hyper-games #65 NEXT item 2. It asked for an audit of the two batch-path nodes' push constants, on the grounds that nothing gates that the index a node reads is the slot the buffer manager fills. The audit found four defects, and the framing was off in both directions: Simplex Noise is not a batch-path node at all, and Set Position -- which is -- was broken worse than either. THE ROOT CAUSE IS ONE SENTENCE. `pc.paramN` means a push-constant SLOT to compile_batch, whose shader template declares the block and whose executor refills it every dispatch. It means SOCKET N to compile_to_function, which has no block and textually replaced `pc.paramN` with socket N's default. Two conventions, one spelling, and every node that used both was wrong in the density path. Measured on the emitted text, 2026-09-06: Displace vec3 npos = vec3(-pos.x, pos.z, pos.y) * 1.0; float n = bp_perlin_fbm_3(npos, float(int(5.0)), 4.0, 0.5, ... pos += normal * n * 0.0; Every argument one socket early -- Strength read as Scale, Scale as Detail, Detail as Roughness -- and `pc.param0` is socket 0, which is GEOMETRY and has no float literal, so the displacement was multiplied by zero. A no-op. SimplexNoise vec3 nsp = pos * 0.0; ... ps_fbm_noise_3d(nsp.x, nsp.y, nsp.z, int(5.0), 4.0, pc.param3) * 0.5 + 0.5; The same shift, and `pc.param3` addressed a fourth socket the node has not got, so the substitution never reached it and it survived VERBATIM into a function where `pc` is not declared. No density graph containing this node has ever compiled, which is also why fixing it can regress nothing. It also read the shader's `pos` where execute() reads socket 0. SetPosition pos += vec3(0.0, 0.0, 0.0); Sockets 0, 1 and 2 are Geometry, Bool and Vector; none has a float literal, so all three slots came out 0.0. A no-op. AND ONE THAT WAS LIVE ON THE GPU. GraphExecutor packed Set Position's block with `read_vector(node_idx, 1)`, commented "offset vector (socket 1)". Its sockets have been Blender's -- Geometry, Selection, Position, Offset -- since they were matched to Blender's, so socket 1 is SELECTION and the offset is socket 3. Every GPU-batched Set Position displaced by a boolean and ignored Offset entirely. The CPU path is a different function, so nothing could see it. THE FIX IS TO STATE THE MAPPING ONCE. ProcNode::get_push_constant_sockets() returns the input sockets the batch block carries, in slot order. GlslEmitContext::param() walks it to spell `pc.paramN`, GraphExecutor walks it to pack the bytes, and the per-type `if` in the executor is gone. Float, Int and Bool take one slot, Vector takes three. A node with no map -- Simplex Noise -- gets in() everywhere and behaves as if none of this existed. The two paths must still emit DIFFERENT text, and that is not cosmetic: the batch pipeline is cached in the ExecutionPlan and set_input_default() emits no `changed`, so a literal baked there would freeze at the value the plan was built with. GlslEmitContext::push_constants says which path is being emitted, and it is the only thing a node has to know. Displace sockets 1,2,3,4,5 -> param0..param4, the text it always emitted SetPosition socket 3 (Offset) -> param0..param2, likewise SimplexNoise no map; in() on sockets 0, 1, 2 and its two properties baked A value the shader cannot see now sends the node to the CPU. The batch shader's only channel is the push-constant block, so a non-geometry input that is connected and not in the map is a value the dispatch would silently ignore -- Set Position's Selection and Position are exactly that, and the CPU applies `new = Selection ? (Position + Offset) : old` per element. Falling back is the only answer that is not a different function. GATED, in demo/glsl_golden_gate.gd, 14/14 -> 24/24: H no emitted `pc.` anywhere 30 GPU-capable nodes, swept by ClassDB I the map names sockets that exist, are not Geometry, and fit 7 floats J the density path reads the right socket, proved with values nothing else has: scale 22, detail 33, roughness 44, lacunarity 55, strength 11 K the batch path still spells SLOTS, in the declared order, with no default baked in Five mutations, every one caught by the case that should catch it: Displace density back to pc.paramN slots Displace socket order Displace map, sockets 1 and 2 swapped Displace batch slots SetPosition reads socket 1 (the old packer) 3 cases, incl. batch slots SimplexNoise reads `pos`, not socket 0 SimplexNoise socket order SetPosition map with Geometry socket 0 map sanity + batch slots ⚠ CASE H CATCHES THE LEAK, NOT THE SHIFT, and the sweep is what said so. Reverting Displace to `pc.param1`/`pc.param2` leaves H green -- both indices are inside the socket count, so the substitution consumes them and the text is merely reading the wrong sockets. Only Simplex Noise's `pc.param3` was ever visible to it. J is what catches a shift. The docstring claiming otherwise was written before the sweep and is corrected in the file. ⚠ THE CPU FALLBACK IS NOT GATED. It lives in the dispatch path and reaching it needs a RenderingDevice, which the golden gate has not got. It is verified by reading only, and case I asserting that Selection and Position are absent from the map is the nearest thing to a proof there is. ProcNodeGraph::compile_batch_glsl() exists so the gate can read the batch text at all -- it was previously reachable only through a GPU dispatch, which is how the executor's socket index came to be stale with nothing to notice. Unmoved: golden map_dc_erosion_density 44140 chars byte-identical, emitted-op parity 770/770 on the RTX 2070, all three targets build. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>