I see, that can be confusing. The circled area that is shown in the screenshot is level 2 of the tree. But only level 3 contains geometry data.
The intention behind this behavior is to support level of detail: You’d basically try to find a low-detail representation of the geometry that will be shown in level 3, and put this low-detail representation into level 2, and even lower detail in level 1, and even lower detail in level 0.
For level 0 (the whole city), the geometry could probably just be a single polygon of the city outline. On level 3, it will be the buildings that you already have. But I see that it could be hard to find sensible “lower-detail representations” for the steps in-between…
(Maybe this could be some sort of “convex hull” of all the buildings in that area? Or just 2D polygons of the building shapes (not extruded to 3D)? - it’s hard to imagine what these approaches would “look and feel” like when zooming and panning…)
If the intention is to either show no geometry or all geometry (at the highest level of detail), then setting a constant geometric error via the semantic is one (and I think the only) way to achieve that.
But again: This means that when the user zooms in, then it suddenly tries to load all B3DM files at once. The data set here may still be small enough for this to not be too critical, but one should be aware of what it means when this approach is applied with “many” buildings, and the buildings are more complex, with detailed geometry and textures - maybe with 100000 buildings where each building is a several-MB B3DM file…