I think you have it wrong. Wolfram's claim is that for a wide array of small (s,k) (including s <= 4, k <= 3), there's complex behavior and a profusion of (provably?) Turing machine equivalent (TME) machines. At the end of the article, Wolfram talks about awarding a prize in 2007 for a proof that (s=2,k=3) was TME.

The `s` stands for states and `k` for colors, without talking at all about tape length. One way to say "principle of computational equivalence" is that "if it looks complex, it probably is". That is, TME is the norm, rather than the exception.

If true, this probably means that you can make up for the clunky computation power of small (s,k) by conditioning large swathes of input tape to overcome the limitation. That is, you have unfettered access to the input tape and, with just a sprinkle of TME, you can eeke out computation by fiddling with the input tape to get the (s,k) machine to run how you want.

So, if finite sized scaling effects were actually in effect, it would only work in Wolfram's favor. If there's a profusion of small TME (s,k), one would probably expect computation to only get easier as (s,k) increases.

I think you also have the random k-SAT business wrong. There's this idea that "complexity happens at the edge of chaos" and I think this is pretty much clearly wrong.

Random k-SAT is, from what I understand, effectively almost surely polynomial time solveable. Below the critical threshold, it's easy to determine in the negative if the instance is unsolvable (I'm not sure if DPLL works, but I think something does?). Above the threshold, when it's almost surely solveable, I think something as simple as walksat will work. Near, or even "at", the threshold, my understanding is that something like survey propagation effectively solves this [0].

k-SAT is a little clunky to work in, so you might take issue with my take on it being solveable but if you take something like Hamiltonian cycle on (Erdos-Renyi) random graphs, the Hamiltonian cycle has a phase transition, just like k-SAT (and a host of other NP-Complete problems) but does have a provably an almost sure polynomial time algorithm to determine Hamiltonicity, even at the critical threshold [1].

There's some recent work with trying to choose "random" k-SAT instances with different distributions, and I think that's more hopeful at being able to find difficult random instances, but I'm not sure there's actually been a lot of work in that area [2].

[0] https://arxiv.org/abs/cs/0212002

[1] https://www.math.cmu.edu/~af1p/Texfiles/AFFHCIRG.pdf

[2] https://arxiv.org/abs/1706.08431

P vs. NP and the Difficulty of Computation: A ruliological approach