Independent clamping of face-averaged volume fractions can inflate mixture properties

[sbryngelson]

Diffusive face fluxes weight a mixture property by the face-averaged volume fraction, clamping each component independently:

alpha_avg(fl) = 0.5_wp*(q_prim_vf(adv%beg + fl - 1)%sf(...) + q_prim_vf(adv%beg + fl - 1)%sf(...))
! Raw cell-centered alphas can under/overshoot near interfaces; clamp to [0,1]
alpha_avg(fl) = min(max(alpha_avg(fl), 0._wp), 1._wp)

Three sites share this pattern:

• src/simulation/m_riemann_state.fpp:880 (cylindrical viscous source flux)
• src/simulation/m_riemann_state.fpp:1091 (Cartesian viscous source flux)
• src/simulation/m_conduction.fpp (Fourier conduction face conductivity, added in Fourier heat conduction in the energy equation #1906)

The problem. Clamping each component independently does not preserve the sum. With three or more fluids, a physically-summing set can clamp to a sum greater than one:

(0.6, 0.6, -0.2)  sums to 1.0
clamped ->
(0.6, 0.6,  0.0)  sums to 1.2

The mixture property is then inflated by up to that factor — it is no longer a convex combination of the per-fluid values, so it is not bounded by [min_i prop_i, max_i prop_i].

Scope.

• Unreachable when mpp_lim = T: s_convert_species_to_mixture_variables_kernel already clamps and renormalizes (alpha_K = alpha_K/max(alpha_K_sum, sgm_eps)), so cell alphas sum to one and the face clamp is a no-op.
• Sum-preserving for two fluids whose alphas sum to one — (1.05, -0.05) clamps to (1.0, 0.0), still one.
• Real for three or more fluids with mpp_lim = F (the default).

Suggested fix, applied to all three sites together so the two diffusive terms keep treating interfaces the same way:

prop_face = sum_i(alpha_face(i)*prop(i)) / max(sum_i(alpha_face(i)), sgm_eps)

Identical to the current result whenever the clamped alphas sum to one, and a true convex combination otherwise. Doing it in one place for both viscous and conduction avoids the two diverging; doing it for conduction alone would be worse than leaving it.

Regenerating whichever viscous goldens move is part of the work.

Raised by an automated review on #1906. The conduction site copies the existing viscous convention deliberately, so this is a property of the shared pattern rather than something that PR introduced.