Phase 3 — prefix-cache affinity (8 sessions x 20 turns, growing distinct prefixes)
Per-request pod attributed by vllm:request_success_total +1 (selector-agnostic), sequential sender.
EPP mode:   per-session retention 100.0% (all 8 sessions pinned to one pod for all 20 turns); prefix-cache hit 95.3%
Control RR: per-session retention  46.9% (sessions scatter across 3 pods);                  prefix-cache hit 87.4%
Cross-check: EPP-log 'Request handled' endpoint retention also 100% (earlier run).
Conclusion: Q3 PASS. Affinity is produced by EPP prefix-cache scheduling, not loxilb.
