The batch admitted three node types BY NAME and ran one node per dispatch, whose only channel for a socket was the push-constant block. So a field feeding a batched geometry node -- Set Position's Offset, the ordinary way to displace anything -- had nowhere to go, and both it and its consumer fell to the CPU. It was not missing GLSL: every one of the twenty-one nodes involved emits GLSL, and GpuCompiler::compile_batch has always looped over a node LIST threading values through GlslEmitContext. The executor simply never handed it more than one node. Now a batch is a geometry node PLUS the run of value nodes feeding it, emitted into one shader and dispatched once. Three pieces: GpuBatch::fused_nodes the graph nodes in this shader, topological, geometry node last GlslEmitContext knows who is in the shader with it, so param() resolves a connected socket to the producer's LOCAL instead of a push-constant slot -- and only when that producer is actually emitted here _execute_gpu_batch value nodes contribute arithmetic and nothing else: no dispatch, no buffers, no CPU execution execution report before { MeshGrid: cpu, Position: cpu, SetPosition: cpu } after { MeshGrid: cpu, Position: gpu (fused), VectorMath: gpu (fused), SetPosition: gpu } CPU vs GPU worst |dp| = 0.0 MUTATION-CHECKED: making param() ignore is_emitted() reproduces the exact divergence this started from, worst |dp| = 3.191786289, and fails that one line while everything else stays green. ⚠ Two restrictions, both stated rather than discovered later: * CONTIGUOUS ONLY. A batch is a SPAN of the execution order and the main loop skips it by advancing the index, so a cone with an unrelated node interleaved cannot be expressed without changing that contract. Topological order groups a producer with its consumer in the ordinary case; when it does not, the value node stays on the CPU. Slower, never wrong. * EVERY CONSUMER MUST BE INSIDE THE BATCH. A fused node's CPU execute() never runs, so it writes no output socket -- absorbing one that something downstream also reads would hand that reader a silently empty value. The walk stops at the first candidate that fails this. Unmoved by the change: glsl_golden_gate 32/32, volume_mesh_gate 4/4, sdf_grid_gate 3/3, and the bridged mountain still bakes 48149 verts across y -20.46 .. -0.05. ⛔ A trap worth the line: `git checkout --` to undo a MUTATION also reverts the FIX in the same file when the fix is not yet committed. It took the param() change with it and the gate was rebuilt against a library missing the very thing it was proving. Check what the file contains, not that the mutation is gone. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>