nx_wgsl_front_candidate_t217.nx

nx_wgsl.nx -- THE SHADER BACKENDS: NishiLang in, GLSL and WGSL out. CONTRACT SYMBOL: wgsl_emit_module (declared _ABSENT_ on the graphics compare board until now). THE SOURCE IS NISHILANG, NOT GLSL (operator 2026-08-23). This organ is NOT a GLSL->WGSL transpiler. A shader is authored ONCE against the nx_shader_ir representation; glsl_emit_module and wgsl_emit_module are two BACKENDS over that one source, exactly as nx_compile_x86 and the WAT lane are two backends over one NishiLang front-end. Neither dialect is authored, so neither can drift from the other -- disagreement is impossible BY CONSTRUCTION rather than by discipline. RUNG 1, SHIPPED HERE: the fullscreen-triangle VERTEX STAGE. It is not a toy -- it is the exact stage that ships TODAY, hand-written TWICE in nx_game_page_emit: GLSL VSH : void main(){vec2 v=vec2((gl_VertexID<<1)&2,gl_VertexID&2);gl_Position=vec4(v*2.-1.,0.,1.);} WGSL vs : @vertex fn vs(@builtin(vertex_index) vi:u32)->@builtin(position) vec4f{...} Those two hand copies are the duplication this rung deletes: below, ONE NishiLang source (shsrc_fullscreen_tri) emits both. WHY THIS STAGE PROVES THE THESIS. The two dialects are not spelling variants of each other: * GLSL ASSIGNS gl_Position; WGSL RETURNS @builtin(position). Different STRUCTURE. * GLSL gl_VertexID is i32; WGSL vertex_index is u32 and must be converted before i32 maths. * WGSL requires a u32 shift amount (1u); GLSL takes a plain int. * WGSL requires 0.0, never 0. ; GLSL accepts either. Every one of those is dialect knowledge the BACKEND owns. A translator would have to rediscover them from GLSL text; a backend simply knows them, once. COVERED SUBSET (rung 1) -- anything outside REFUSES BY NAME via sir_refuse and returns -1. A NAMED REFUSAL IS A CONTRACT; A SILENT MISTRANSLATION IN A SHADER BACKEND IS THE WORST AVAILABLE FAILURE: it compiles, it links, it runs, and it draws the wrong picture forever. decls : uniform (scalar + array); attribute and varying are GLSL-covered and, since S1 (2026-09-04), WGSL-REFUSED BY NAME (K_ATTRIB anywhere, K_VARY in a fragment stage) until the entry-parameter derivation that WGSL needs for them lands; a texture-typed uniform refuses in WGSL too -- it was being emitted INSIDE struct U S3 (2026-09-04): a TEXTURE is its own kind (K_TEXTURE / sir_texture) in BOTH backends: GLSL `uniform highp usampler3D`, WGSL a @group/@binding var OUTSIDE struct U with its binding DERIVED from declaration order; only texture_3d<u32> is covered, other texture types refuse S4 (2026-09-04): the FRAGMENT entry signature is DERIVED from the declared parameters (no longer a literal): a sir_param_position parameter is `@builtin(position) name:vec4f` in WGSL and gl_FragCoord in GLSL; a position READ with no such parameter, and any entry parameter a backend cannot bind (plain fragment params, any vertex/compute param), refuse by name stmts : var, assign, return, if/else, discard