Every weekday at 09:00, a training portal receives a sharp traffic spike. New instances take ten minutes to initialize, and reactive scaling causes slow responses at the start of each session. Which operational change most directly prevents the predictable capacity gap?

1. Create scheduled scaling actions that raise capacity before 09:00 and lower it after the training window.
2. Increase the health-check grace period and leave the instance count unchanged.
3. Shorten the scaling cooldown but start all additional launches only after the 09:00 CPU spike.
4. Use only a CPU target-tracking policy with a target of 95%.