1\section{Ablations}2\label{sec:ablations}3The ablation grid of the protocol (predictor $\times$ depth $\times$ budget $\times$ policy $\times$ placement $\times$4transport $\times$ kernels $\times$ batch) is executed gate by gate. In this version only the single-node factors have logs:5\begin{itemize}6 \item \textbf{Install path} (per-miss synchronisation vs lazy batched installs): $+80$ to $+90\,\%$ decode throughput at 16 slots7 (Section~\ref{sec:implementation}).8 \item \textbf{Fast path} (device-side slot map when a layer is complete): 203 $\to$ 489~\tps{} at 64 slots.9 \item \textbf{Eviction policy} (LRU vs LFU): within 0.01 hit rate at every capacity in simulation.10 \item \textbf{Storage layout} (scattered safetensors vs expert-contiguous records) and \textbf{reader parallelism}: 8--10 $\to$ 18 $\to$ 36~\tps{}11 on the cold tier at 16 slots.12 \item \textbf{Predictor} (none / prev-rank / transition$_d$ / probe$_d$ / oracle), \textbf{depth} and \textbf{prefetch width} $m$: Table~\ref{tab:sim}13 (simulation) and Tables~\ref{tab:cold}--\ref{tab:attr} (online); the best online configuration ties with no prefetch.14 \item \textbf{Kernel fusion} (up+gate): no measurable effect (Table~\ref{tab:fused}).15 \item \textbf{Placement} (local vs odd experts remote over 1~GbE): $-80\,\%$ throughput (Figure~\ref{fig:twonode}); frequency- and16 cost-aware placement are exercised in Gate~F.17\end{itemize}18