1\section{Conclusion}2\label{sec:conclusion}3We have built an MLX-native MoE runtime whose bounded expert residency is semantically invisible (bit-identical logits to mlx-lm)4and whose all-resident path costs nothing (489 vs 494~\tps{}), characterised the cluster it will run on as it actually is5(gigabit Ethernet, no Thunderbolt), and obtained a first negative result that the formulation explains: one-layer-ahead6prefetching cannot pay when the fetch latency exceeds the per-layer compute time. The next steps are fixed by the protocol:7depth-$d$ and token-horizon prefetching against the cold-NVMe tier, the medium model on nodes that cannot hold it, the8two-node TCP transport with static placement, the cost-aware placement of Algorithm~\ref{alg:placement}, profiling-driven Metal9work, and the energy frontier. Each will be reported with its raw logs, including the cases where \titan{} loses.10