Ran a game-day last night, killed our vector DB cluster in us-east for 15 minutes, and learned our autoscaler starved GPUs during failover. How are you testing regional loss for token-streaming services without doubling your footprint — any scalable patterns that keep p95 stable under load spikes?