fix: use dedicated _shutting_down flag for cancel-time checkpoint restore

The previous guard used self._running, which is False both before run()
starts and during shutdown. Tests and process_direct() dispatch messages
with _running=False, so cancelled turns there silently skipped the
checkpoint restore and lost partial context. Introduce _shutting_down,
set only by stop(), and gate the restore on it.

nanobot/agent/loop.py

@@ -378,6 +378,7 @@ class AgentLoop:
         )
         self._unified_session = unified_session
         self._running = False
+        self._shutting_down = False
         self._mcp_servers = mcp_servers or {}
         self._mcp_stacks: dict[str, MCPConnection] = {}
         self._mcp_connecting = False
@@ -1224,11 +1225,11 @@ class AgentLoop:
                     # it into session history now makes it visible in the
                     # next conversation turn.
                     #
-                    # During gateway shutdown (self._running is False), skip
-                    # this so the pending markers survive and
+                    # During gateway shutdown (self._shutting_down is True),
+                    # skip this so the pending markers survive and
                     # recover_stale_sessions() can notify the user and
                     # re-trigger the interrupted turn on next startup.
-                    if self._running:
+                    if not self._shutting_down:
                         try:
                             key = self._effective_session_key(msg)
                             session = self.sessions.get_or_create(key)
@@ -1318,6 +1319,7 @@ class AgentLoop:
     def stop(self) -> None:
         """Stop the agent loop."""
         self._running = False
+        self._shutting_down = True
         logger.info("Agent loop stopping")
 
     async def _process_message(