Fix issues #179, #181-#185, #188-#190

[asalmgren]

Fixes #179. Fixes #181. Fixes #182. Fixes #183. Fixes #184. Fixes #185. Fixes #188. Fixes #189. Fixes #190.

(#180 is not included -- it was already fixed on development by c44cbcf.)

What's in here

Issue	File	Fix
#179	Source/NS_derive.cpp	The EB one-sided vorticity stencils in dermgvort read velocity two cells away without checking those cells are uncovered -- isConnected only certifies the immediate neighbor. Now the intermediate cell is certified before taking the quadratic form, with a two-point difference when only one cell is available and zero when both sides are disconnected (2D x/y, 3D x/y/z).
#181	Tutorials/TaylorGreen/benchmarks/ViscBench.cpp	The loop over global grid indices dereferenced MultiFabs by global index, and FillVar(FArrayBox*,...) fills on one rank only, so the tool could not work under MPI. Switched to the distributed FillVar(MultiFab&,...) overload, drove the exact-solution loop from MFIter, and moved the prints to amrex::Print() so the collective norm* calls stay on every rank.
#182	Source/NS_getForce.cpp	The getForceVerbose min/max loops indexed State and force by absolute state component, reading past the end of the velocity-only State FAB and the 0-indexed force FAB that scalar callers pass. Each FAB is now indexed by its own layout; the printed label keeps the absolute component.
#183	Source/main.cpp	The regrid-on-restart test lacked the max_step < 0 guard the time loop applies, so with only stop_time set it regridded and wrote a spurious plotfile/checkpoint before stepping anyway.
#184	Tutorials/HIT/NS_getForce.cpp	The coarse hi bound was one node short whenever ihi mod ff_factor was in [1, ff_factor-2], so the interpolation read past ff_force. Now ihi/ff_factor + 1, exactly the node the kernel touches (the three ihi == ff_ihi*ff_factor adjustments become dead and are removed). ff_force also gets an Elixir so its device memory is not freed while the asynchronous kernels still read it.
#185	Tutorials/HIT/TurbulentForcing_def.H	Nothing constrained the mode indices writing into the flat stack array tmp, so large nmodes or a high-aspect domain corrupted the stack; the mode counts are now asserted to fit array_size. The GPU copy out of tmp was asynchronous with no sync before the stack frame died -- switched to the synchronous htod_memcpy. Also fixes the verbose block, which printed nxmodes in place of nymodes and nzmodes.
#188	Source/prob/prob_init.cpp	meanFlowDir = +/-2 assigned vel_x = v_vort and vel_y = u_vort, giving a non-solenoidal, essentially irrotational field instead of a vortex; un-swapped so only the mean flow term moves to the y component. "No mean flow" is now an explicit case 0, with a host-side Abort for out-of-range directions that the bare default silently accepted.
#189	Source/MacProj.cpp	The known-edgestate mac_sync_compute overload passed a default-constructed MultiFab as the state, which ComputeAofs indexes unconditionally -- now an alias of Sync. mac_sync_compute built the divu constraint at prev_time only, omitting the 0.5*dt*dsdt half-time correction that velocity_advection and scalar_advection apply, so the sync reconstructed edge states with a different divu than the advance used. test_umac_periodic did host-side FArrayBox work on device-resident data; it now gets pinned host copies.
#190	Source/NavierStokesBase.cpp	Both limiters seeded the running maximum with numeric_limits<Real>::min(), the smallest positive normal rather than lowest(), so the upper clip bound was wrong for any all-negative stencil. With lowest() binding, covered cells are skipped so they keep the COVERED_VAL invariant. do_denminmax called ConservativeScalMinMax with Density as both the scalar and the normalizing density, clamping rho/rho == 1 against rhoo/rhoo == 1 -- a provable no-op; it now calls ConvectiveScalMinMax on the raw density.

Heads-up: benchmarks need regolding

The do_denminmax fix makes ns.do_denminmax=1 change results for the first time (it was previously a bit-exact no-op). Exec/eb_run2d/regtest.2d.hotspot and Exec/eb_run3d/regtest.3d.hotspot enable the flag, so their benchmarks will need to be regolded. On eb_run2d hotspot at step 4, tracer max goes from 1.0012879887 to 1.0000060234 and density max from 1.0011269247 to 1.0011264227 -- i.e. the limiter now does its job.

Build and test

Built clean: run2d, run3d, eb_run2d, eb_run3d, Tutorials/HIT (both USE_FAST_FORCE=TRUE and FALSE), and ViscBench with USE_MPI=TRUE.

All shipped regtests were run in 2D and 3D, EB and non-EB. eb_run2d/regtest.2d.shock_past_cylinder aborts with "MLMG failed" -- this fails identically on unmodified development, so it is pre-existing and unrelated.

Targeted verification against baseline binaries built from development:

• init_ConvectedVortex: meanFlowDir=+/-2 swaps u/v; meanFlowDir=0 leaves velocity unset #188 -- meanFlowDir=2 now gives max|u| = 0.0832, max|v| = 1.0832, the exact mirror of meanFlowDir=1. meanFlowDir=4 aborts with the new message.
• main(): regrid-only branch misses max_step < 0 guard, fires when only stop_time is set #183 -- baseline restart with only stop_time set writes 3 plotfiles and regrids level 0 twice; with the fix, 2 plotfiles and one regrid.
• HIT init_turbulent_forcing: unchecked nmodes*step overruns tmp; async H2D copy unsynced #185 -- turb.nmodes=40 now trips the assertion instead of writing ~264 KB past the end of the stack array.
• HIT fast-force interpolation reads past ff_force and frees it while the kernel runs #184 -- HIT runs correctly at turb.ff_factor=6 (a remainder that was previously out of bounds), and tracks the exact-forcing build.
• dermgvort: EB one-sided vorticity stencil reads covered cells two away in thin gaps #179 -- EB mag_vort is sane in plotfiles (max 2619 in 2D, 2199 in 3D), no 1e40 values.
• getForceVerbose min/max indexes State/force by absolute component, reads past FAB end #182 -- with ns.getForceVerbose=1, scalar calls print State comp 0/1 from the velocity-only FAB instead of reading past it.
• MacProj: known-edgestate sync overload segfaults, sync divu omits dsdt, umac test on host #189 -- mac_proj.check_umac_periodicity=1 passes on a periodic 3D case under MPI.

Plotfile comparisons vs. baseline at step 4: poiseuille, traceradvect_bds, eb_double_shear_layer and eb_flow_past_cylinder are bit-identical. ns_hotspot / ns_hotspot_rz differ at 1e-12 / 1e-9 relative, from the dsdt term in #189 (tiny here because temp_cond_coef = 1e-8). eb_hotspot differs at ~0.2%, the intended #190 change described above.

Nothing here was run on GPU -- the GPU-specific parts of #184, #185 and #189 are compile-checked and reasoned, not executed on device.

🤖 Generated with Claude Code