Excellent point - FIR is way more flexible .. but as you know comes with it's set of problems too.
To get back to the point of using derivatives - the variable duration model works well for detection but not so much when comparing multiple conditions against each other. I also agree (but this is mentioned in Grinband's paper) that other areas where there is no modulation will be less well modeled. So far (but I have to admit Jeanette and I have done limited (independent) simulation work on it) taking the mean over all trials for duration (do note I say all trials across all conditions - not condition specific) to model each regressor and add modulation by RT specific to each regressor seems to work best (as in control better your type 1 error / power) in the case you want to compare/contrasts multiple conditions -- hope this clarify a little the context
Best
Cyril