Federated Learning (FL) enables collaborative model training across distributed IoT and edge devices while preserving data locality, making it attractive for privacy-sensitive and resource-constrained environments. However, the integration of Differential Privacy (DP) introduces a critical trade-off between privacy guaranties and model utility, which is further intensified by client heterogeneity, non-IID data distributions, and irregular participation. Existing DP-enabled FL approaches typically apply uniform noise budgets and static aggregation strategies, overlooking differences in client uncertainty, reliability, and contribution. In this work, we address this limitation by proposing a liminality-aware federated learning framework that adaptively assigns privacy noise and aggregation weights based on client-level uncertainty and participation behavior. Liminality is defined as a lightweight epistemic uncertainty measure derived from softmax entropy and prediction confidence, which requires no access to raw data and minimal additional computation. The framework combines uncertainty and behavior-aware signals to dynamically redistribute learning responsibility under DP constraints. The proposed approach is evaluated in highly heterogeneous synthetic FL settings with severe non-IID label and data size distributions across 80 clients. Experimental results show that the liminality-awareness strategy can improve the average accuracy by up to 7% over uniform DP baselines, while also improving fairness for lower-performing client groups. These findings demonstrate that liminality-aware aggregation provides a practical and effective solution to balance privacy, utility, and fairness in federated IoT systems.