I don't see why a high order polynomial is objectionable per se.
Would you object to a series of sines and cosines? Legendre polynomials are well behaved -- we don't use x^0, x^1, x^2, ... , x^10 for a set of functions, for example -- that would be bad.
1) Scanners do drift with long acquisitions, particularly because of heating from the gradient currents. This effect can be noticeable, and we used to have big problem with this on our scanners until GE put in a software fix. Now it's a smaller problem -- but still there. I don't think SNR changes much with time, but the image can become slightly distorted, for example. No citation, though.
2) If you model 25 min as separate runs (say 5x5 min), then you are going to use 2 or 3 polynomials per run, and then allow for possible discontinuities between runs. Since there are unlikely to be such discontinuities, then the baseline model actually would contain degrees of freedom that wouldn't be used. As long as you are careful to split the runs smack in the middle of the control task, then you could probably get away with the analysis. It would be interesting to compare the results to the analysis just using -polort 10, though. Probably not very large, I'd think.
3) Unless you have a task that requires a long run, I don't see why a run of more than 10 minutes would be used.
