I realized a few minutes later that I was mistaken to write:
OTOH maybe that is not necessary because probably the whole array does not have to be set up in order to tell whether a given key will work. 1/3 of the entries in the table are fixed once they have been swapped once, so if you checked after doing the first 20 entries, say, about 7 should have their final values, and we can perhaps reject a key already in a known plaintext situation just from that. So actually the large table size may not help against exhaustive key search. (The mod I suggested to the key setup would defend against this possibility, which raises the question of whether this design aspect was chosen to allow for export approval.)
Just knowing several of the first few entries in the table doesn't allow you to quickly reject keys because the algorithm selects entries from throughout the table to xor with the data stream. So this does not imply that keys can be rejected quickly, nor does it suggest that the particular setup algorithm used is particularly weak or was chosen for export approval. Sorry about the error. Hal