In internet of things (IoT) networks, resource-constrained devices can offload data to mobile edge computing (MEC) servers for processing. Nevertheless, offloading incurs additional energy consumption, making energy-efficient offloading a critical issue. Furthermore, the mismatch of transmission and computation traffic rates further exacerbates energy inefficiency. Additionally, ensuring quality of service (QoS) requires not only meeting the per-slot processed data volume thresholds of devices but also satisfying their long-term energy consumption limits imposed by limited battery capacity. The coexistence of these heterogeneous constraints renders algorithm design more challenging. In this paper, we propose a safe deep reinforcement learning (Safe-DRL)-based resource allocation strategy for traffic matching of transmission and computation in QoS-guaranteed MEC-IoT networks, aiming to maximize energy efficiency. Specifically, we formulate an energy efficiency maximization problem that captures the interdependence among transmission, buffering, and computation traffic, subject to both data processing and energy consumption constraints. To reduce complexity, closed-form solutions for a subset of variables are derived via mathematical analysis. Subsequently, we develop a Safe-DRL algorithm, termed action projection augmented Lagrangian soft actor-critic (APAL-SAC), which integrates a penalty-based action projection mechanism to enforce per-slot constraints and a Lagrangian dual method to ensure long-term constraint satisfaction. Simulation results demonstrate the effectiveness of APAL-SAC.