That’s a good question. linalg has evolved from just linear algebra to more of a payload dialect that represents structured (perfectly nested or not) operations on large data (tensor/memref), so it’s not totally out of line.
But the convolution list explosion made us re-think that strategy. If we could common up into a tree (as you were trying to do partially), even if the parent wasn’t generic, it would be ok to be in the same dialect, because we can still do a lot of linear algebra (albeit with some creativity) on them.