Fix issues #233-#239

[asalmgren]

Fixes #233, #234, #235, #236, #237, #238, #239.

What's in here

Issue	File	Change
#233	src/convection/incflo_compute_MAC_projected_velocities.cpp	The direction_dependent inflow tests read a zeroed scratch MultiFab, whose ghost cells hold the interior copy (0) on the outflow part of an in/out face. The non-strict >=/<= tests treated that as zero inflow and overwrote the extrapolated outflow face velocity with 0, after which Hydro aborts with "no outflow from the direction dependent boundaries". Tests made strict.
#234	src/prob/prob_bc.H	The Burggraf (probtype 16) lid profile 16x²(1-x)² is the tangential velocity u on the high-y wall, not the normal velocity v. The PR #122 refactor assigned it to norm_vel, so benchmark.burggraf ran a uniform-speed lid with a spurious normal velocity through it. Pre-#122 behaviour restored.
#235	src/prob/prob_init_fluid.cpp	probtype 43 shared probtype 42's zero initial velocity, so the outflow part of its direction_dependent x faces started at 0 and InitialProjection's in/out solvability check saw no outflux. probtype 43 now starts from the zero-net-flux profile u = 6y(1-y)-1 that IncfloVelFill imposes on those faces.
#236	src/incflo_compute_dt.cpp	The explicit-diffusion limit used the constant m_mu, ignoring the strain-rate dependent viscosity and the tracer/temperature diffusivities, which are advanced explicitly by the same m_diff_type switch. The effective diffusivity is now built per cell from compute_viscosity_at_level, max_n mu_s[n] and mu_T/cp, divided by rho, with covered cells zeroed.
#237	src/incflo_regrid.cpp, src/incflo.cpp	particleData.Redistribute() inside RemakeLevel/MakeNewLevelFromCoarse runs before AmrCore::regrid installs the new BoxArrays, DistributionMappings and finest_level, so it redistributed against the old layout and the first step after a regrid indexed u_mac with stale grid indices. Now redistributes once after regrid() returns.
#238	src/incflo_apply_predictor.cpp	ApplyPredictor advected the tracer particles during InitialIterations. The fields are restored after each initial iteration but the particle positions are not, so particles started the run m_initial_iterations*dt ahead of the fluid. Skipped when incremental_projection is set.
#239	src/prob/prob_bc.H	IncfloDenFill/IncfloTracFill/IncfloTempFill never wrote the outflow part of a direction_dependent face (filcc has no case for that BC type and the functors were inflow-only), and decided inflow/outflow from the raw input velocity rather than the probtype profile IncfloVelFill uses. The prescribed normal velocity is factored into incflo_bc_normal_velocity(), used by all four functors, and the scalar functors gained the same "outflow: copy the first interior cell" branch IncfloVelFill has.

Verification

Baseline below is pristine ab031bed; both built DOUBLE/optimized. All five build configurations (2D/3D × EB/no-EB, plus USE_PARTICLES=TRUE) compile with no new diagnostics.

#233 + #235 — benchmark.inout: baseline aborts "Cannot enforce solvability, no outflow from the direction dependent boundaries"; with these fixes it runs all 10 steps in both 2D and 3D.

#234 — Burggraf u error at steady state (stop_time=20):

N	baseline Norm0	fixed Norm0	fixed Norm2
16	1.55	1.042e-2	3.580e-3
32	1.02	3.258e-3	9.529e-4
64	—	9.140e-4	2.407e-4

Second order (Norm2 orders 1.91, 1.98) where the baseline error was O(1) and non-convergent. 3D at N=16 agrees: 0.428 → 0.0102.

#236 — Bingham (mu=1, tau_0=1, papa_reg=1e-3) with diffusion_type=0: baseline dt = 4.883e-5, fixed dt = 4.878e-8 — a factor of 1001, exactly eta(sr→0)/mu.

#237 — rayleigh_taylor with amr.max_level=2 amr.regrid_int=2 and tracer particles: baseline segfaults (exit 11) on the step after the first regrid; fixed runs 12 steps through 5 regrids.

#238 — plt00000 particle data with initial_iterations 0 vs 3 (Godunov): baseline differs, fixed is identical.

No regressions — plt files bit-identical to baseline (fcompare) on 24 benchmarks: 5 in test_no_eb_2d, 5 in test_2d, 9 in test_no_eb_3d, 10 in test_3d — including every 3D-only probtype routed through the new helper (311, 32, 322, 33, 333, 1102, 41, 42).

Pre-existing issues noticed while testing (not addressed here)

1. init_burggraf (prob_init_fluid.cpp) uses x = (i+0.5)*dx - 0.5, i.e. a domain shifted to [-0.5, 0.5], while DiffFromExact and the probtype-16 forcing use x = (i+0.5)*dx. The IC is not the exact solution it is meant to be.
2. benchmark.burggraf sets stop_time = 5, which does not reach steady state — the 32² error is 0.040 at t=5 versus 0.0033 at t≥15. test_convergence_burggraf therefore measures a transient and needs stop_time ≈ 15-20 to show the convergence above.
3. benchmark.inout's fixed_dt = 0.01 exceeds the CFL limit by ~7× with the corrected IC (the warning is printed; the run is stable since the problem is inviscid).
4. test_no_eb_3d/benchmark.burggraf sets a 2-component amr.n_cell, so it aborts in 3D before doing anything; at 32×32×8 the nodal MLMG then fails identically in baseline and with these fixes.
5. In single precision (PRECISION = FLOAT) MLMG cannot reach its default rtol, so most benchmarks abort with "Failed to converge" regardless of this branch — e.g. untouched benchmark.double_shear_layer stalls at resid/bnorm = 1.65e-6.

🤖 Generated with Claude Code