Without this, Ninja requires a two-step process because it doesn't know
what generates version.h it sees from its depfile support and when it
changes in one run, it has already pruned the dependent nodes from its
internal graph and requires a second run to catch it.