Phase 2 forced-bias (EPP pool restricted to 1 pod via InferencePool selector {vrfy-pin:A})
A) loxilb sel=0 (RR), 60 req  -> loxilb dest 60/60 = 198.51.100.11 ; EPP handled {198.51.100.11:60}
B) loxilb sel=8 (CHWBL), 40 req -> loxilb dest 40/40 = 198.51.100.11 ; EPP handled {198.51.100.11:40}
Conclusion: loxilb follows EPP decision 100%; does not override with its own selector. Q2 PASS.
