The change in limits seemed to do the trick.

Incidentally, increasing the number of test points helped, but not enough to justify the increased time to check the problems. Increasing to 25 test points led to my test problem marked correct about 85% of time. This was substantially better than my theoretical computations, which indicated that should the test points be chosen uniformly, the correct answer would be marked correct about 20% of the time. However, 85% is still too low for my taste.

Playing around with specifying the "solutions" attribute has left me puzzled. It seems to have no effect on the problem. I entered "abs(x-1)+abs(y-1)=0.5" for the question which asked for a circle with center (1,1) and radius 0.5. Despite only having the cardinal points as indicated solutions, this answer was correctly rejected 100% of the time (in over 50 trials).