fix: correct MTHINC normals on non-uniform grids

[sbryngelson]

Summary

Builds on #1303 (MTHINC implementation). Fixes two bugs found during review:

• Fixes MTHINC: interface normals inaccurate on non-uniform (stretched) grids #1395 — MTHINC interface normals inaccurate on non-uniform grids
• Fixes THINC/MTHINC: int_comp incompatible with viscosity and surface tension (reconstruction split silently disables compression) #1396 — int_comp incompatible with viscous / surface-tension flows

Fix 1: MTHINC normals on non-uniform grids (#1395)

The reference-space normal in s_compute_mthinc_normals used a uniform-grid assumption (* 0.5) that gives the wrong direction when cell widths vary. The correct weighting is:

n_ξ ∝ (∂α/∂x) · Δx_j  =  [α(j+1) - α(j-1)] · (x_cb(j) - x_cb(j-1)) / (x_cc(j+1) - x_cc(j-1))

On uniform grids this ratio equals 0.5 exactly, so all existing tests pass unchanged.

Fix 2: int_comp with viscous / surface-tension flows (#1396)

int_comp > 0 requires full (non-split) reconstruction, but the viscous and surface-tension RHS paths were calling split reconstruction, silently disabling interface compression. Fixed by routing those paths through the full reconstruction branch and adding a runtime @:PROHIBIT guard.

New regression test (7A1719C6)

A new golden-file test exercises Fix 1 on a non-uniform grid:

• 2D domain with cosh-based x-stretching, creating non-uniform cell widths
• Circular fluid-2 bubble creates a curved interface with diagonal normals (both nr_x ≠ 0 and nr_y ≠ 0) — only diagonal cells are sensitive to the Δx_j weighting correction; axis-aligned interfaces always normalize to ±1 regardless of scaling
• Uniform vel(1)=0.5 advects the bubble so MTHINC reconstruction produces non-trivial fluxes; 129+ distinct values at t=50 confirm genuine interface evolution

Test plan

• All 4 existing uniform-grid MTHINC tests pass unchanged (5126B21F, 4E4FECA9, 4F3722DB, 4C4F339C)
• New non-uniform grid MTHINC test passes (7A1719C6)
• ./mfc.sh precheck -j 8 passes
• ./mfc.sh build -t simulation -j 8 compiles cleanly

[github-actions]

Claude Code Review

Head SHA: ae74a6a

Files changed:

• 16
• src/simulation/m_thinc.fpp (new)
• src/simulation/m_global_parameters.fpp
• src/simulation/m_muscl.fpp
• src/simulation/m_weno.fpp
• src/simulation/m_rhs.fpp
• src/simulation/m_start_up.fpp
• src/simulation/m_mpi_proxy.fpp
• src/common/m_constants.fpp
• toolchain/mfc/case_validator.py
• toolchain/mfc/params/definitions.py

---

Findings

1. GPU correctness bug — Gauss-quadrature constants uninitialized on device (MTHINC only)

File: src/simulation/m_thinc.fpp

real(wp) :: gq3_pts(3) = [-5e-1_wp*0.7745966692414834_wp, 0._wp, 5e-1_wp*0.7745966692414834_wp]
real(wp) :: gq3_wts(3) = [5._wp/18._wp, 8._wp/18._wp, 5._wp/18._wp]
real(wp) :: ln2 = 0.6931471805599453_wp
$:GPU_DECLARE(create='[gq3_pts, gq3_wts, ln2]')

GPU_DECLARE(create=...) allocates device memory but does not copy the host values. The host initializers (set at declaration time) are never transferred to the device — s_initialize_thinc_module has no $:GPU_UPDATE(device='[gq3_pts, gq3_wts, ln2]') call.

gq3_pts and gq3_wts are accessed inside GPU_PARALLEL_LOOP kernels via the GPU_ROUTINE(parallelism='[seq]') functions f_mthinc_volume_integral, f_mthinc_volume_integral_dd, and f_mthinc_face_average. On any GPU build with int_comp=2, these routines read uninitialized device memory, producing garbage Gauss-quadrature results and silently incorrect MTHINC reconstruction.

Fix: add $:GPU_UPDATE(device='[gq3_pts, gq3_wts, ln2]') in s_initialize_thinc_module, or change the declare to GPU_DECLARE(copyin='[gq3_pts, gq3_wts, ln2]'). This does not affect int_comp=1 (THINC), which does not use these arrays.

---

2. Dead device function — f_log_cosh compiled for GPU but never called

File: src/simulation/m_thinc.fpp

function f_log_cosh(x) result(res)
    $:GPU_ROUTINE(parallelism='[seq]')
    ...
end function f_log_cosh

f_log_cosh is not called anywhere in the module. f_thinc_integral_1d calls f_log_cosh_diff directly (via atanh(tanh(a)*tanh(h))), bypassing f_log_cosh entirely. The function and its sole dependency ln2 are dead code that will nonetheless be compiled into device code on GPU builds. Both should be removed, along with ln2 from the GPU_DECLARE list.