“Standard sandbox-based training methods for coding agents do not typically capture long-term software health metrics such as maintainability, ownership boundaries, migration costs, or future debugging burdens.”