In viscosity.f90, we had a piece:

\begin{verbatim}
        if (ldynamical_diffusion .and. lvisc_hyper3_mesh) then
          diffus_nu3 = p%diffus_total3 * sum(abs(dline_1),2)
        else
          diffus_nu3 = p%diffus_total3*dxyz_6
        endif

which was not correct, because it makes diffus_nu3 finite in cases that
have nothing to do with hyper3. I noticed this because of its timestep
constraint, and this could have been a problem also in a few other cases.
I noticed that this piece of code goes back to 2011-08-02 and it seems
that it was coded by Chao-Chin with the message

  By fixing the mesh Reynolds number, dynamically adjust the mesh hyper3
  coefficients for density, magnetic, and viscosity.  Generally work with
  non-equidistant grids.  cdtv3 may be as high as 0.9, or even 1.

after earlier work by Wlad. I then disentagled the loop into two, but
realized that diffus_nu3 will be needed in a few other cases. For example
in samples/MRI-turbulence_hyper we have lvisc_hyper3_rho_nu_const_symm=T,
so I included this and it worked. I have now added all 14 logicals
that seem to be needed.

        if (lvisc_hyper3_mesh .or. lvisc_hyper3_simplified .or. lvisc_hyper3_simplified_tdep .or. &
            lvisc_hyper3_mesh_residual .or. lvisc_hyper3_mesh .or. lvisc_hyper3_rho_nu_const .or. &
            lvisc_hyper3_mu_const_strict .or. lvisc_hyper3_nu_const .or. &
            lvisc_hyper3_cmu_const_strt_otf .or. lvisc_hyper3_rho_nu_const_symm .or. &
            lvisc_hyper3_rho_nu_const_aniso .or. lvisc_hyper3_nu_const_aniso .or. &
            lvisc_hyper3_rho_nu_const_bulk .or. lvisc_hyper3_nu_const) then