Design ChatGPT streaming with resumable responses and long conversations

You have already sketched a ChatGPT-style chat product. Now go deeper on three follow-ups: how replies stream to the browser (SSE or something else, and why), how a reply survives the client disconnecting mid-stream and resumes on reconnect or on a second device without losing or repeating tokens and without regenerating, and how conversations far longer than the model's context window keep answering well and fast. About 100M weekly users and 5k new replies per second at peak.