Here's a sequence that shows the race:
Thread-1 (Users of TThreadedServer): Calls TThreadedServer::stop(), which calls interruptChildren and initiates the tearing down of client connections.
Thread-2: In TServerFramework::serve(), broke out of accept, and now blocks in TThreadedServer::serve() waiting to drain all the clients.
Thread-3 (The connected client thread created by TThreadedServer): In disposeConnectedClient, running because the server is shutting down and the shared_ptr specified this function to be the cleanup function for the client. This thread just returned from onClientDisconnected and now context switches.
Thread-2: TThreadedServer::serve() is notified that all of the clients have disconnected and completes.
Thread-1: Joins on Thread-2 and destroys the server object because it is done.
Thread-3: Finally gets a chance to run, but now encounters undefined behavior because it is still executing a member function of an object that has been destroyed.
You can force this race in action if you put sleep(1) before onClientDisconnected() in disposeConnectedClient