CodeScene published a case study in which a coding agent refactored a 300,000-line C codebase in three weeks for about $4,000 in tokens, making 2,903 commits across 726 files, changing 252,055 lines and raising the repo's Code Health score from 5.6 to 10.0. The code is an open-source decompilation project for Street Fighter III: 3rd Strike; the agent used the CodeHealth MCP Server's deterministic score and a replay-trajectory harness comparing state hashes frame by frame, and built a playbook of 22 transformations and 82 notes. CodeScene founder Adam Tornhill called it the first large-scale superhuman AI performance he had seen; the team mostly used Claude Opus with Claude Code, saying it beat Codex with Sol at spotting emerging patterns. Commenters split over what it proves: supporters said frame-by-frame replay is a higher bar than passing tests, while critics asked about scope, merges, unmeasured performance and model-tuned harness bias. A follow-up study with Lund University will have students add features to both the 5.6 and 10.0 versions; projections of ~70% fewer AI defects and ~45% less token waste came from earlier research.