papers/lp-201-zap-universal-wire/sections/proposal.tex

% sections/proposal.tex -- ZAP-everywhere migration proposal
\section{Proposal: ZAP everywhere}
\label{sec:proposal}

\subsection{Tier model}

Each non-ZAP surface is classified by frequency on the call graph, not by
team or repository:

\begin{description}[leftmargin=0pt,style=nextline]
  \item[Tier~1 (Hot path).] High-frequency wire: per-block, per-tx,
    per-RPC, per-tool-call. Cost of a non-ZAP byte is paid in CPU,
    allocations, GC pressure, and tail latency on every transaction.
    \emph{Action: URGENT migrate.}
  \item[Tier~2 (Warm path).] Per-session, per-handshake, per-page-load.
    Cost is paid once per user session but compounds at fleet scale.
    \emph{Action: SCHEDULED migrate.}
  \item[Tier~3 (Cold path).] Boot, config, admin, one-shot CLI. Cost is
    negligible per call but format heterogeneity is a maintenance tax.
    \emph{Action: OPPORTUNISTIC migrate.}
  \item[Tier~4 (External boundary).] Talks to third parties whose wire
    we do not own (Binance, Stripe, OpenAI, Plivo). \emph{Action: keep
    external format; tunnel internal hops in ZAP only.}
  \item[Tier~5 (Already ZAP).] Verified zap-native today.
\end{description}

\subsection{Top-5 URGENT Tier-1 targets}

The following surfaces are simultaneously (a)~highest-frequency wire on
the call graph and (b)~currently zero-ZAP. They unlock the largest
absolute bench delta per engineering hour invested.

\begin{enumerate}[leftmargin=*]
  \item \textbf{P-Chain blocks + tx codec.}
    \texttt{lux/protocol/p/\{txs,block,warp,state\}/*.go} plus the
    \texttt{vms/platformvm} composition root. LP-023 already measured
    $7.50\times$ Parse geomean on the zap\_native subtree
    benchmarks~\cite{lp023zapnative}. Blocked on \texttt{lux/utxo}
    completing its ZAP migration so the cross-chain UTXO wire
    stabilises.
  \item \textbf{X-Chain UTXO wire.}
    \texttt{lux/protocol/x/txs/*.go}~--- identical shape to the
    P-Chain migration; same $7.50\times$ Parse projection applies. Can
    be migrated in parallel with~(1) since the codec is structurally
    isomorphic.
  \item \textbf{Hanzo Agent control-plane.}
    \texttt{hanzo/agent/control-plane/internal/\{handlers, services,
    cli, mcp, infrastructure/communication\}}. Currently
    JSON-RPC$+$HTTP$+$gRPC heterogeneous mix on every agent-to-agent
    tool call. ZAP/MCP-bridge bench shows $17\times$ serialisation,
    $345\times$ parse, $91\%$ memory reduction
    (\texttt{zap/README.md}~\cite{zapreadme}); apply that ratio to the
    $\sim$$140{,}000$ calls/sec baseline measured in
    \texttt{BenchmarkMCPOrchestrator20Tools}.
  \item \textbf{Hanzo MCP server.}
    \texttt{hanzo/mcp/go/*.go} plus \texttt{rust/}, \texttt{src/}. The
    canonical MCP wire is JSON-RPC over stdio; a ZAP transport variant
    (the \texttt{luxfi/zap/mcp} bridge) already exists. The migration
    is a transport swap, not a protocol change~--- tool authors keep
    writing the same MCP schemas; only the bytes on the wire change.
  \item \textbf{Hanzo Jin tensor wire.}
    \texttt{hanzo/jin/jepa/} multimodal model server uses
    framework-native serialisation (JSON manifests + pickle/pt tensor
    blobs). A ZAP tensor schema (typed fixed-stride leaves over the
    ZAP segment) replaces both manifest and blob with a single
    zero-copy frame. The bench delta is not yet measured but the
    structural argument is identical to the
    \texttt{BenchmarkZAPOrchestrator20Tools} result:
    deserialise-by-pointer instead of allocate-and-decode.
\end{enumerate}

\subsection{Phased rollout, by dependency}

\begin{description}[leftmargin=0pt,style=nextline]
  \item[Phase~A~--- consensus completion.] Finish what LP-023 and the
    \texttt{utxo} ZAP migration started: empty the
    \texttt{FINAL\_CODEC\_RIP} ledger, delete \texttt{luxfi/codec} from
    the production tree, and rip the duplicate \texttt{lux/proto/}
    snapshot.
  \item[Phase~B~--- AI/ML hot wire.] Agent, MCP, Jin migrations in
    parallel. These have no inter-dependency on Phase~A; gating only
    requires the canonical \texttt{luxfi/zap} reference
    (\texttt{zap-es} for TS clients) which is already shipped.
  \item[Phase~C~--- gateway internal tunnels.] \texttt{zen/gateway}
    keeps OpenAI/Anthropic/Together JSON outbound (Tier~4) but
    Hanzo$\to$Hanzo back-plane traffic moves to ZAP-tunnelled hops.
  \item[Phase~D~--- legacy cleanup.] Delete \texttt{lux/proto/} (Tier~3
    duplicate). Re-verify zero \texttt{luxfi/codec} imports across the
    entire workspace.
\end{description}

\subsection{Constraints and external boundaries}

\begin{itemize}[leftmargin=*]
  \item Tier~4 boundaries (market data, payments, third-party LLM
    inference) are kept on their vendor formats. The ``ZAP everywhere''
    claim is bounded by what Hanzo, Lux, and Liquidity \emph{own}.
  \item The Quasar activation at unix~$1766708400$
    (2025-12-25T16:20:00-08:00) is the cut-over point: every L1 EVM
    boots from canonical genesis with ZAP-only wire from block zero
    forward (memory \texttt{network\_activation\_2025\_12\_25}).
    Pre-activation P/X blocks remain decodable via the v0 legacy
    decoder retained in \texttt{node/vms/platformvm/block/v0}.
  \item No backwards-compat shim is introduced. Following the
    workspace rule ``forwards perfection only,'' migrated code paths
    have one wire (ZAP) and one composition root.
\end{itemize}