fix: CAPT false negatives during affordance propagation

[nrkumar93]

Summary

Fix two defects in CAPT::subdivide that can omit points from leaf affordance buffers. The omissions cause both CAPT::collides and CAPT::collides_simd to return false negatives for point-cloud collision queries.

A motion planner using a CAPT-backed environment could consequently accept a configuration where a query sphere overlaps a stored point.

Primitive collision geometry is unaffected.

Root cause

After partitioning a cell, each child must inherit points near the split plane from the opposite half.

Two issues made this propagation incomplete:

1. The low half was scanned from the wrong end.

   Each half is sorted along the split dimension. The high child must inherit the low half's suffix, the points closest to the split, but the previous implementation scanned and copied from the low half's beginning.

2. The propagation band excluded the point radius.

   The required distance is r_max + r_point, but the propagation checks used only r_max. This omitted points that the later Volume::affords check would otherwise accept.

Fix

• Use the existing max_affordance_l1 value (r_max + r_point) as the propagation band.
• Walk backward from the split to find the low half's near-plane suffix.
• Continue walking forward from the split to find the high half's near-plane prefix.
• Copy those two ranges into the corresponding child affordance buffers.
• Preserve the existing finite-value checks used to exclude padding points.

Both changes are required. Testing the changes separately still produced false negatives:

Variant	False negatives
Unpatched	1972
Band fix only	1633
Scan-direction fix only	69
Both fixes	0

Verification

Added CAPT regression coverage comparing tree results against a brute-force sphere/point overlap test for:

• a uniform random point cloud;
• a regular grid with points on median split planes;
• r_point = 0, isolating the propagation-direction defect;
• small clouds containing 1, 2, 3, 5, and 8 points.

Results after the fix:

• 0 false negatives and 0 false positives across approximately 110,000 queries;
• scalar and SIMD query paths agree;
• CAPT::is_valid() holds for every tested tree;
• all existing Catch2 tests pass;
• the single-point-tree edge case remains valid.

Before the fix, missed overlaps included deep penetrations—not only floating-point boundary cases. The worst observed missed overlap depth was approximately 10 cm.

Performance impact

Correct propagation stores more affordance entries and therefore increases CAPT construction time and memory use. For a 12³ grid containing 1,728 points:

	Affordance vectors	Build time
Before	23,149	6.8 ms
After	69,579	16.2 ms

CAPT construction is a per-scene cost. Query throughput was unchanged to slightly improved in the measured workload.

[nrkumar93]

Follow-up observation

This fix assumes queries respect CAPT's documented [r_min, r_max] radius range. Generated robot FK code may query the environment with link-culling spheres larger than Robot::max_radius, so callers should ensure that the CAPT r_max bounds the largest environment-query radius, not merely the largest leaf collision sphere.

This is independent of the affordance-propagation defects fixed here and can be addressed separately through documentation or generated robot metadata.