\newpage
\begin{appendices}
\section{Registered hypotheses}
\label{app:hypotheses}
\input{appendix/hypotheses_verbatim}

\section{Source audit and unexecuted methods}
\label{app:audit}
Thirty-four sources were verified on 2026-10-29 from their arXiv pages, repositories and documentation (\texttt{literature/sources.csv};
notes in \texttt{literature/audit\_group\{1,2,3\}.md}). Table~\ref{tab:audit} summarizes the core methods.
\begin{table}[H]\centering\scriptsize
\caption{Audited core methods: availability of code and checkpoints, backend, question access, and WP30 status.}\label{tab:audit}
\begin{tabular}{@{}llllll@{}}\toprule
Method & Code & Checkpoint & Backend & Question access & WP30 \\\midrule
SwaT (2610.15095) & yes (no licence) & none & torch/CUDA & blind & not run \\
Attention Matching (2602.16284) & MIT & n/a (optimization) & vLLM 0.11 / torch & blind & not run (MLX port out of scope) \\
C2C (2510.03215) & Apache-2.0 & Qwen3 fusers & torch/CUDA & blind & not run \\
LatentMAS (2511.20639) & Apache-2.0 & n/a & HF/vLLM, CUDA & blind & not run \\
KIVI (2402.02750) & MIT & n/a & CUDA kernels & n/a & motivation tested via B2 \\
SnapKV (2404.14469) & Apache-2.0 & n/a & torch & aware & faithful re-implementation (B5) \\
PyramidKV (2406.02069) & MIT & n/a & torch & aware & budget rule re-implemented (B6) \\
H$_2$O (2306.14048) & no LICENSE file & n/a & torch & cumulative & not run \\
StreamingLLM (2309.17453) & MIT & n/a & torch & blind & retention pattern only (B3) \\
LLMLingua-2 (2403.12968) & MIT & yes (HF) & torch (MPS ok) & blind & official (A8) \\\bottomrule
\end{tabular}
\end{table}

\section{Correctness tests}
\label{app:tests}
\texttt{tests/test\_cache\_correctness.py} (12 tests, all passing on Qwen2.5-1.5B-Instruct-bf16): prefix equivalence in float32 and bf16
(distributional tolerance of Section~\ref{sec:correctness}, identical greedy continuation); no duplicated boundary token; identity
round trip with exact logits and measured bytes equal to \eqref{eq:mkv}; loader rejection of a wrong prefix hash and of a different
checkpoint; 100\,\% retention exact (shared and per-head indices); sparse mask path versus token-by-token append; preserved versus
relabelled positions differ; packed quantization bit widths and execution for 8/4/2 bits; no shared mutable state across branches;
attention selection keeps the window; no-message control. \texttt{tests/test\_leakage.py} (3) and \texttt{tests/test\_scoring\_and\_data.py} (5)
cover evaluator isolation, scorer behaviour and generator determinism.

\section{Prompts}
\label{app:prompts}
\input{appendix/prompts_verbatim}

\section{Pseudocode}
\label{app:pseudo}
\begin{algorithm}[h]\caption{Trial scheduler (resumable, append-only)}\footnotesize
\begin{algorithmic}[1]
\State load config, manifest, resources; $done \gets$ trial keys with status success in \texttt{runs/raw/<exp>.jsonl}
\For{document $d$ in selected items} \For{model pair $(S,R)$} \For{condition $c$} \For{seed}
  \State $P \gets$ questions of $d$ whose trial key $\notin done$; \textbf{continue} if $P=\emptyset$
  \If{$c$ is question-blind} build message/cache $m$ once for $d$ (reuse count $=|P|$) \EndIf
  \For{question $q \in P$} build $m$ if question-aware; run receiver; score with evaluator-only key; append validated record \EndFor
\EndFor\EndFor\EndFor\EndFor
\end{algorithmic}\end{algorithm}
\begin{algorithm}[h]\caption{Cache handoff pipeline (Track B/C)}\footnotesize
\begin{algorithmic}[1]
\State $\text{pre} \gets$ chat tokens of (system, document) that are a true prefix of the full (system, document, question) prompt
\State $C \gets$ receiver.prefill(pre) \Comment{dense bf16; measured doc prefill}
\State $C' \gets$ compress($C$): quantize / select positions (sparse cache with logical offset $|\text{pre}|$) / map from sender
\State payload $\gets$ serialize($C'$, manifest); optionally codec
\State at receiver: check manifest; deserialize; install; prefill fresh query tokens with retained-positions mask; decode greedily
\end{algorithmic}\end{algorithm}
\begin{algorithm}[h]\caption{Evidence retrieval (A6) and repairable chains}\footnotesize
\begin{algorithmic}[1]
\State refs $\gets$ A5 facts without values; rank refs by BM25(question, entity+attribute)
\For{ref in ranked order while lookups $<$ max and bytes $<$ budget} fetch source span $\pm$ context; count bytes and lookups \EndFor
\State chains: for codes mentioned in the previous message, fetch source sentences containing the code under the same meters
\end{algorithmic}\end{algorithm}
\begin{algorithm}[h]\caption{Adaptive policy (E8)}\footnotesize
\begin{algorithmic}[1]
\State on dev: for each (family, context bin, pair, access) and action $a$: $\hat q_a$, lower bootstrap bound $\underline{q}_a$, mean cost $\hat c_a$
\State feasible $\gets \{a : \underline{q}_a \ge \tau\,\max_{a'}\underline{q}_{a'}\}$; choose $\arg\min_{a\in\text{feasible}}\hat c_a$; fallback to the dev-best static action
\State freeze; on test: look up the cell (nearest bin fallback counted), apply the chosen action, add router time
\end{algorithmic}\end{algorithm}

\section{Per-attempt record schema}
\label{app:schema}
Each attempt is one JSON line validated against \texttt{wp30-attempt-1}: identifiers (study, experiment, trial, attempt, run group,
protocol and config hashes, git commit, timestamp), task (family, document id and hash, question id, split, cluster), models (sender and
receiver revisions, tokenizer hashes, backend versions, hardware id, device state, weight precision), method (id, fidelity, compression
config with realized values, question access, topology, handoff index, reuse count, cache hit, compatibility hash, seed), sizes (source
tokens under both tokenizers, query and message tokens, UTF-8 bytes, cache tensor bytes, serialized payload bytes, side information,
fetched bytes, transport bytes), timing spans and TTFT, usage and provider cached usage, memory (process RSS, MLX active/peak, prefill and
decode peaks, mapper memory), cost with source, prediction and evidence artifacts, key reference, scores, status, error code, retry link,
log path, payload and output hashes.

\section{Additional tables}
\label{app:tables}
\input{appendix/extra_tables}
\end{appendices}