docs(arch): stream-client split + total-bytes cap for subscription forwards (FWD-15, FWD-14)

http-adapters.md: forwarding-handler step 5 documents the time/bytes
split for Sub forwards (stream client without the total request
timeout, total streamed-bytes cap, read timeout as the staleness
guard); the HTTP Client section documents the two derived clients
(request vs stream), the reqwest 0.13 per-request-override limitation
that forces the derived-client design, and that only the total timeout
differs between them.

http-server.md: the per-endpoint dispatch deadline paragraph now covers
the outbound half — unbounded time, bounded bytes on both the gateway
and the forwarding side (ADR-049/021 note: unbounded time by design for
subscriptions, bounded bytes per subscription).

Verified: cargo doc --no-deps clean, full suite green.

docs/architecture/http-adapters.md

@@ -181,7 +181,7 @@ The forwarding handler is stored in the `HandlerRegistration` as a
    wraps it in a `ResponseEnvelope`, returns. Registered as
    `HandlerKind::Once` — a `Handler` returning a single
    `ResponseEnvelope`.
-5. For a `Sub` (`text/event-stream` response): streams
+ 5. For a `Sub` (`text/event-stream` response): streams
     `call.responded` events as the SSE chunks arrive (same SSE parsing as
     the TS `parseSSEFrames`), then the stream ends on SSE close (which
     becomes `call.completed` on the wire). Registered as
@@ -189,6 +189,22 @@ The forwarding handler is stored in the `HandlerRegistration` as a
     `BoxStream<ResponseEnvelope>` (alkcall ADR-021). Each SSE `data:`
     frame becomes a `ResponseEnvelope::ok()`; an HTTP error (non-2xx)
     becomes a single `ResponseEnvelope::error()` and ends the stream.
+    The streaming send rides the shared HTTP client's
+    **stream client** (see HTTP Client below): the same config minus
+    the total request timeout (FWD-15) — a subscription is unbounded in
+    *time* by contract (alkcall ADR-021 sets `deadline: None`), so the
+    outbound half must not die at the 30 s request deadline the
+    request/response half carries. Time-unbounded does not mean
+    memory-unbounded: the SSE parse loop enforces a **total
+    streamed-bytes cap** per subscription
+    (`HttpClientConfig.stream_total_byte_cap`, default 1 GiB),
+    accumulated across every chunk fed to the parser (FWD-14);
+    exceeding it terminates the stream with a single terminal error
+    envelope (`HTTP_413`), the same stream-ends semantics as the other
+    terminal arms. The bounded-time companion is the client's read
+    timeout, which stays armed on the stream client: it bounds
+    upstream *staleness* (no byte for 30 s → terminal), not stream
+    lifetime.
 6. On HTTP error (non-2xx): maps to the declared `ErrorDefinition` by
    HTTP status code (see Error Fidelity below), returns a `CallError`.
 
@@ -205,6 +221,29 @@ connection pooling, keep-alive, TLS, and a retry stack. The shared type is
 both retry and Retry-After are middleware on the stack, and middleware
 requires the `ClientWithMiddleware` wrapper.
 
+The shared `SharedHttpClient` exposes **two derived clients** from one
+config, rebuilt-and-swapped together (the same atomicity rule as the
+config itself):
+
+- **Request client** (`SharedHttpClient::client`) — carries the total
+  request timeout (default 30 s, anchored to the gateway's Once-op
+  deadline). Every request/response forward (`from_openapi`/
+  `from_jsonschema` Queries and Mutations) sends through it.
+- **Stream client** (`SharedHttpClient::stream_client`) — built from
+  the same config with the total request timeout removed and the
+  connect + read timeouts retained (FWD-15). Streaming (SSE)
+  subscription forwards send through it. reqwest 0.13's per-request
+  timeout override can lengthen a client-level total timeout but never
+  clear it (the request-scoped `None` falls back to the client
+  default), so a timeout-free subscription send requires the derived
+  client rather than a request extension. Unbounded *time* per
+  subscription is the contract (alkcall ADR-021); the complementary
+  bounds are the read timeout (staleness) and the total streamed-bytes
+  cap (see the forwarding-handler step 5 above).
+
+Both clients carry the same middleware stack, redirect policy, and TLS
+trust — the total request timeout is the only delta between them.
+
 The middleware stack has two layers:
 
 1. **`RetryTransientMiddleware`** (from `reqwest-retry`) — exponential

docs/architecture/http-server.md

@@ -385,7 +385,16 @@ registry invoke in `tokio::time::timeout`); a hung handler surfaces as
 a `TIMEOUT` error (`504`, `retryable: true`), not an indefinitely-held
 HTTP request. Streaming invokes (`/subscribe`) are unbounded —
 subscriptions are long-lived by contract (alkcall ADR-021 sets
-`deadline: None` for the streaming branch).
+`deadline: None` for the streaming branch). The same time/bytes split
+governs the **outbound** half of an imported subscription:
+`from_openapi`/`from_jsonschema` SSE forwards send through the shared
+HTTP client's stream client, built without the total request timeout
+(FWD-15) so a healthy subscription is not killed at the 30 s outbound
+deadline, while a total streamed-bytes cap
+(`HttpClientConfig::stream_total_byte_cap`, default 1 GiB) bounds the
+bytes a single subscription may pull (FWD-14) and the client's read
+timeout remains the upstream-staleness guard. Unbounded time, bounded
+bytes — on both the gateway and the forwarding side.
 
 ### `/healthz` (raw route)