Timeout-induced retry amplification cascade

A config push lowers the client-side request timeout on service A (calling downstream B) from 5s to 800ms, intending to 'fail faster.' B's p99 is normally 600ms. After the push, A's error rate jumps, B's CPU and request rate both climb sharply, B's p99 balloons past 2s, and the whole A to B path degrades into near-total failure even though B was healthy before. Dashboards: B's inbound request rate is ~3x the rate A's users actually generate; A's retry counter is spiking. A retries failed/timed-out calls up to 3x with no backoff. How do you triage and mitigate?