initial overlaping sphere sweep with MTD fails inside generateEEMTD

[pkusilence]

We are using PhysX 4.1.2. Our QA reports a bug, that a sphere character get stuck in a geometry and I found that is caused by initial overlap with distance = 0 (exactly), and the normal is always (-1)*(sweepDir). The sphere looks just touching the flower bed (red line in image below), which is a triangle mesh, and the contact feature should be the edge.

I walked through the calculations inside physx::Gu::computeCapsule_TriangleMeshMTD, and at line 278:
hadContacts = calculateMTD(....)
it returns no contacts. So setupSweepHitForMTD report that sweepHit.normal = -unitDir;

Then I checked PCMCapsuleVsMeshContactGeneration::processTriangle, found that

1. generateContacts returns no contact, because the sphere center's project point on the triangle is not inside (con0/con1 is false)
2. generateEEContactsMTD also returns no contact, because for a sphere, PCMCapsuleVsMeshContactGeneration::generateEEMTD always return false at line 818, the comment says "If both points in the same side as the plane, no intersect points".

Now I just work around to check at first, if it is a sphere, just calculate contact using pcmDistancePointTriangleSquared.

	const FloatV sqInflatedRadius = FMul(inflatedRadius, inflatedRadius);

	// Sphere fast path: degenerate capsule (p0==p1). generateEEMTD always exits early for
	// this case because pq=(0,0,0) => npq=0, and generateContacts fails when the diagonal
	// projection lands outside the triangle face. Use direct point-triangle query instead.
	// W stores raw distance (consistent with generateContacts / generateEEMTD convention),
	// so the outer loop's distV = mtd - capsuleV.radius gives the correct signed penetration.
	const Vec3V capsulePQ = V3Sub(capsule.p1, capsule.p0);
	if(FAllEq(V3Dot(capsulePQ, capsulePQ), zero))
	{
		Vec3V closestP;
		bool generateContact = false, faceContact = false;
		const FloatV sqD = pcmDistancePointTriangleSquared(
			capsule.p0, p0, p1, p2, trigFlag, closestP, generateContact, faceContact);
		if(FAllGrtr(sqInflatedRadius, sqD) && generateContact)
		{
			const Vec3V contactNormal = faceContact
				? n
				: V3Normalize(V3Sub(capsule.p0, closestP));
			manifoldContacts[numContacts].mLocalPointA = capsule.p0;
			manifoldContacts[numContacts].mLocalPointB = closestP;
			manifoldContacts[numContacts].mLocalNormalPen =
				V4SetW(Vec4V_From_Vec3V(contactNormal), FSqrt(sqD));
			manifoldContacts[numContacts++].mFaceIndex = triangleIndex;
		}
		return true;
	}

	FloatV t, u, v;
	const FloatV sqDist = pcmDistanceSegmentTriangleSquared(capsule.p0, capsule.p1, p0, p1, p2, t, u, v);

But I'm not that sure, and wonder what is the best way to fix this.

[pkusilence]

Update:
When selectNormal return false, pointOnTriangle has v+w = 1+4.8×10⁻⁶ > 1 + FEps , results in con0 = isValidTriangleBarycentricCoord(v0, w0) => false, con1 also false.

#define FLT_EPSILON 1.192092896e-07F // smallest such that 1.0+FLT_EPSILON != 1.0

It is too small for a mesh that size is greater than 200m.