Accurate puck tracking from broadcast ice hockey video remains challenging due to the puck’s small size, frequent occlusions, motion blur, and rapid camera motion. Recent advances in puck detection have improved per-frame localization, but robust tracking is often undermined by geometric instability arising from image-space motion modelling. We argue that representation stability, rather than tracker complexity, is the primary bottleneck in broadcast puck tracking. We propose a three-stage pipeline that combines reliable puck detections with a temporally stable planar representation of the rink surface. Using monocular 3D reconstruction, we estimate a single, consistent rink plane across an entire broadcast sequence and map image-space puck detections into this stable planar coordinate system. Unlike traditional rink registration approaches, the proposed representation prioritizes temporal and geometric consistency over absolute metric scale or orientation, enabling smooth and coherent puck-motion modelling without relying on per-frame homographies or camera calibration. Geometry-aware motion estimation is then performed directly in the planar space using simple sequence models. Experimental results demonstrate that enforcing temporal geometric consistency enables robust puck tracking with simple models, particularly under severe occlusion, and provides a strong alternative to complex image-space tracking pipelines.