The problem, I believe, is that there are such few voxels in the mask (94 voxels) that random noise generated in the alignment process's masking matches the data well, and this is sensitive to these changes in the constraints. The optimization will continue to find the minimum cost. You can see what that cost is by using "-verb". The "-twopass" option is a solution for avoiding local minima, but with the small ROI and random noise, it would be "fooled" by the noise. I found good results that are less sensitive to the shift limits by using the "-nomask" option that should avoid the random noise masking. Also, using "@Align_Centers -cm" makes for a good starting point.