Fallback model should add evidence, not just a new opinion

A note on fallback model rules for coding work: repeated failure, missing capability, and high-risk review.

A fallback model rule is useful only if the fallback changes the evidence situation. If a coding agent fails twice in the same shape, another model can help. But the prompt should ask for a different role: inspect the patch, find missing tests, check official docs, or review a specific risk. Asking for the same broad solution again often produces a second confident answer with the same blind spot. The fallback should reduce uncertainty, not just make the workflow feel busy.