SWE-Bench Performance Reaches 50.8% Without Tool Use: A Case for Monolithic State-in-Context Agents