> is it statistically sound to have 4 task regressors for some participants, and only 3 for others?

If by "statistically sound" you meant at group analysis, this is a relevant issue indeed. If you take both beta and t-statistic for group analysis with 3dMEMA, it should not matter whether you have equal number of time points in the time series for each regressor across subjects or not since the t-statistic automatically contains such information. However, this may become moot if you only take beta values for group analysis with the conventional approach because it carries two underlying assumptions: (1) within-subject variability is mall relative to cross-subject variability; and (2) within-subject variability is about the same across subjects. You might be OK especially with assumption (2) if you have large number of time points in the time points.

Gang