nx_wgsl_compute_gate.nx

nx_wgsl_compute_gate.nx -- GE43 REFEREE: the compute stage of the ONE NishiLang shader backend emits EXACTLY the WGSL the kernel contract names, and the GLSL side REFUSES it BY NAME. SUBJECT: the DEPLOYED nx_wgsl.elf (end-to-end -- this gate forks the artifact a browser page would be built from, not the library). `nx_wgsl compute` prints the depth-min kernel through both backends; every tooth below is a KAT over those bytes. A KAT is the right shape here: a shader backend that drifts by one character produces a program that compiles and draws the wrong picture, so "contains the right idea" is not a bar -- the bytes are. WHY THE GLSL REFUSAL IS LOAD-BEARING: WebGL2 has no compute stage, no storage buffers and no atomics. A backend that silently dropped the storage declarations would emit a fragment program that compiles, links, runs and draws nothing forever. Two teeth pin the refusal (the STAGE refusal on a compute emit, the DECLARATION refusal when the same kernel is asked for as a fragment stage), so the refusal cannot degrade into a later accident and still read GREEN. CONTROLS: the rung-1 contract (fullscreen triangle, trace_equal=1, uncovered_rc=-1) must still hold on the same binary -- a compute edit that broke rung 1 must read RED here, not on some other board weeks later. neg-control-* teeth assert refusals and absences. nx_wgsl_compute_gate [subject-elf] default ./nx_wgsl.elf, then _offc/nx_wgsl.elf license_tier: ORIGINAL No hw writes (Rule 26).