Agreed - that trap is very real. The open question for me is what we do when atomic, 5min readable diffs are the right goal but not realistically achievable always. My gut says we need better deterministic signals to reduce noise before human review. Not to replace it.