> How critical is the choice of duration for TENT(zero)/SPLIN(zero) etc. models?

It's really hard to pinpoint an exact number because several factors play roles about the response shape and length, and it may vary across cognitive states, stimulus types, tasks, brain regions, subjects, and groups. Too small a number would render an inaccurate characterization of the response, while overparameterization through too many basis functions pays the price in degrees of freedom, and may run the risk of over-fitting (leaking physiological noise or head motion into the response estimate) as well as multicollinearity problem.

You may want to try a few numbers on a couple of subjects, and see how they pan out.

Gang