The thing is, we can't really know what the expected result should have been, since we don't know the true matrix. You want to make it possible to get help, by posting the matrix in a .mat file, not as only what was displayed by default. pinv is a pretty simple code, so the difference between what you got and what should have happened will be easy to track down.
I would guess the matrix has a singular value very near the cutoff point (but just below it), and depending on the tolerance for pinv relative to that singular value, then pinv effectively decides the matrix is singular in single precision. In that case, pinv zeros out that singular value when computing the pseudo-inverse.
The above scenario does not happen when you convert to double though, so everything works. But this is not really a bug in pinv at all, merely a reflection of the tolerance chosen.
