Achieving stable and accurate imitation of human motions by humanoid robots remains challenging, particularly for tasks requiring firmly standing while executing diverse upper-body movements. While reinforcement learning (RL) offers promise for whole-body motion control, directly tracking upper-body motions often lead to the difficulty in resolving the conflict between stability and similarity. In this paper, we present a RL framework for humanoid robots to stably imitate diverse upper-body motions while standing. Our approach employs a motion retargeting network to translate human motions into humanoid targets. A decoupled RL policy is then trained to control the lower body for standing while tracking upper-body motions. However, large-amplitude motions outside training data may lead to loss of balance. To address this, we introduce an Executable Motion Prior (EMP) network that adjusts potentially unstable actions into more feasible motions while minimizing changes to the intended movement amplitude, enhancing standing robustness without retraining. We evaluate our framework through simulation and real-world tests, demonstrating its practical applicability.