Truncated conversation history breaking agent state

The agent forgot user preferences after turn ten because I sliced the message list by token count instead of message count. This split a multi-part tool response in half and corrupted the state. Now I enforce complete message boundaries before truncating any history.