When two or more SNMPv3 hosts polled by the same Zabbix server or proxy present an identical snmpEngineID, polling becomes unreliable or fails entirely. That is because the underlying net-snmp USM layer caches engine boots/time and localized keys per engineID. When two different IP endpoints response with the same engineID but different engineBoots/engineTime values, the library treats responses as out-of-time-window or spoofed and discards them.

Please conside one or more of the following:

1. Isolate SNMPv3 USM state (engineID/boots/time cache and localized keys) per host interface rather than per poller process, so that duplicate engineIDs on different IP:port endpoints do not collide. Since Zabbix 7.0 introduced its own asynchronous SNMP polling implementation, this may now be feasible without being fully constrained by net-snmp's shared LCD behavior.
2. Alternatively, provide a per-interface or per-host option (e.g. a macro or interface flag) to disable engineID caching / force engine rediscovery per request for affected hosts, accepting the minor overhead.
3. If neither is feasible short-term: add anti-affinity rules to proxy groups (e.g. "these hosts must never be assigned to the same proxy"), so HA pairs can at least coexist with automatic load balancing.