Model providers go down. They rate-limit you, they have capacity blips, a specific model gets deprecated, and sometimes a request just fails. If a model call is on a path your users care about, you want a fallback: if the first model errors, try another, ideally from a different provider so a single vendor's bad day does not become yours. The problem is that building that fallback the usual way means owning the differences between providers. Each has its own SDK with its own client setup, its...