This is exactly the gap I'm worried about. human review still matters, but linear reading breaks down once the diff is mostly machine-generated noise. Summarizing what actually changed before reading feels like the only way to keep reviews sustainable.