Issue #14: implement event streaming and backpressure

docs/gateway-process-design.md

@@ -206,13 +206,23 @@ accounting and a clear fan-out policy.
 Behavior:
 
 1. Validate session id and authorize event access.
-2. Attach a stream cursor to the session event channel.
-3. Send events in worker sequence order.
-4. Stop on client cancellation, session close, or session fault.
-5. Emit a terminal status when the session faults if gRPC status alone cannot
+2. Attach the single active subscriber lease for the session.
+3. Read worker events into a bounded public stream queue.
+4. Send events in worker sequence order.
+5. Stop on client cancellation, session close, or session fault.
+6. Emit a terminal status when the session faults if gRPC status alone cannot
    preserve the required details.
 
-The gateway must not reorder events from one worker.
+`EventStreamService` owns subscriber tracking and public stream backpressure.
+The default policy allows one active subscriber per session. A second subscriber
+is rejected with `EventSubscriberAlreadyActive`. Stream cancellation releases
+the subscriber lease so a later stream can attach to the session.
+
+The gateway must not reorder events from one worker. `EventStreamService` writes
+mapped events to a bounded first-in, first-out queue and faults the session with
+`EventQueueOverflow` if the queue fills. The gateway does not synthesize
+`OperationComplete`; it forwards that family only when the worker reports a
+native MXAccess `OperationComplete` event.
 
 ## Web Dashboard
 
@@ -584,7 +594,8 @@ worker MXAccess event
   -> worker outbound event queue
   -> worker pipe writer
   -> gateway read loop
-  -> session event channel
+  -> worker client event queue
+  -> EventStreamService bounded stream queue
   -> gRPC StreamEvents
 ```
 
@@ -598,13 +609,15 @@ The gateway should record:
 
 Default backpressure policy for parity testing should be fail-fast:
 
-1. If the session event channel fills, fault the session.
+1. If the worker client event queue fills, fault the worker client.
+2. If the public stream queue fills, fault the gateway session.
 2. Preserve the overflow details in logs and metrics.
 3. Do not silently drop data-change events.
 
-Do not set a production event-rate target before measurement. Emit event rate,
-queue depth, stream send latency, and overflow metrics. Later production modes
-may support explicit coalescing by item handle as an opt-in behavior.
+Do not set a production event-rate target before measurement. `GatewayMetrics`
+records received event counts by family, queue depth, stream disconnects, and
+overflow counts. Later production modes may support explicit coalescing by item
+handle as an opt-in behavior.
 
 The gateway should not synthesize `OperationComplete` from write completion,
 command replies, ASB completion queues, or completion-only status frames. Forward