Developing humanoid robots that can emulate the versatility, agility, and robustness of human movement in complex, unstructured environments has long been a central pursuit in robotics research [1, 2, 3, 4, 5, 6]. Achieving this vision requires not only the ability to execute diverse motor skills, but also the capacity to maintain balance under challenging conditions. Studies in neuroscience and motor control suggest that human balance relies on intricate sensorimotor loops involving the vestibular system, proprioception, and high-level planning [7, 8], making it a particularly demanding aspect of motor control to replicate in robotics. This difficulty is exemplified by the Swallow Balance task shown in Figure 1, in which a humanoid must maintain stability in an extreme single-legged pose with the upper body extended horizontally. Such movements require full-body coordination, precise control of the center of mass, and robustness to perturbations--highlighting the demanding nature of humanoid balance. In recent work on learning-based humanoid control [4, 5, 9, 10, 11, 12, 6], a common approach for enabling humanoids to perform diverse motions is to train a control policy to track reference poses.