Semantic 3D city models (s3DCM) are a key component of urban digital twins (UDT) because they integrate geometry and semantic information in three-dimensional environments. However, generating geometrically consistent and interoperable models remains challenging due to data heterogeneity and the sensitivity of the reconstruction processes. This research identifies and analyzes the critical factors that influence the quality of s3DCM in the Colombian context of mountainous cities, while also developing a methodological workflow based on the Roofer algorithm for LoD2 reconstruction to generate energy-oriented UDT by integrating building footprints and lidar data. The results show that footprint quality and consistency, together with lidar classification, are the main determinants of 3D reconstruction accuracy. In addition, mountainous terrain and local construction practices require further adjustments of the algorithm parameters that better reflect real-world conditions. Multiple adjustments to the input data and algorithm parameters show that model quality depends on rigorous preprocessing, appropriate geometric validation, and, in mountainous areas, accurate estimation of building base height. Finally, we propose a reproducible workflow for s3DCM generation based on the identified critical factors. The workflow provides a foundation for urban analysis and an energy-oriented UDT implementation in mountainous cities.