Two defects the bridged mountain found, one of them mine from the fusion. `n30_s0` UNDECLARED IDENTIFIER. compile_batch set ctx.emitted_nodes to the whole batch up front, so param() resolved a socket to a producer's local before that producer had emitted -- or when it had emitted NOTHING, because compile_glsl returned empty and the loop skips it. The consumer then named a local that does not exist and the whole shader failed to parse: ERROR: 0:46: 'n30_s0' : undeclared identifier WARNING: ProcNodes: GPU pipeline creation failed for SetPosition The list is built as the loop goes now, so a node can only name a producer that has already put its declaration above it. Execution order is topological, so a real producer is always there; anything else falls back to the push constant, which is what it did before the fusion existed. A NON-INDEXED SURFACE IS STILL TRIANGLES. ProcMesh::from_array_mesh copied ARRAY_INDEX only when present, so a surface without one produced a ProcMesh with positions and NO indices -- triangle_count() == 0, and every node downstream saw a mesh with no faces: proc-nodes: Mesh to SDF Grid got no triangles, so there is no surface proc-nodes: SDF Grid Fillet got no grid proc-nodes: Grid to Mesh got no grid ProcNodes [GPU=true]: execute=130.493ms to_array_mesh=0.0ms on a cage that had 60516 corners. blender-bridge builds exactly that shape: Blender exports per-loop-triangle CORNERS so modifier normals survive, and never writes an index array. For a non-indexed triangle list the fan IS the index array -- 0, 1, 2, ... -- and it is synthesised now. glsl_golden_gate 32/32 sdf_grid_gate 3/3 volume_mesh_gate 4/4 gpu_batch_gate 6/6 Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>