A few questions :)

[Sven]

SEE would be way too slow to compute for captures as well as quiets. It should only be used for captures, and even then one should be conservative with using it.

Agreed, and therefore in qsearch one can restrict applying SEE only to moves that actually have potential to lose material, i.e. only in "higher piece captures lower piece" cases. There is no point in calculating SEE for a move that captures a queen, for instance.

[hgm]

Since just a few percent of the nodes are full-width nodes, and non-captures are only considered there, (and then only if hash move, null move or a capture didn't produce a cutoff), it seems to me you could almost afford anything for non-captures.