\section{Conclusion}
\label{sec:conclusion}
We asked under which conditions each way of passing information between model executions preserves what matters at the lowest measured cost,
and we answered with a single harness on commodity hardware, real cache tensors, measured bytes and synchronized clocks, released records for
every attempt, and a protocol frozen before the test split was touched. The bounded findings are these. (1) Lossless representations are
interchangeable: text, its zstd encoding, the native cache, the reloaded cache and the 8-bit packed cache give identical answers; the choice
among them is a pure cost question, and the cost ranking flips between the cold case (text wins: no build) and the reused case (cache wins:
no prefill), with a 1~GbE link and a 3B model already favouring moving an 8-bit cache over re-reading at two thousand tokens.
(2) Every lossy choice we measured is model- or regime-dependent: 4-bit per-token quantization is free for Llama-3.2-3B and fatal for
Qwen2.5-3B; attention-based eviction is excellent with the question and useless without it, and no eviction survives a code specification;
LLMLingua-2 drops exactly the identifiers private-fact questions ask for. (3) At a 512-token sender budget, prose summaries carried more than
structured evidence with provenance, in single handoffs, in eight-hop relays and in fan-in, because provenance is paid in the sender's tokens;
relays themselves were nearly lossless after the first hop and invented almost nothing, and metered repair bought a few points for kilobytes of
re-reading. (4) A same-family linear KV mapper that reconstructs most of the receiver's cache variance transfers no content, and produces the
most dangerous kind of failure---coherent abstention---while an invalid cache concatenation produced plausible answers; compatibility manifests
and negative controls are therefore part of any cache-transfer protocol, not an afterthought. (5) A router fitted on development data matched
the best static policy at half the amortized latency when questions were unknown and offered nothing when they were known.

Open questions follow directly. The structured-evidence result should be re-run with budgets that separate provenance from values and with
senders large enough to fill them; the break-even map needs a faster receiver and a slower link to locate its crossing; the transfer ladder's
cross-family rung needs a trained mapper with generation and distillation terms, and its evaluation must keep the sender-alone and
receiver-alone references and an identity control beside it; and the fixed-point behaviour of structured relays suggests that the first
message, not the chain, is where compression research for multi-agent systems should concentrate.