Close up no longer falls apart: how near the nearest thing is goes by each object's box (not a ball round it: inside a big or long flat one's ball it came to nothing, the near plane with it, and surfaces fought each other in blocks, jittering, the sky and grid gone), the near plane a quarter of the way to it; reference pictures count too. Grid lines meet exactly where their pieces join (no crawling dots): a streak's ends each face the eye, each its own width; pieces a quarter further from the eye each. The depth pre-pass waited on before the scene uses it; the same depth in both (invariant)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

engine/shaders/mesh.vert

@@ -5,6 +5,9 @@ layout(location = 0) out vec3 v_normal;
 layout(location = 1) out vec2 v_uv;
 layout(location = 2) out vec2 v_lightmapUv;
 layout(location = 3) out vec3 v_world;
+// The same depth in the depth pre-pass and the scene pass (else the compiler, dropping what the pre-pass
+// doesn't use, can round it differently and the scene pass fails its depth test in patches).
+invariant gl_Position;
 
 void main() {
     Vertex vx = VERTEX(VERTEX_ID);