Bipedal robots can navigate complex, unstructured environments, but their low inherent stability and sensitivity to model errors make robust locomotion across varied terrains challenging. In this work, we present a unified end-to-end deep reinforcement learning framework designed to enable terrain-adaptive bipedal locomotion. Our approach builds upon a hybrid internal model and contrastive learning to implicitly infer both environmental context and proprioceptive state. To enhance short-horizon adaptability, the policy receives additional historical observations, improving responsiveness to sudden terrain changes. Furthermore, we propose a human-inspired stepping strategy reward that encourages high leg lifts for traversing elevation differences and promotes passive ankle compliance on rugged terrain. Extensive simulation results demonstrate that our method improves dynamic adaptability, reduces energy consumption, and enhances terrain generalization, enabling robust locomotion over slopes, uneven terrain, and continuous stairs.