Mid-stream failover made my chat API answer the same prompt twice — switch models before the first token or not at all