Several threads here describe the same thing: a Realtime agent on a phone speaker triggers speech_started on its own voice and answers itself, even with the platform’s echo cancellation on and a mic gate. This is what we measured in a production React Native…