This branch adds a minimal local-only LatentMAS execution path to OpenClaw. The goal is to reduce text-token overhead during multi-step reasoning by moving the internal collaboration phase from visible text into model latent states.