Open-set gait recognition presents a critical challenge for real-world identity authentication systems, requiring simultaneous identification of known users and detection of unknown users under practical deployment conditions. However, in practical Wi-Fi sensing environments, signal noise, clothing variation, and multipath interference often blur the boundary between known and unknown classes, making traditional closed-set methods inadequate. To address this challenge, GUARD is proposed as a unified open-set and closed-set gait recognition framework based on feature reconstruction. The core idea is that known-class samples can be accurately reconstructed under matched label conditions, while unknown samples yield significantly higher reconstruction errors due to label mismatch, thereby providing a discriminative signal for open-set recognition. To enhance the stability and discriminability of features, GUARD integrates a Global Temporal Attention (GTA) mechanism to capture long-range temporal dependencies, and introduces a Pseudo-Gaussian Enhanced Self-Attention (PGESA) module that models dynamic attention distributions via Gaussian approximation, enabling selective emphasis on salient temporal features while effectively suppressing background noise. Additionally, a feature extractor locking strategy is employed to freeze identity-relevant representations once closed-set performance is optimized, preventing degradation during open-set training. Experimental results show that GUARD achieves over 20% improvement in open-set recognition rate, while maintaining approximately 95% closed-set accuracy, demonstrating superior robustness and generalization in complex sensing environments.