsibilities. AI-heavy repos show initial over-aggregation: one adapter, one dataloader, one plugin file with too many methods. AI-heavy projects also had about 2x the testing issue density.
You can explore the full analysis of all the projects in the showcase at https://app.octokraft.com/showcase/BENCHMA785