Why make clamp to (tx,ty) during computing cov2d? Why the backward grad dJ_dz does not change whether the clamp happens or not?

[archershot]

diff-gaussian-rasterization/cuda_rasterizer/backward.cu

Lines 168 to 176 in 59f5f77

	const float limx = 1.3f * tan_fovx;
	const float limy = 1.3f * tan_fovy;
	const float txtz = t.x / t.z;
	const float tytz = t.y / t.z;
	t.x = min(limx, max(-limx, txtz)) * t.z;
	t.y = min(limy, max(-limy, tytz)) * t.z;
	const float x_grad_mul = txtz < -limx || txtz > limx ? 0 : 1;
	const float y_grad_mul = tytz < -limy || tytz > limy ? 0 : 1;

Clamp (tx,ty) seems to make Gaussians closed to view space (ndc). And for remote huge Gaussians, it would contribute for rendering when its radius is larger than 0.3*tanfov. Why preprocess gaussian like this? Isn't that a bug?

diff-gaussian-rasterization/cuda_rasterizer/backward.cu

Lines 262 to 264 in 59f5f77

	float dL_dtx = x_grad_mul * -h_x * tz2 * dL_dJ02;
	float dL_dty = y_grad_mul * -h_y * tz2 * dL_dJ12;
	float dL_dtz = -h_x * tz2 * dL_dJ00 - h_y * tz2 * dL_dJ11 + (2 * h_x * t.x) * tz3 * dL_dJ02 + (2 * h_y * t.y) * tz3 * dL_dJ12;

And for the later grad computing of dL_dtz. It should be affected by the clamp, like clamp J[0,2] from "-fxtx/(tztz)" to "-fx/tz*clamp(fov)". But the above implementation ignore the clamp for dL_dtz.