Isn't the core assumption of MVT that there exists a one dimensional spectrum from which to calculate a median in the first place? Voting for candidates is hardly one-dimensional; even a particular ballot measure may have multiple choices, or another question on the ballot of a related but somewhat conflicting nature.

I think the assumption/result is that even though you only get one out of two candidates to choose from and there are hundreds of independent issues, in aggregate, after all forces and processes have had their effect, resulting policy in most major issues is surprisingly close to what the median voter wants.

Caveat: I'm just some guy who reads web pages, not a voting scientist.