So instead of putting the position at the center of the earth we're putting it at 1,1,1 and that fixes it? Seems kind of odd to me and maybe masking a bigger issue?

Random idea, is the problem here that positionEC is actually a vec4 and not a vec3 and something like:

positionEC.xyz = vec3(0.0);
positionEC.w = 1.0;

Is the more correct fix?

My thinking is that positionEC.xyz = vec3(1.0); puts the point behind the camera since negative Z is forward in eye space. It should get clipped afterwards in the pipeline.

Probably a clearer fix is to do positionEC.xyz = vec3(0.0, 0.0, 1.0); and have an inline comment explaining why.

What's the range for eye coordinates again? Is 1.0 on the near plane? Maybe use 2.0?

Makes sense, I was reading poistionEC as earth-centric, now eye coordinates. (forgetting we use WC for world coordinates)

It's still a little strange that positionEC.xyz = vec3(0.0); doesn't work. That should be putting the point behind the near plane. I just noticed that the bug only happens when log depth is true. And it doesn't matter if I change the log-depth near plane default from 0.1 to 1.0.

Still, positionEC.xyz = vec3(0.0, 0.0, 1.0); should be safe because it is 1 meter behind the camera, and either 1.1 or 2.0 meters behind the near plane (for log-depth and multi-frustum respectively).