Hybrid 0th-order creation: clipped boundary cells, and cells that straddle a membrane

[jcschaff]

Fixes #24 and #26. Both are in the same loop: the OPTION_VCELL branch of zeroreact in smoldyn-2.38/source/Smoldyn/smolreact.c, which handles field-dependent 0th-order creation in a compartment (reaction_cmpt c r 0 -> A k*B;). The loop moved into a helper, zeroreactcmptmesh.

What changed

• Hybrid 0th-order particle creation uses full cell volume at boundary nodes (overproduces by Σfull/Σactual) #24, boundary cells. Nodes on the domain walls own half, quarter or eighth cells. The Poisson count now uses each node's cell clipped to the domain, the same region randomPosInMesh fills, instead of the full dx·dy·dz.
• Hybrid 0th-order creation in a compartment skips cells whose centre is outside (under-creates at curved membranes) #26, cells that straddle a membrane. A cell now draws if its node or any neighbouring node (face, edge or corner) is inside the compartment. The existing posincompart thinning then keeps exactly the molecules in cell ∩ compartment. Previously a cell drew only if its centre was inside, so straddling cells with the centre outside never created.
• Where the rate is evaluated. For a straddling cell whose own node is outside, the rate comes from the nearest neighbouring node inside. The field at the outside node belongs to another region, so evaluating it there would read the wrong value.
• Wall nodes. Membership is tested with each node nudged just inside the domain, because posincompart is ambiguous for points exactly on a wall.

Measured. Native runs through libvcell → pyvcell-fvsolver, 8 seeds each, B a static catalyst, comparing created A with the exact expectation. Before: the stock 0.10.7 wheel, numbers from the issues. After: this branch, built locally.

case	before	after
11×11×3 slab, box walls only (#24)	1.813	0.995 ± 0.004
ball R = 4 µm on a 37³ mesh, relative to the volume of Smoldyn's compartment (#26)	≈0.978	0.996 ± 0.002

The remaining ~0.5% is a separate start-up artifact: Smoldyn skips 0th-order creation for about the first 2 of 200 steps.

Not covered

• I haven't run the C++ test suite locally; CI covers it.
• The cases were run from viva-pde-particle: Studies A3 two-way-exchange and B2b sphere-cosim-vs-native.

Related: #25 (particle binning at the same membranes) and #23 (a second solve in one process); separate PRs.

🤖 Generated with Claude Code