\section{Robustness}
\label{sec:robustness}
Held-out prompt families give hit rates within 0.01 of the fit families at every capacity (v1 traces), and the transition
predictor's recall drops from 0.53 (fit) to 0.50 (held-out), so the locality statistics are not an artefact of the fitting
prompts. Throughput measurements on the laptop vary by less than 1\,\% across repeats for the all-resident configurations and by
up to 40\,\% across prompt families for demand loading (code prompts route more repetitively than general prose, hence higher
hit rates). Thermal state and competing processes are not controlled in the smoke tests; the cold-NVMe sweeps were run alone on the
laptop. Table~\ref{tab:audit} gives bootstrap confidence intervals for every comparison used in the text
(\code{scripts/audit\_stats.py}, 10\,000 resamples, seed 0, raw logs only): the all-resident parity with mlx-lm holds within
$-5.9$ [$-7.1$, $-4.8$]~\tps{} on the laptop and $-0.1$ [$-0.3$, $+0.3$] on M3U96a; every prefetch configuration's interval
excludes a gain except the depth-2 transition predictor at 16 slots, whose interval straddles zero ($-1.4$ [$-7.3$, $+6.0$]); the
fused-kernel differences are within $\pm 1$~\tps{}; and the two-node loss is $-66$ [$-72$, $-55$]~\tps{}.

\input{tables/audit}