How Naive Failover Between Inference Providers Makes Outages Worse

I built a fault-injection rig and ran it for six hours on a DigitalOcean H200 GPU Droplet. Naive and backoff-only failover multiplied backend load 1.48x-3.99x and did not return to baseline latency in 3 of 6 scenarios. A disciplined retry-budget-and-circuit-breaker policy held load near 1x and recovered every time, at the cost of in-fault success as low as 47.8 percent.