1\begin{appendices}2\section{Reproducibility}3\label{app:repro}4\paragraph{Software.} Python 3.12 virtual environment with MLX 0.32.4 and mlx-lm 0.32.2 (\code{requirements-lock.txt});5reference repositories audited at colibri \code{0a5504a} (2026-10-31), mlx \code{bd63d38} (2026-11-01), mlx-lm \code{736ab0d}6(2026-10-30), llama.cpp \code{a91eab6} (2026-10-31).7\paragraph{Commands.} Correctness: \code{.venv/bin/python -m pytest tests -q}. Capacity sweep:8\code{benchmarks/scripts/bench\_decode.py --backend \{mlx\_lm,titan\} --capacity C --predictor P --traces T}. Traces:9\code{benchmarks/scripts/collect\_traces.py}. Simulation: \code{benchmarks/scripts/simulate\_cache.py}. SSD tier:10\code{benchmarks/scripts/ssd\_bandwidth.py} after \code{sudo purge}. Node preparation: \code{scripts/cluster\_setup.sh}.11\paragraph{Raw data.} \code{benchmarks/raw/gateA/} (smoke baseline, SSD), \code{gateB/} (design microbenchmarks, capacity sweep,12tests), \code{gateC/} (traces, online prefetch, miss-cost breakdown, Qwen3 baselines); processed simulations in13\code{benchmarks/processed/}. Logs are append-only JSON lines.14\paragraph{Prior-art table.} \code{research/prior\_art.csv} (59 rows, evidence URL per row) and15\code{research/RELATED\_WORK.md}.16\paragraph{AI-assistance disclosure.} The software, experiments and this manuscript were produced by Claude Code (Anthropic,17model Claude Fable 5.1) acting as an autonomous research engineer under the author's written mandate18(\code{docs/MANDATE.md}); the author set the research question, hypotheses, integrity rules and gates, and is responsible for19the content. Every quantitative claim traces to a logged measurement; the agent's decisions are recorded in20\code{research/DECISIONS.md} and its chronology in \code{research/LOG.md}.21\end{appendices}22