A financial services company runs a stateless order-matching service on an EC2 Auto Scaling group behind an Application Load Balancer. During market-open bursts, demand spikes within seconds. The application AMI is large and requires roughly 6 minutes to boot, run configuration scripts, and warm its JVM cache before it can pass ALB health checks. Scale-out driven by target-tracking on request count is too slow, causing dropped orders during the first few minutes of each spike. The team wants to reduce scale-out latency without permanently running (and paying for full price on) many idle instances. Which approach best meets these requirements?