Reservoir computing (RC) provides an efficient framework for processing temporal information. However, spin–orbit–torque (SOT)-based physical RC systems are fundamentally limited by a trade-off between stochasticity in device responses and the richness of reservoir dynamics. Here, we address this challenge through a device–architecture co-design strategy. At the device level, we demonstrate that edge modulation of Hall-bar SOT devices effectively suppresses stochasticity in magnetization switching by regulating domain-wall depinning dynamics, thereby improving the stability of reservoir states. At the architecture level, we introduce a multi-port configuration that integrates physically distinct nonlinear responses from different terminals, enhancing the dimensionality and dynamic richness of the reservoir. Within a time-division multiplexed RC framework, the reduced stochasticity achieved by edge modulation leads to improved accuracy in speech recognition and waveform classification tasks. Furthermore, in the Mackey–Glass chaotic time-series prediction task, the multi-port reservoir achieves longer effective prediction horizons and lower normalized root-mean-square error compared with single-device reservoirs. These results demonstrate that jointly optimizing device-level stability and architecture-level reservoir diversity enables simultaneous enhancement of state reliability and representational capacity, providing an effective route toward high-performance spintronic physical RC systems.