Shouldn’t parallel aggregate come into play regardless of scan selectivity?
I'd say that the costing should take into account the estimated number of groups.
The more tuples that make it into each group, the more attractive parallel grouping should seem. In the extreme case if there's 1 tuple per group, then it's not going to be of much use to use parallel agg, this would be similar to a scan with 100% selectivity. So perhaps the costings for it can be modeled around a the parallel scan costing, but using the estimated groups instead of the estimated tuples.
