Millimeter-wave (mmWave) radar enables non-contact human identification by exploiting gait micro-Doppler signatures without capturing visual imagery. Nevertheless, practical deployment encounters several critical challenges: strong torso echoes that obscure faint limb-motion features, channel-dependent responses caused by unconstrained walking styles, and computational limitations on edge hardware. This paper proposes a lightweight gait-based human identification method for multiple-input multiple-output (MIMO) radar systems, which integrates adaptive Gamma correction, multi-transmit-channel differential encoding, and an asymmetric-pooling convolutional recurrent neural network (Asym-CRNN). Specifically, the differences across the three transmit-channel responses of the three-transmit four-receive (3T4R) radar are encoded into a three-channel spatiotemporal representation, and the compact network retains temporal resolution to facilitate efficient gait discrimination. Evaluated on a self-collected five-subject dataset with acquisition-sequence-level train-test separation, the proposed method achieves closed-set accuracies of 89.62% and 99.52% under 1 s and 6 s observation windows, respectively. For open-set scenarios, the method attains a registered-versus-unregistered area under the receiver operating characteristic curve (AUROC) of 0.9869. With a maximum softmax probability (MSP) threshold of 0.85, the method rejects 91.88% of unregistered human samples and all tested pet-interference samples. Furthermore, the model achieves an accuracy of 86.98% using only single-channel IDRad inputs. The exported model requires merely 4.32 MB of storage. Deployed on a Raspberry Pi 5 platform, the acquisition-processing pipeline runs using only two CPU cores with CPU-only execution: one core continuously receives and buffers raw radar data, whereas the other processes previously acquired observation windows. The per-window processing latency reaches 715.7 ms for the 1 s window and 4366.2 ms for the 6 s window, including network inference times of 6.7 ms and 15.2 ms, respectively. The non-causal bidirectional long short-term memory (Bi-LSTM) generates one identity decision once each observation window is fully collected. These results verify the feasibility of the proposed method for privacy-preserving indoor monitoring at small scales on resource-constrained edge devices.