Low-slow-small (LSS) target recognition in low-altitude radar surveillance remains challenging because radar echoes are often affected by background clutter, weak micro-Doppler components, and aspect-dependent variations. In addition, practical radar applications require models that can preserve discriminative radar information while maintaining compact model size and efficient inference under resource-constrained conditions. To address these requirements, we developed a deployment-oriented lightweight multi-task framework for K-band frequency-modulated continuous-wave (FMCW) radar to jointly perform target-type and aspect-angle classification. The proposed input representation was built upon the complex in-phase/quadrature short-time Fourier transform (I/Q-STFT) domain and organized the real part, imaginary part, and normalized magnitude map into a task-oriented three-channel input representation. This design aimed to preserve complementary phase-related and energy-distribution cues for dual-label radar recognition. Based on this representation, a compact MobileViT-style backbone was adapted with large separable kernel attention, ConvNeXtV2 blocks, and linearly scheduled stochastic depth to improve long-range time-frequency modeling, intermediate feature encoding, and robustness under joint optimization. Design-choice and sensitivity analyses were conducted to examine the effects of task-weight selection, input-channel composition, and drop rate. Experimental results showed that the proposed method achieved 99.90% accuracy for target-type recognition and 93.37% accuracy for aspect-angle recognition, while maintaining a compact model size of 2.230 M parameters and an inference throughput of 245.3 samples per second. These results demonstrated that the proposed framework provided an effective trade-off between recognition performance and deployment efficiency for resource-constrained K-band radar applications.