Fix issues #250-264

[asalmgren]

Fixes #250, fixes #251, fixes #252, fixes #253, fixes #254, fixes #255, fixes #256, fixes #257, fixes #258, fixes#259, fixes #260, fixes #261, fixes #262, fixes #263, fixes #264.

Correctness / physics

• set_inflow_velocity / prob_set_inflow_velocity have had no caller since PR #145, and the probtype 33/333 branches write velocity component 2 outside any AMREX_SPACEDIM==3 guard (out-of-bounds in a 2D build if ever revived) #250 Delete the dead set_inflow_velocity / prob_set_inflow_velocity, which have had no caller since PR CC Projection Outflow Bugfix #145 and wrote velocity component 2 with no AMREX_SPACEDIM guard for probtypes 33/333.
• incflo.use_mac_phi_in_godunov is half-implemented: it drops grad p from the MAC face prediction, solves into mac_phi and doubles it, but nothing ever reads mac_phi; the "macphi" plot variable is uninitialised memory on regridded or restarted levels #251 Refuse incflo.use_mac_phi_in_godunov: it only drops grad p from the MAC face prediction and is never read back as a pressure. Also zero mac_phi in the LevelData constructor so the macphi / error_mac_p plot variables are defined on levels created by regrid or restart.
• compute_convective_term / compute_MAC_projected_velocities: with use_tensor_correction + Godunov + godunov_include_diff_in_forcing, the "viscous term" added to the edge-state forcing is divtau_o = (tensor - scalar) difference, not div(eta grad u)/rho #252 use_tensor_correction now forces godunov_include_diff_in_forcing off, because divtau_o then holds the (tensor - scalar) difference rather than the full explicit viscous term the edge-state forcing expects.
• InitialPressureProjection projects the body-force field (rho-rho0)/rho*g with set_inflow_bc=true, so ApplyNodalProjection writes the inflow *velocity* into its ghost cells (and runs enforceInOutSolvability on it) before taking the divergence #253 InitialPressureProjection passes set_inflow_bc = false; the field being projected is the body force (rho-rho0)/rho*g, not a velocity.
• ApplyCCProjection uses the caller's u_mac/v_mac/w_mac as scratch, so evolveTracerParticles afterwards sees the projected new-time face field instead of the half-time MAC velocities #254 ApplyCCProjection uses its own face scratch instead of the caller's u_mac/v_mac/w_mac, which the caller still needs for tracer particle advection.
• init_jump: probtype 1100 writes velocity component 2 and probtype 1102 reads domain.length(2) with no AMREX_SPACEDIM guard, so a 2D build with either probtype corrupts memory in prob_init_fluid instead of being refused #256 Refuse probtypes 1100/1102 in a 2D build and guard the z-direction code in init_jump, which otherwise wrote past the end of a 2-component FAB.

Defensive gaps and diagnostics

• get_diffuse_tensor_bc has no BC::mixed case: a mixed BC with the default use_tensor_solve aborts at the first diffusion solve with "undefined BC type" instead of at read time; get_diffuse_velocity_bc's abort names the wrong function #255 Reject a mixed BC combined with use_tensor_solve at read time (neither MLTensorOp nor MLEBTensorOp implements Robin BCs), add the BC::mixed case to get_diffuse_tensor_bc, and fix get_diffuse_velocity_bc's copy-pasted abort message.
• All analytic EB builders call EB2::Build(gshop, geom.back(), 0, 100): with amr.max_level > 0 the coarse AMR levels are optional coarsenings with no EB data in domain ghost cells, and a failed coarsening surfaces later as "Geometry not found" #257 Every EB builder now requires max_level coarsenings, as the STL builder already did, so coarse AMR levels get EB data in their domain ghost cells.
• MakeEBGeometry: an unrecognised incflo.geometry (a typo, a 3D-only name in a 2D build, or csg in a build without CSG_EB) falls through to make_eb_regular and the run proceeds with no embedded boundary #258 An unrecognised incflo.geometry aborts instead of silently building a regular geometry; only an empty name or all_regular selects make_eb_regular.
• EB is built in the incflo constructor from the inputs' prob_lo/prob_hi, but ReadCheckpointFile resets the physical domain from the checkpoint header with no check: a restart with edited prob_lo/prob_hi keeps the EB at the wrong physical location #259 Refuse a restart whose checkpoint prob_lo/prob_hi differ from the inputs the EB was built from.
• Three accepted options do nothing or abort later: incflo.test_tracer_conservation (consumer commented out; 8 decks set it), incflo.ic_p (sets m_p000, which nothing reads), amr.plt_divu / "divu" (documented, aborts "xxxxx TODO" at the first plotfile) #264 Refuse the unimplemented divu plot variable when the inputs are read rather than at the first plotfile, and refuse a non-zero incflo.ic_p, which only sets m_p000 and would otherwise be silently dropped.

Build coverage, docs and shipped decks

• No CI job builds particle support: no workflow passes INCFLO_PARTICLES=ON, no GNUmakefile sets USE_PARTICLES; src/particles/* and every #ifdef INCFLO_USE_PARTICLES block are never compiled or run (nine particle bugs in two audit rounds, none catchable) #260 Build particle support in one GCC and one CUDA CI job, run the 2D smoke test with tracer particles, and add USE_PARTICLES to the four GNUmakefiles.
• Docs/Particles.rst still documents the i%2==0 && j%2==1 quarter-cell seeding removed by PR #213 and advertises a tracer_particles_mass_density plot variable that nothing can produce (massDensity/GetMeshPlotVar have no caller) #261 Drop the stale quarter-cell seeding note removed by PR fix issues 171 176 181 188 189 191 #213 and the tracer_particles_mass_density plot variable that nothing can produce.
• Docs document seven input keys under names/prefixes the code never queries (incflo.refine_particles, amr.write_eb_surface, incflo.rho_0, {nodal_proj,mac_proj,diffusion}.rtol/atol, diffusion.*): setting them as documented is silently ignored #262 Document the input keys the code actually queries (ro_0, incflo.write_eb_surface, amr.refine_particles, mg_rtol/mg_atol, the scalar_diffusion/tensor_diffusion prefixes and the mg_ iteration keys).
• Shipped decks set keys nothing reads: incflo.v (vortex_in_sphere, convergence decks), amr.plot_p (inputs.channel_sphere), cylinder.speed (three rotating/Couette-cylinder decks, their only flow driver), cylinder.height (19 decks) #263 Rename incflo.v, amr.plot_p and incflo.use_godunov in the shipped decks, drop the unread cylinder.height, and mark cylinder.speed (and the two decks it was the only flow driver for) as not implemented.

Testing

• 2D/3D x EB/no-EB x with/without particles: five configurations, all build with no errors or warnings.
• Every new guard confirmed to fire at runtime: use_mac_phi_in_godunov, ic_p, divu (all three entry paths), 2D probtype 1100, typo'd and 3D-only geometry names, mixed BC + tensor solve, and the EB restart domain mismatch.
• Runs clean: benchmark.split (mixed BC, max_level = 1), the eb_chkptfile write -> read round trip, a 3-level rayleigh_taylor with macphi under amrex.init_snan=1, use_cc_proj = 1 with tracer particles, and a sweep of shipped 2D/3D decks.
• One pre-existing failure, unrelated to this branch: test_3d/inputs.channel_sphere aborts with MLMG failed. on unmodified source too (and its amr.max_level = 0, so the All analytic EB builders call EB2::Build(gshop, geom.back(), 0, 100): with amr.max_level > 0 the coarse AMR levels are optional coarsenings with no EB data in domain ghost cells, and a failed coarsening surfaces later as "Geometry not found" #257 change is a no-op there).

🤖 Generated with Claude Code