[adrian-andrzjewski]

Summary

• Allow scale-relative rounding tolerance when ray slab entry and exit distances nearly coincide, preventing false misses on zero-extent BoundingBox instances.
• Keep parallel-slab checks exact so rays outside a flat box's plane remain misses.
• Add regressions for both clipping directions, a nearby miss, and a parallel ray outside the plane.

Fixes #2556.

Validation

• gradlew.bat :jme3-core:test passed on Windows.
• The corner case failed against the original implementation (0 collisions) and passes with the fix (1 collision).
• The issue's TeaVM-specific path was not built; the same miss reproduces in the JVM regression.

[adrian-andrzjewski]

@riccardobl @pspeed42 — could one of you review this fix for #2556 when you have a chance? It adds a narrow scale-relative tolerance for nearly coincident slab distances, with regressions for flat-box corner hits and out-of-plane misses. Thanks.

[jaime-jmebot]

This also relaxes the ray-origin lower bound t[0] = 0, so a ray that starts just outside a slab and travels away from the box can be reported as a hit: its exit parameter is slightly negative, passes this tolerance check, and the collision code then returns a contact at the origin. The tolerance should only bridge rounding between slab clipping distances, not admit an interval entirely behind the ray. Could you keep the zero-time bound exact?

Suggested change

	if (newT < t[0] - clipTolerance(newT, t[0])) {
	if (newT < t[0] && (t[0] == 0.0f || newT < t[0] - clipTolerance(newT, t[0]))) {

[jaime-jmebot]

Hi, thanks for tackling this corner case and adding focused regressions. The change is nicely scoped, and keeping the parallel-slab test exact is important for preserving misses on flat boxes. I reviewed both ray-collision paths, which share this clipping helper, and the added tests cover the reported rounding issue from both clipping directions.

I did find one edge case to address before merging: the tolerance in the negative-denominator branch also applies when t[0] is still the ray's initial lower bound of zero. That can accept a slightly negative exit distance for a ray that starts just outside a box and points away from it; the collision path then treats the changed clipping interval as a hit, even though the entire interval is behind the ray origin. The inline comment suggests keeping that zero-time bound exact while still allowing tolerance between actual slab entry and exit distances.

Once that is addressed, this looks close. Thanks again for the targeted regression coverage!