Industrial digital twins operating across geographically distributed manufacturing facilities must simultaneously contend with concept drift, workload surges, and infrastructure failures, conditions that expose a fundamental gap in static, periodically retrained architectures: they treat adaptation, resource provisioning, and resilience as independent, offline concerns rather than as jointly managed
operational properties. This paper addresses that gap with a cloud-native systems architecture that closes the adaptation-provisioning resilience control loop in a single, continuously operating coordination plane. The architecture integrates four tightly coupled infrastructure mechanisms: stream-resident Stream Adaptation Agents (SAA) that perform partition-level drift monitoring and quorum gated orchestration; a Shared Adaptive Inference Layer that maintains a streaming ensemble model updated in-place without service interruption.