Hi, Don. I had met exact same question before. I have thought about this. I think it is valid to use 3dANOVA2 to analyze this type of experiment.
To answer why this way produce so much more significant t-stats, we need to check the difference of the t-test in 3dttest and t-test in 3dANOVA2. The numerators for these two t-tests are the same, which is the average difference between the 2 stimulus conditions. The denominators for these two t-tests are different. 3dttest uses the sum of within group sample variance as the estimate of sigma^2. 3dANOVA2 uses the MSE as the estimate of sigma^2, and MSE has taken account of the effect of subjects. If the subjects have so much difference from each other, then MSE will be much smaller than the estimate of sigma^2 in 3dttest. That will make the t-stats in 3dANOVA2 much larger than t-stat in 3dttest.
The document for 3dANOVA2 dividing estimate of sigma by sqrt(b*n), not sqrt(b), because in that notes, n is the replication of each subject for each condition. For your case, n=1. Hope this will not confuse you.