nx_wgsl_gate.nx

nx_wgsl_gate.nx -- IS THE SHADER BACKEND REAL, OR A RUBBER STAMP? SUBJECT: ./nx_wgsl.elf by default, or ./nx_wgsl.sov.elf.new with `staged`. Both modes execute the actual binary; e2e fork, never an in-process re-derive -- an in-process gate would prove the SOURCE compiles, not that the SHIPPED artifact behaves). THE PRE-DECLARED ACCEPT RULE (written before the backend was built): the SAME NishiLang source compiled to BOTH dialects must produce the SAME RESULT. On this box that cannot be proven at the pixel level -- headless WebGPU wedges inside requestAdapter/createShaderModule (measured; it is why nx_game_page_emit carries a GPU_BOOT_MS race at all) -- so equivalence is asserted at the EMITTED-IR level and the pixel gap is DECLARED, never papered over: * TRACE IDENTITY: each backend records the node ids it emits, in order. Two backends that consumed the same source completely produce identical traces. * NON-VACUITY: a module with one statement removed must NOT trace the same. Without this leg trace identity would be satisfied by two backends that both emit nothing. * REFUSAL: a construct outside the covered subset must refuse BY NAME. A silent mistranslation in a shader backend compiles, links, runs, and draws the wrong picture. * DIALECT SEPARATION: the GLSL output must carry GLSL's own spelling and the WGSL output WGSL's -- proving each backend emitted its dialect rather than copying the other. * SHIPPING EQUIVALENCE: the emitted GLSL must carry the same operators and operands as the hand-written VSH that ships in nx_game_page_emit today, which is the migration reference. * DECLARATION REFUSALS (S1, 2026-09-04): a texture-typed uniform, a K_ATTRIB and a fragment K_VARY each REFUSE BY NAME in WGSL (they used to be emitted wrong with no refusal), and a scalar uniform is still ADMITTED by the same pass -- the positive control a deny-guard must carry. * UNIFORM QUALIFICATION (S2, 2026-09-04): WGSL reads a uniform as `u.<name>` (it lives in struct U), GLSL reads it bare; a parameter or local of the same name shadows the uniform and stays bare in both. Trace identity must survive a uniform being in play. * TEXTURES ARE THEIR OWN KIND (S3, 2026-09-04): K_TEXTURE emits a @group/@binding var OUTSIDE struct U in WGSL and a sampler uniform in GLSL, read bare in both; the deprecated texture-typed K_UNIFORM refuses in BOTH dialects naming K_TEXTURE; texture_2d<f32> and a binding collision refuse by name. * DERIVED FRAGMENT SIGNATURE (S4, 2026-09-04): a sir_param_position parameter yields `@fragment fn main(@builtin(position) pos:vec4f)->@location(0) vec4f{` and GLSL reads it as gl_FragCoord; zero parameters derive to the old literal exactly; a position READ with no declared parameter and a plain (unbound) fragment parameter refuse by name -- an undeclared input is never a placeholder. WHAT THIS GATE DOES NOT MEASURE, STATED SO IT CANNOT BE MISREAD: no tooth renders a frame. GREEN here means the two dialects are emitted from one source and agree structurally. It does NOT mean a GPU drew the same pixels from both. That is nx_wgsl_pixel_gate, and it is OWED.