Perhaps this is what you mean by "deterministic", but isn't it possible for some collations to treat multiple byte sequences as equal values? And those multiple byte sequences wouldn't necessarily occur sequentially in C collation, so it wouldn't be possible to work around that by having the grouping node use one collation but the sorting node use the C one.
If my memory is incorrect, then this sounds like an intriguing idea.
I could see the value in a hash aggregate on C-collation that then passes itself as a partial aggregate up to another step which applies the collation and then finalizes the aggregation before sorting
