We introduce CAPA, a parameter-efficient test-time optimization framework for adapting pre-trained 3D foundation models to depth completion from sparse geometric cues. CAPA freezes the model backbone and updates a small set of parameters through parameter-efficient fine-tuning, grounding the model’s geometric prior in scene-specific observations. For videos, sequence-level parameter sharing improves temporal consistency. The method is model-agnostic for ViT-based foundation models and achieves state-of-the-art performance across indoor and outdoor depth-completion settings.