fix: release the WebSocket listener even when CloseAllClients throws (fixes #141)

[1shuqing]

fix: release the WebSocket listener even when CloseAllClients throws

Fixes the root cause behind #141 (Failed to start WebSocket server: Port <PORT> is already in use,
cleared only by restarting Unity).

Target: CoderGamester/mcp-unity @ 382a43a (v1.5.0)
Patch: mcp-unity-port-leak-fix.patch (single file: Editor/UnityBridge/McpUnityServer.cs)

---

Root cause

McpUnityServer.StopServer() closes clients and the listener inside one try block, and then
clears the field in finally:

try
{
    CloseAllClients(closeCode ?? 1000, closeReason ?? "Server stopping");   // ← can throw
    _webSocketServer?.Stop();                                              // ← then never runs
}
catch (Exception ex) { McpLogger.LogError(...); }
finally
{
    _webSocketServer = null;                                               // ← reference dropped
    _socketOwner = null;
    Clients.Clear();
}

CloseAllClients walks live sessions and calls Close(...) on each one. When a client is in a
half-closed state (CLOSE_WAIT — exactly what happens when the MCP client times out and drops the
socket during a reload) it throws. The exception is swallowed, so:

1. WebSocketServer.Stop() is never called → the underlying TcpListener stays bound.
2. finally drops the only reference to the server object.
3. The domain then unloads; finalizers do not run for objects in an unloaded AppDomain, so the
   socket handle is only released when the editor process exits.

From that point the port is permanently "alive but unserviced": TCP connects succeed (kernel backlog)
while every WebSocket handshake hangs, and every later bind fails with AddressAlreadyInUse
(WSAEACCES in some Windows states). Only restarting Unity clears it — the behaviour reported in #141.

The liveness probe added for #141 correctly detects this state, but as its own remarks state, it
deliberately does not repair it; the repair has to happen before the socket is orphaned.

Evidence

Instrumented with file-based lifecycle logging (Debug.Log output is dropped during domain reload):

beforeAssemblyReload: instance=True, owner=True, ownerListening=True
StopServer enter: server=True, listening=True
ForceCloseListener: type=TcpListener, wasBound=False      <-- Stop() had been skipped

and, on the next start attempt, ForceCloseAllKnownListeners: known=1, closed=0 with the port still
held by the process — i.e. the bound socket was no longer reachable from any managed object.

Changes

1. StopServer() — close the listener unconditionally. Client shutdown, server.Stop() and the
   explicit listener close each get their own try/catch; the listener is always released before
   the fields are cleared. This is the actual fix.
2. ForceCloseListener() — closes the listener even if Stop() was a no-op or threw, by taking
   WebSocketServer._listener (TcpListener) via reflection and calling Stop() + Socket.Close().
   Also called from CleanupFailedStart().
3. _socketOwner + OnBeforeAssemblyReload — the reload hook no longer depends on the singleton
   being alive; if the singleton is gone it closes the recorded socket directly.
4. _knownServers + ForceCloseAllKnownListeners() — every server created in the domain is
   tracked, so a start failure can reclaim a listener that an earlier failed start left behind.
5. Retry budget — DelayedStartMaxAttempts 10 → 150 with longer tail delays (≈15 min instead of
   ~32 s). A listener held by a just-killed editor process can take minutes to disappear on Windows;
   the server now waits for it instead of giving up.
6. SocketError.AccessDenied is retried too — Windows reports WSAEACCES rather than
   AddressAlreadyInUse when the port is held by a socket created with SO_EXCLUSIVEADDRUSE.
7. Optional diagnostics — lifecycle trace to Temp/mcp_bridge_diag.log. Reviewers may drop this;
   it is independent of the fix. WebSocketServer.ReuseAddress was evaluated and deliberately not
   enabled: it produces WSAEACCES, and new connections can be delivered to the stale listener.

Verification (Unity 6.3 LTS / 6000.3.23f1, Windows 11)

Scenario	Before	After
Script recompile (domain reload)	port leaked, MCP dead until editor restart	beforeAssemblyReload → StopServer → rebind 1.2 s later ✅
Enter Play Mode	port leaked	rebind 4 s later ✅
Exit Play Mode	port leaked	bridge stays up ✅
Editor process restarted	port unusable for minutes, no recovery	old socket released ≈3 min later, retry binds automatically ✅

How to test

1. Apply the patch, open Unity with the package installed (AutoStartServer: true).
2. Confirm the MCP client connects.
3. Trigger a script recompile (recompile_scripts or just edit a script) and immediately retry an MCP
   call — it must succeed within a few seconds.
4. Enter and exit Play Mode, retry — it must still succeed.
5. netstat -ano | findstr 8090 should show exactly one LISTENING socket owned by the editor, and
   Temp/mcp_bridge_diag.log (if diagnostics kept) should show START OK: ... listenerBound=True
   after each reload.