Neural decoding of spatial navigation trajectories based on neural data is an important research direction. In practical scenarios, such tasks encounter substantial challenges due to cross-domain distribution shifts, particularly when dealing with limited training samples or unseen behavioral states. Currently, most distribution shift task is focused on classification rather than regression in neural decoding. To address this gap, we attempt to define a task-regression label distribution shift issue that commonly arises and then propose a Dual-Phase Adaptation framework with Pre-balanced training and Test-Time Learning (DPA-PTTL) as a solution. This framework operates through two synergistic phases: (1) Pre-balanced Training: We mitigate intra-domain sample imbalance in the source domain by employing SMOGN-based synthetic minority oversampling, thereby generating neuro-behaviorally consistent trajectories while preserving motion continuity; (2) Test-Time Learning: we develop an unsupervised neural trajectory validator (e.g., unsupervised algorithm) that identifies cross-domain anomaly samples by analyzing the consistency of movement trends between pseudo-behavioral outputs (or pseudo-label metric). This dual mechanism ensures simultaneous enhancement of source-domain representativeness and cross-domain adaptability, achieving harmonization of sample-space distributions without requiring target-domain labels. Experimental results on an existing hippocampal dataset with long-range attributes demonstrate that the proposed scheme can mitigate the label distribution shift in decoding positions and improve performance by up to 4.2% compared with similar methods.