Measure first is the advice I will actually use. Separating queue time, setup time, and run time before changing anything stops the usual guess about which step is slow. The restore-keys detail also stood out, because a cache miss that falls back to the closest older cache is easy to skip and expensive to lose. The concurrency group tip fits the AI written code problem best, since two pushes an hour apart should not both run to the end when only the second one matters.