Retry amplification from slow upstream

At 14:02 your payments API's p99 latency to the downstream fraud-scoring service jumps from 80ms to 9s, then the fraud service starts returning 503s. Dashboards: fraud-service CPU pinned at 100%, its inbound request rate is 6x normal while your own user traffic is flat, and the fraud service's own upstream (a feature-store Redis) shows a 200ms blip ending at 14:01. Your gateway's retry metric shows attempts climbing from 1.1 to 4.0 per logical request. The fraud team deployed nothing today. How do you triage and mitigate?