Cache immutable Wasm server configuration encoding

docs/wasm-runtime-architecture.md

@@ -87,9 +87,11 @@ the boundary).
   children yielding `const DValue&`.
 - The request context (`params`/`get`/`post`/`cookies`/`session`, the raw body
   `in`, and—for WS—the connection context) is marshalled into a single `ctx`
-  DValue, UCEB-encoded, and handed to the workspace. The response (body,
-  headers, status, and any `meta` such as `ws_commands`) comes back the same
-  way.
+  DValue and UCEB-encoded. Immutable server configuration is UCEB-encoded once
+  per worker as a flat string map; both byte ranges are written into one guest
+  buffer. The guest decodes the flat map directly into its fresh `Server` and
+  the dynamic tree into its fresh `Request`. The response (body, headers,
+  status, and any `meta` such as `ws_commands`) comes back as UCEB.
 
 See [`docs/wasm-phase1-dvalue-abi.md`](wasm-phase1-dvalue-abi.md) for the wire
 format details.
@@ -160,10 +162,11 @@ drops one empty workspace. This preserves Wasmtime's fork boundary while
 preventing the first request assigned to each worker from paying engine, linker,
 or pre-instantiation startup.
 Server configuration is immutable by that point. The worker therefore retains
-one native `DValue` view of it; each request's temporary context tree references
-that view while UCEB is encoded, avoiding a repeated native tree copy. The guest
-still decodes the bytes into its fresh request tree. Request parameters, body,
-cookies, session, call data, and response state are never retained this way.
+one native `DValue` view and one UCEB encoding of it. Each request transfers
+those cached bytes and the guest decodes the flat scalar map directly into its
+fresh `Server`, avoiding both repeated native encoding and an intermediate guest
+`DValue` tree. Request parameters, body, cookies, session, call data, and
+response state are never retained this way.
 Startup duration or failure is written to the service log. The serialized core
 module lives in the configured writable cache root rather than beside the
 possibly root-owned deployed `core.wasm`; freshness still uses the deployed
@@ -285,7 +288,8 @@ request rather than being deep-copied twice; the historical by-value
 artifacts stay ABI-compatible. `request_perf()` subdivides birth into policy,
 import materialization, core instantiation, export/table lookup, and initialization,
 and context transfer into bytes, host encode, guest allocation/write,
-guest decode/application, and free.
+guest decode/application, and free. The byte profile separately reports the
+worker-cached server-configuration portion.
 
 ### Task callbacks and workspace lifetime