To address the limitations of traditional analytical modeling in capturing complex surface topographies, this paper presents comprehensive research on the error sensitivity mechanism, loaded tooth contact analysis (LTCA), and load-bearing contact characteristics of curved face gears based on high-precision point cloud modeling. The primary objectives are threefold: (1) to establish a high-fidelity topological reconstruction framework using Non-Uniform Rational B-Splines (NURBS) to bridge the gap between discrete data and finite element analysis (FEA); (2) to reveal the inherent mechanical response and sensitivity mechanism to spatial installation misalignments; and (3) to evaluate the contact performance and transmission error fluctuations under operational loads. Specifically, an analytical discretization method is proposed for point cloud generation, followed by a dual-path validation system integrating “rigid tooth contact analysis (TCA)” and “loaded FEA”. The results demonstrate that the proposed reconstruction achieves a superior accuracy with a Root Mean Square Error (RMSE) of 2.2 × 10⁻³ mm. Furthermore, shaft angle error is identified as the dominant sensitivity factor affecting transmission smoothness and edge contact, exerting a more significant influence than offset and axial errors. Compared with existing research on arc-tooth and helical face gears, this work provides a more robust closed-loop verification for curved profiles, revealing that material elastic deformation increases transmission error amplitude by 10.1% to 17.2%. These insights offer a theoretical reference for the high-precision assembly and tolerance allocation of helicopter transmission systems.