A soft fork is always better if possible because of the "danger of two competing blockchains", and I don't see how a hard fork can solve the quadratic hashing problem here. (edited, see below) But legacy transactions cannot be permitted in an unrestricted way after the 2MB activation like today, as Lauda correctly points out:


So the "grace period" between the Segwit activation and the 2MB activation must be as long as possible (~1 year) to correctly evaluate the danger of "full block spam" (in this case "4 MB block spam"). Edit: And the 2 MB hard fork should contain also a restriction for legacy transactions. Maybe they can be prohibited and all people owning bitcoins on non-Segwit keys have to transfer them to Segwit addresses. But we don't have to do that already when the Segwit soft fork is activated (so we don't need a hard fork then), the 2 MB hard fork 1 year later would be early enough.

Regarding the possible 8 MB block spam, as I have already discussed in another thread, I would support going forward slower (first go to 1,5 MB/6 MB max. after a year and then to 2 MB after 1-2 years more if everything works right).