LLM Fallbacks Break Agent Pipelines

Standard model swapping during rate limits forces production AI to accept corrupted data. A basic fallback router shows 100% completion but drops schema integrity to 0%. The author built a zero-dependency, state-aware recovery layer that catches errors, rebuilds payloads for backup models, and saves agent progress before swapping.