Future 6G networks are expected to support highly heterogeneous and latency-critical services over dynamic radio and compute infrastructures spanning the edge-to-cloud continuum. In this setting, user association can no longer be treated as a purely radio-centric problem, as end-to-end performance depends jointly on radio access conditions and service-specific latency, both of which may vary rapidly over time. Based on the architecture of the EU-funded project ORIGAMI and its Dynamic Fairness and Load Balancing (FLB) use case, this paper studies a service-aware user association problem under non-stationary conditions, where a controller assigns each user to a serving cell while keeping the application-dependent (from transportation, processing, etc) and potential handover (if there is a change in association decisions) delays to a minimum. Importantly, these latencies are not only unknown at the time the user associations are determined but are also assumed to be time-varying and even chosen by an adversary aiming to disrupt the user's quality of experience/service. To solve this problem and address these challenges, we propose a framework based on online meta-learning that combines the decisions of multiple learners (or algorithms, experts), each with different learning rates (as the problem is difficult to tackle) via a meta-learning algorithm that tracks and chooses the best performing one(s) for the experienced conditions. Finally, we establish a theoretical dynamic regret guarantee against a powerful benchmark (i.e., oracle) and validate the approach using real-world data from a top-tier mobile network operator in Europe.