"Just add a fallback model" is the kind of advice that sounds complete until you've actually shipped it. A single backup model handles one failure mode (the primary is down) and quietly ignores the other four that show up in production: slow responses, malformed output, rate limits, and a fallback model that doesn't accept the same context shape as your primary. Here's the pattern breakdown that actually holds up. The five patterns, and when each one is the right call P...