HTTP/2 head-of-line blocking

After enabling HTTP/2 between an edge proxy and an upstream service (replacing many HTTP/1.1 connections with a single multiplexed H2 connection per proxy), p99 latency for fast endpoints regressed badly even though throughput and average latency improved. Dashboards show fast 'GET config' requests (normally 5ms) sometimes spiking to hundreds of ms, and the spikes correlate in time with a few slow, large streaming responses on the same upstream. The upstream's own per-request processing time for the fast endpoints hasn't changed. Connection count to the upstream dropped sharply after the H2 rollout. Walk through triage and the fix.