after a long struggle, it appears that i finally have kociemba's
algorithm working properly, with the improvement that dik mentioned.
the first pattern i tested was "anaconda" (bandelow's terminology),
which dik called "walker's worm" (singmaster's terminology?).
(i think that's the same pattern.) within 10 minutes, most of which
was initializing tables, the program produced the following 14 quarter
turn maneuver, which is better (in the quarter turn metric) than the two
maneuvers bandelow gives in his book. here it is:
anaconda: (14 qt) B1 R1 D3 R3 F1 B3 D1 F3 U1 D3 L1 F1 L3 U3.
more results as they come in. my program isn't quite finished yet;
hopefully i won't screw it up attempting to improve it.