Yes, the epoxide opening step bothers me too. It's the only mechanism I can come up with right now though.
The cyanide exchanges I think are ok. We see reversible cyanide addition to carbonyls in the Benzoin condensation, so maybe it's no so unreasonable.
I should have used equilibrium arrows here, because then I think it makes more sense - the epoxide might open at the other end most often, but this just regenerates the previous intermediate and the epoxide can re-form. Even if opening of the epoxide as I've proposed is rare, the protonation of the final carbanion is going to lie heavily to the product side, since this is the most basic intermediate in the mechanism by several orders of magnitude - perhaps this step is "essentially irreversible" like the thermodynamic sink of a deprotonated Claisen condensation product?
I don't know, I see the holes too and I'm not entirely convinced either, but I think it's plausible at least. It's a neat transformation in any case.