input%ellow does not cut away eigenvalues
Christoph told me that he once had a version of fleur in which eigenvalues below input%ellow were cut away.
This can be useful if a very large valence basis is used (which may be required for GW) and this basis provides a reasonable but not perfect representation for core states. You can then decide to not use the valence representation of the core states but to describe them with the core solver.
It is probably not difficult to re-implement this feature. We should do it if nothing prevents us from doing it. Maybe we have to define a new variable for this if the meaning of input%ellow changed too much.