> How does the official spm software handle this? maybe there are some hard
> constraints how much of the temporal/spatial derivative can be added.
The three basis functions coded in option SPMG3 of 3dDeconvolve are exactly the same as implemented in the SPM. Over-fitting is the caveat that can't be easily avoided with the approach of multiple basis functions since little constraint is imposed. The issue of over-fitting aside, the fitting itself with your data seems reasonable. To see how the three components end up with the blue curve, you can plot out each of the three components and see how they look like.
Gang