Yes. You are right. Assuming that conditions are constant in neural activity duration, modeling under 2s doesn't make much of a difference. However, as soon as you have variable condition durations, I believe you need to account for the differential durations. My point was more that, once you pass 2s, you need to model the duration even for constant duration events.
With regard to (4), I just checked the most recent SPM release (r6906) and 0s is different from 0.0001s is different from 0.1s duration events, so there still seems to be some differences.