I wonder if my guess on end game is correct.
By changing one letter in 20a, I get a name who is associated with other names in 1a, 1d and 20a (by changing letters).
I followed a hunch and went with symmetry (in a broad sense - looking at symmetrical rows and columns, eg top row/bottom row) and that worked quite well - apart from there being no thematic name in the 5th column (there are two columns that have a pair of names in each of them).
Yes, I think you're right about the construction, hankir - it must have been tough cramming it all in with real words throughout (and some rare ones at that).