OpenClaw token optimization guide. Reduce per-turn costs by 70%+ via context pruning, file splitting, and model routing. Tested on Windows/QClaw environment.