nx_craft_tile_gate.nx

nx_craft_tile_gate.nx -- GE30 REFEREE: the frame cut into row bands is BIT-IDENTICAL to the frame drawn by one thread, and a band writes NOTHING outside the rows it owns. WHY THIS SUBJECT NEEDS ITS OWN GATE. nx_wasm_craft_gate proves the WORLD (raycast ground truth, causal edits, determinism, the input edge contract) and nx_wasm_craft_vm_gate proves the SHIPPED WASM fits and paints inside the memory it declares. Neither can see the defect GE30 introduces: a frame assembled by N workers over ONE shared linear memory is correct only if the band pass writes disjoint pixels and reads no scratch its neighbour writes. The old renderer failed the second half by construction -- five wray outputs lived in ONE shared slot (O_ST 46-50), which is invisible with one thread and a race with two -- so the decomposition is the subject here. THE LOAD-BEARING TOOTH IS A FULL PIXEL COMPARE, NOT A DIGEST. Two fresh arenas, same seed, same camera: arena A drawn by render_impl (one thread, the shipped path), arena B drawn as N bands plus one overlay. Every word of the WRITTEN EXTENT must agree -- rw x rh, the compact buffer the one-pixel-per-ray law made the single owner of, never W*H, because a comparison over unwritten memory is a claim about uninitialised bytes. A tiling defect lives at BAND SEAMS, so the band counts include ones that do NOT divide the row count evenly, and a stride-sampling checksum is deliberately not used anywhere in this file. ANTI-VACUITY, STATED BECAUSE "IDENTICAL" IS THE EASIEST GREEN TO FAKE: two blank frames are also identical. So the gate first proves the frame is VARIED (many distinct colours) and that the overlay pass genuinely CHANGES pixels; only then does agreement mean anything. nx_craft_tile_gate license_tier: ORIGINAL No hw writes (Rule 26).