I just read this argument in this paper (PDF).
It suggests that, from variational principles, you can show that you can always lower the energy of a state by making the phase constant, thus resulting in a ground state that must have no phase change. I know this to be true for 1D systems (ground state is always real), but I'm not so convinced for the general $N$-dimensional case, though that's what the paper says as it uses the nabla for spatial derivatives.
My problem with their argument is that they split the trial wavefunction as
$$ \psi = f(\mathbf{r})e^{i\chi(\mathbf{r})} $$
but even if you send $\chi$ to zero, what proof do we have that an $f$ that, alone, solves the Schrödinger equation always exists? Is that a separate theorem? It doesn't seem trivial to me.