fix(gateway): offload evaluate_after_turn to thread executor

evaluate_after_turn() calls judge_goal() which makes a synchronous
HTTP request to the auxiliary LLM. Running it on the event-loop
thread blocks Discord heartbeats for 10-40s, causing connection
flaps and gateway instability.

Offload to the default thread-pool executor so the event loop
stays responsive during evaluation.

gateway/run.py

@@ -19424,10 +19424,21 @@ class GatewayRunner(GatewayAuthorizationMixin, GatewayKanbanWatchersMixin, Gatew
         except Exception:
             _bg_procs = None
 
-        decision = mgr.evaluate_after_turn(
-            final_response or "",
-            user_initiated=True,
-            background_processes=_bg_procs,
+        # evaluate_after_turn calls judge_goal() which makes a synchronous
+        # HTTP request to the auxiliary LLM.  Running it on the event-loop
+        # thread would block Discord heartbeats for 10-40 s and cause
+        # connection flaps, so we offload it to a thread-pool executor.
+        # _run_in_executor_with_context (not bare run_in_executor): the
+        # profile secret scope and auxiliary runtime context are contextvars,
+        # and a default-executor hop would drop them — aux-client provider
+        # resolution would then read credentials unscoped and fail under
+        # multiplexing (same pattern as compression in slash_commands.py).
+        decision = await self._run_in_executor_with_context(
+            lambda: mgr.evaluate_after_turn(
+                final_response or "",
+                user_initiated=True,
+                background_processes=_bg_procs,
+            ),
         )
         msg = decision.get("message") or ""