1) If the collinearity is only in baseline (AKA null hypothesis) components (e.g., -polort or -stim_base stuff), then -GOFORIT should be fine. It's when the collinearity involves a signal component that -GOFORIT usage needs to be thought through carefully.
2) The 0th column IS the constant term, so you must be getting a nearly constant in time global average. The program checks for "near collinearity" between column pairs -- I assume you are getting a message like
"!! * Columns 0 [label] and xx [label] are nearly collinear!"
-- so the global regressor is not necessarily EXACTLY constant but must be pretty close to constant for this message to appear. Are you calculating this regressor using 3dmaskave? I don't know why some subjects would be more constant than others, but it could be that they are all very close to constant but some are just over the threshold for reporting and others aren't. Or it could be that you did something different when computing the global regressor for those subjects, and don't have a record of it handy.
Probably it would be best to remove the mean from the global regressor time series before using it, anyhoo.