Your option 2) seems reasonable to me. Or you may further differentiate the incorrect trials between tasks A and B if you believe the brain may respond differently between the two types of incorrect trials.
You may also consider using BLOCK(4,1) instead of GAM since the stimulus duration is longer than one TR. The impulse response function (with option -iresp) from 3dDeconvolve would be boring when you use GAM or BLOCK(d,p) as your basis function since that's just a standard curve.
HTH,
Gang