When Lz < 4·rMaxLimit, the stencil cannot reach every reacting pair

[yingyue2030699]

When Lz < 4 · rMaxLimit, the stencil cannot reach every reacting pair

z = std::max(4, ...) forces four z-cells regardless of the box thickness, and check_dimensions() reapplies the same floor rather than correcting it.

Below Lz = 4 · rMaxLimit, the cells become thinner than the interaction range, and the ±1 stencil no longer covers the full range. Pairs separated by two cells along z are therefore never passed to check_bimolecular_reactions().

Measured using the rev_3D model—with 2,000 molecules, rMaxLimit = 16.70 nm, 1,000 steps, and 19 snapshots—while varying the box thickness and counting every pair by brute force:

Box Lz	cells_z	Z-cell size	rMaxLimit	Pairs in range	Reached by stencil	Missed
30 nm	4	7.50 nm	16.70 nm	29,123	26,951	2,172 (7.5%)
50 nm	4	12.50 nm	16.70 nm	20,475	20,391	84 (0.41%)
100 nm	5	20.00 nm	16.70 nm	11,350	11,350	0
940 nm	30	31.33 nm	16.70 nm	1,328	1,328	0

The boundary is exactly where the arithmetic predicts:

• Zero missed pairs whenever the z-cell size is greater than or equal to rMaxLimit.
• Nonzero missed pairs as soon as the four-cell floor forces the z-cell size below rMaxLimit.

This produces a systematic bias in the association rate, not a statistical fluctuation, and it occurs silently.

Origin

This issue is pre-existing. The implementations in Dimensions() and check_dimensions() are byte-identical to those on master; the nerdss-optimized branch neither introduced nor fixed the problem.

[yingyue2030699]

Suggested fix

1. Drop the max(4, …) floor; assert cell size ≥ rMaxLimit
   Derive each dimension from max(1, floor(L/rMaxLimit)) and assert the resulting edge is not below the cutoff. Fixes a silent correctness bug worth up to 7.5% of reacting pairs in thin boxes.

2. Remove the maxPairs downscaler, or make it isotropic
   It produced 125 nm cells against a 28 nm cutoff and 98× over-inclusion. Its stated purpose — keep cell overhead below N²/2 — is better served by fix 4, which makes that overhead O(occupied) instead of O(total). N is also read once at setup and never revisited as the system grows.