Phase 2 PR 4 — close the 4 open high/medium MXAccess findings from exit-gate-phase-2-final.md. High 1 (ReadAsync subscription-leak on cancel): the one-shot read now wraps subscribe→first-OnDataChange→unsubscribe in try/finally so the per-tag callback is always detached, and if the read installed the underlying MXAccess subscription itself (the prior _addressToHandle key was absent) it tears it down on the way out — no leaked probe item handles when the caller cancels or times out. High 2 (no reconnect loop): MxAccessClient gets a MxAccessClientOptions {AutoReconnect, MonitorInterval=5s, StaleThreshold=60s} + a background MonitorLoopAsync started at first ConnectAsync. The loop wakes every MonitorInterval, checks _lastObservedActivityUtc (bumped by every OnDataChange callback), and if stale probes the proxy with a no-op COM AddItem("$Heartbeat") on the StaPump; if the probe throws or returns false, the loop reconnects-with-replay — Unregister (best-effort), Register, snapshot _addressToHandle.Keys + clear, re-AddItem every previously-active subscription, ConnectionStateChanged events fire for the false→true transition, ReconnectCount bumps. Medium 3 (subscriptions don't push frames back to Proxy): IGalaxyBackend gains OnDataChange/OnAlarmEvent/OnHostStatusChanged events; new IFrameHandler.AttachConnection(FrameWriter) is called per-connection by PipeServer after Hello + the returned IDisposable disposes at connection close; GalaxyFrameHandler.ConnectionSink subscribes the events for the connection lifetime, fire-and-forget pushes them as MessageKind.OnDataChangeNotification / AlarmEvent / RuntimeStatusChange frames through the writer, swallows ObjectDisposedException for the dispose race, and unsubscribes in Dispose to prevent leaked invocation list refs across reconnects. MxAccessGalaxyBackend's existing SubscribeAsync (which previously discarded values via a (_, __) => {} callback) now wires OnTagValueChanged that fans out per-tag value changes to every subscription ID listening (one MXAccess subscription, multi-fan-out — _refToSubs reverse map). UnsubscribeAsync also reverse-walks the map to only call mx.UnsubscribeAsync when the LAST sub for a tag drops. Stub + DbBacked backends declare the events with #pragma warning disable CS0067 because they never raise them but must satisfy the interface (treat-warnings-as-errors would otherwise fail). Medium 4 (WriteValuesAsync doesn't await OnWriteComplete): MxAccessClient.WriteAsync rewritten to return Task<bool> via the v1-style TaskCompletionSource-keyed-by-item-handle pattern in _pendingWrites — adds the TCS before the Write call, awaits it with a configurable timeout (default 5s), removes the TCS in finally, returns true only when OnWriteComplete reported success. MxAccessGalaxyBackend.WriteValuesAsync now reports per-tag Bad_InternalError ("MXAccess runtime reported write failure") when the bool returns false, instead of false-positive Good. PipeServer's IFrameHandler interface adds the AttachConnection(FrameWriter):IDisposable method + a public NoopAttachment nested class (net48 doesn't support default interface methods so the empty-attach is exposed for stub implementations). StubFrameHandler returns IFrameHandler.NoopAttachment.Instance. RunOneConnectionAsync calls AttachConnection after HelloAck and usings the returned disposable so it disposes at the connection scope's finally. ConnectionStateChanged event added on MxAccessClient (caller-facing diagnostics for false→true reconnect transitions). docs/v2/implementation/pr-4-body.md is the Gitea web-UI paste-in for opening PR 4 once pushed; includes 2 new low-priority adversarial findings (probe item-handle leak; replay-loop silently swallows per-subscription failures) flagged as follow-ups not PR 4 blockers. Full solution 460 pass / 7 skip (E2E on admin shell) / 1 pre-existing Phase 0 baseline. No regressions vs PR 2's baseline.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

src/ZB.MOM.WW.OtOpcUa.Driver.Galaxy.Host/Ipc/GalaxyFrameHandler.cs

@@ -99,9 +99,64 @@ public sealed class GalaxyFrameHandler(IGalaxyBackend backend, ILogger logger) :
         }
     }
 
+    /// <summary>
+    ///     Subscribes the backend's server-pushed events for the lifetime of the connection.
+    ///     The returned disposable unsubscribes when the connection closes — without it the
+    ///     backend's static event invocation list would accumulate dead writer references and
+    ///     leak memory + raise <see cref="ObjectDisposedException"/> on every push.
+    /// </summary>
+    public IDisposable AttachConnection(FrameWriter writer)
+    {
+        var sink = new ConnectionSink(backend, writer, logger);
+        sink.Attach();
+        return sink;
+    }
+
     private static T Deserialize<T>(byte[] body) => MessagePackSerializer.Deserialize<T>(body);
 
     private static Task SendErrorAsync(FrameWriter writer, string code, string message, CancellationToken ct)
         => writer.WriteAsync(MessageKind.ErrorResponse,
             new ErrorResponse { Code = code, Message = message }, ct);
+
+    private sealed class ConnectionSink : IDisposable
+    {
+        private readonly IGalaxyBackend _backend;
+        private readonly FrameWriter _writer;
+        private readonly ILogger _logger;
+        private EventHandler<OnDataChangeNotification>? _onData;
+        private EventHandler<GalaxyAlarmEvent>? _onAlarm;
+        private EventHandler<HostConnectivityStatus>? _onHost;
+
+        public ConnectionSink(IGalaxyBackend backend, FrameWriter writer, ILogger logger)
+        {
+            _backend = backend; _writer = writer; _logger = logger;
+        }
+
+        public void Attach()
+        {
+            _onData = (_, e) => Push(MessageKind.OnDataChangeNotification, e);
+            _onAlarm = (_, e) => Push(MessageKind.AlarmEvent, e);
+            _onHost = (_, e) => Push(MessageKind.RuntimeStatusChange,
+                new RuntimeStatusChangeNotification { Status = e });
+            _backend.OnDataChange += _onData;
+            _backend.OnAlarmEvent += _onAlarm;
+            _backend.OnHostStatusChanged += _onHost;
+        }
+
+        private void Push<T>(MessageKind kind, T payload)
+        {
+            // Fire-and-forget — pushes can race with disposal of the writer. We swallow
+            // ObjectDisposedException because the dispose path will detach this sink shortly.
+            try { _writer.WriteAsync(kind, payload, CancellationToken.None).GetAwaiter().GetResult(); }
+            catch (ObjectDisposedException) { }
+            catch (Exception ex) { _logger.Warning(ex, "ConnectionSink push failed for {Kind}", kind); }
+        }
+
+        public void Dispose()
+        {
+            if (_onData is not null) _backend.OnDataChange -= _onData;
+            if (_onAlarm is not null) _backend.OnAlarmEvent -= _onAlarm;
+            if (_onHost is not null) _backend.OnHostStatusChanged -= _onHost;
+        }
+    }
 }