Currently LevelSet nodes are calculated with aligned origins and spacing, but independent grid boundaries which leads to the scenario you have found. Ideally we would clamp it to the boundary of the SimEmber at some point. That being said, it may not be possible to correctly determine the “inside or outside”-ness of a given sample position if the entire object is not considered.
You’ve discovered the limitations of both representations used by the InVolume and LevelSet nodes respectively. The KD-Tree approach used in InVolume is generally slower but is never in danger of using ridiculous amounts of memory when a poor sample spacing to object size ratio used with a level set.
I’ll add it to the bug list so we can later determine if bounds clamping is feasible.