Looks like it's probably a bug. Can you package up the raster stack and fitList and send to me (off-list) as an .Rdata so I can take a closer look?
In the meantime you should be able to work around this by converting the raster stack values into a data frame (with as.data.frame), predicting with this data frame as normal, and then inserting the results back into a new raster.
Ken