\subsection{Cross-checkpoint transfer: the same-family linear mapper (E3)}
\label{sec:transfer}
Table~\ref{tab:transfer} reports the compatibility ladder's second rung on the synthetic (27 documents of 2k and 8k tokens, 8 questions each) and
executable (40) test items: the receiver Qwen2.5-3B reading the document itself (A1), the sender Qwen2.5-1.5B's own answers with full context
(sender alone), two text handoffs from the 1.5B sender (A4, A5), the identity-transfer negative control (C0: sender cache fed to the receiver
without mapping) and the ridge mapper (C2) alone and followed by question-aware selection (C3). The mapper was fitted on 60 training documents
(248\,569 token positions, 28.3~M parameters, 483~s on M3U96b, held-out reconstruction error $0.10$ for keys and $0.32$ for values, relative
Frobenius; \texttt{runs/checkpoints}); it needs the sender's cache and nothing else at the receiver (no source text, no prefill).

\begin{table}[t]\centering\scriptsize
\caption{Same-family transfer Qwen2.5-1.5B $\to$ Qwen2.5-3B (test split; M3U96a). Accuracy (\%, 95\,\% cluster-bootstrap CI), share of answers
containing ``unknown'', accuracy by question type, median payload and warm per-query latency. Negation questions are yes/no.}\label{tab:transfer}
\adjustbox{max width=\textwidth}{\input{tables/table_transfer}}
\end{table}

\paragraph{The mapped cache is well-formed but empty (H10 confirmed).} The receiver generates fluent, on-task text from the mapped cache---it
answers ``unknown'' 69\,\% of the time and never produces the garbled tokens of the identity control---but it recovers essentially none of the
content: 0\,\% on multi-hop, 2--4\,\% on single facts and exact copies, 0\,\% on code. Its aggregate score of 24.5\,\% comes from the yes/no negation
questions (93\,\%), which a receiver that answers ``no'' by default gets right at the base rate (70\,\% of projects are not cancelled); the sender's
own full-context answers show that the facts were available on the sender side (96\,\%). The 1.5B sender's prose summary transfers more
(28\,\%, 72\,\% on negation, 9--20\,\% on content) and its structured extraction less (12\,\%); the paired difference between the mapper and the
summary is $-3.7$ points (95\,\% CI $[-11.6, +4.2]$, $p=0.37$), and between the mapper and the receiver's own reading $-75$ points. Selection on
the mapped cache (C3) changes nothing. A linear map that explains 68\,\% of the receiver's value variance on held-out documents is therefore
not enough to carry a single identifier across checkpoints: cross-checkpoint transfer needs the generation and distillation terms that
published mappers train with, at a cost we did not have; the measured by-product is that a \emph{wrong} cache can be indistinguishable from an
empty one in the receiver's behaviour (coherent abstention), which is why Section~\ref{sec:fanin}'s manifests matter. Rungs 1 (same checkpoint,
two nodes) and 3 (cross-family) are covered by Section~\ref{sec:network} and the coverage table respectively; the cross-family rung was not
executed.