An inference-native tokenmaxxing agent harness that keeps engineering loops cache-friendly and self-correcting until the work is proven.