No documented, reproducible open-source workflow exists for converting CityGML data from a 3D City Database (3DCityDB) instance into functional Indexed 3D Scene Layer (I3S) Scene Layer Packages (SLPK). This study addresses that gap through systematic empirical evaluation of two conversion architectures tested across four datasets spanning single buildings to a district block of 2,571 features, Levels of Detail from LOD2 to LOD4, and multiple schema extensions. The database-driven pathway via 3D Tiles 1.1 fails due to specification incompatibility between toolchain components. The file-based pathway via 3D Tiles 1.0 succeeds reliably, delivering a 565-fold per-unit efficiency gain at district scale, compression ratios up to 11.4:1, and 100% polygon geometry preservation. Semantic retention is schema-selective at building level (0–100%) and reaches complete annihilation at surface and interior levels. Non-polygonal geometry fails silently with 38–62% data loss and no warning. The study contributes the first functional open-source workflow, the Four Thresholds Framework for practitioner decision-making, and the analytical concepts of loss vs. absence and semantic annihilation. The workflow is recommended for individual building visualization targeting I3S Explorer with AAA® or Planning schema at LOD1–2.