nx_ir_cast_capacity_expanded_t206.nx

nx_shader_ir.nx -- THE SHADER IR: ONE source of truth, N dialect backends. WHY THIS EXISTS (measured 2026-08-23). The estate's shaders are hand-written GLSL string literals inside nx_game_page_emit. Drawing characters on the WebGPU tier therefore demanded a SECOND HAND COPY in WGSL -- ~600 lines (dual-quaternion skinning off a joint texture, instance uniform arrays, the voxel occlusion march, the procedural anatomy/eye/hair/skin paint) of the most actively-edited shader in the estate: three separate feature edits landed in it on the day this was written. Two artifacts that must agree, kept in agreement by discipline, is the exact shape the banked law forbids. This module is the construction that makes disagreement IMPOSSIBLE: neither dialect is authored, both are EMITTED from this representation. HOUSE SHAPE. This mirrors the NishiLang compiler itself -- lex_source -> parse_module -> opt_run -> {nx_compile_x86, nx_compile_wat}: one front-end representation, many backends. nx_glsl.nx and nx_wgsl.nx are backends over this IR, not transpilers bolted to a call site. THE SOURCE OF TRUTH IS NISHILANG (operator, 2026-08-23: "we dont want python or glsl or javascript or all that it should be nishi lang and nishi ecosystem from the first byte up each rung gaining all these capabilities"). GLSL IS A TARGET, NOT AN INPUT. This is NOT a GLSL->WGSL transpiler: translating would enshrine GLSL as the canonical language and make a foreign dialect the artifact we maintain. A shader is authored ONCE in NishiLang; GLSL and WGSL are both OUTPUTS, exactly as x86_64 and WAT are outputs of the same NishiLang front-end. The ~600 lines of GLSL in MVS/MFS are therefore a MIGRATION TARGET, not the source: each rung expresses more of them in NishiLang and PROVES the emitted GLSL equivalent to the hand-written GLSL that ships today. LADDER (each rung widens the covered subset; rung 1 is what this lane ships): rung 1 the fullscreen-triangle vertex stage -- authored in NishiLang, emitting BOTH the GLSL VSH and the WGSL vs that are hand-written and SHIPPING today, proven against them. rung 2 lower the real NishiLang AST (nx_tokenizer/nx_parse) into this IR, so a shader is an ordinary NishiLang function rather than builder calls -- the full 'target, not translator' shape, sharing the front-end with nx_compile_x86 and the WAT lane. rung 3 widen the subset to the MFS paint set; rung 4 the MVS skinning set; then the character pass becomes a compile target and the WebGPU abstention flips from owed to shipped. THE SOURCE OF TRUTH IS NISHILANG (operator, 2026-08-23: "we dont want python or glsl or javascript or all that it should be nishi lang and nishi ecosystem from the first byte up each rung gaining all these capabilities"). GLSL IS A TARGET, NOT AN INPUT. This is NOT a GLSL->WGSL transpiler: translating would enshrine GLSL as the canonical language and make a foreign dialect the artifact we maintain. A shader is authored ONCE in NishiLang; GLSL and WGSL are both OUTPUTS, exactly as x86_64 and WAT are outputs of the same NishiLang front-end. The ~600 lines of GLSL in MVS/MFS are therefore a MIGRATION TARGET, not the source: each rung expresses more of them in