Retry loops that invent phantom wait
A silent client backoff can look like a slow API when you only read the average.
On web APIs, a client that retries without jitter can multiply wait for a single user action. The average climbs; the server never asked those users to wait twice.
Before rewriting the endpoint, walk the path while watching live timings and error rates. Note client-only retries versus server confirmations. Check whether staging shares production clocks and labels.
Latency Budget Mapping stays narrow on purpose: one journey, deep reconciliation, a checklist you can rerun after every release so the phantom wait does not return unnoticed.