Lukas,
I still don't think that you really need any z-score conversion with either percent correct responses or mean reaction time.
The F-statistic part is indeed very confusing. In 3dRegAna, the user tells the program either t or F or both should be reported for each estimate. If you want t value for b0 and b1, that is straightforward. However if you want F for b0 and b1 (which is the case of example 1 in the manual), 3dRegAna provides the F-statistics for significance of the linear regression, which is the same thing for b0 and b1.
In your case with option 'bucket 0', the output file includes the estimate of each regression coefficent plus its t-statistic, F for regression, and R^2. So you didn't ask for F-statistic (and you don't really need it) for b0 and b1. So what it was said in example 1 in the manual does not apply to your case.
Yes, people usually are not interested in the intercept. Instead it is the slope that tells the dependence relationship between the independent variable (i.e., behavioral measure) and dependent variable(s) (i.e., percent signal change due to that behavior at each voxel). By thresholding the t value, you get various lit-up regions that indicate that the slope (dependence) is siginificantly different from 0 at those voxels, which mean there is strong dependence relation between behavioral measure and percent signal change of that behavior. Whether the pattern looks nice is a different story. :)
Gang