The fact is, matrices and vectors **already** contain a seq, which is allocated on the heap (at least in the common case), so there is not much more cost in adding an additional ref.
The only case where avoiding a ref would actually save time is when many matrices or vectors refer to the same underlying seq, but point at different places. A common case is iterating over the rows or columns of a matrix. I have decided to just optimize this case for now, by reusing the same ref, and therefore avoiding allocations (well, there is just a single allocation for the whole iteration)
