Since the polort baseline regressors are built per run, a step
function in one run (a constant added to some motion regressors
in a particular run) will only affect the corresponding baseline
betas (more specifically, just the constant term for that run).

There may be a small effect due to computational truncation that
departs from theory, but that is the only difference you should
see. Correlating with a step function will produce no difference
(except that 3dDeconvolve will whine about multicollinearity or
matrix errors).

Simple test: add 10 to the values in run 2 of one of the motion
regressors. Even that motion beta should be unaffected (again,
except for computational truncation artifacts). The only change
should be in the beta of the constant term for that run.

- rick