The Point Size property is a multiplier applied to the auto-calculated sprite size. The algorithm responsible for calculation uses an estimated local density of the cloud to find the best value to fill the gaps. However, since you set the workaround effectively disables the LOD system, the algorithm cannot do its job correctly - this results in incorrect sizes.
The duplicate handling is only relevant when LODs are generated (it essentially prevents multiple LODs storing points in the same spot), but as with 1. you have disabled the LOD, hence the settings should have no impact on the process. The missing sections, you mentioned, could be a result of the base LOD fighting with itself for the points.
Elevation applies a linear interpolation of colors based on the point’s Z coordinate relative to the Z component of your point cloud’s bounds. If you set those super high, the Z differences between points will be too small to notice
No “correct” solution at this time. Most of the algorithms used are heavily optimized to handle huge, static datasets. We have had some discussion regarding the implementation of something like a dynamic point cloud asset, however, there is no specific timeline for this yet.
If you can estimate your cloud bounds tighter, you should get much better results overall.