nx_glprobe_lib.nx

nx_glprobe_lib.nx -- THE GL-CONTEXT PROBE RULER: one emitted probe, one native classifier, N page emitters. WHY (2026-09-02, operator: "it says no webgl2"): /world/rigview refused with the bare string "NXA REFUSED: no webgl2" and /world/beach fell to its sovereign tier with a bare "no-webgl2-context" -- neither said WHOSE refusal it was. Measured on the operator's laptop the same hour: every process started AFTER a display-driver swap (Intel UMD 32.0.101.6629 -> .8991) created WebGL2 (a fresh headless Edge AND a fresh headless Waterfox, ANGLE on the new UMD; the served rigview rendered 14164 verts / 104 joints in that Edge), while the Waterfox that had stayed open since before the swap still MAPPED THE DELETED 6629 UMD in its GPU process (the module that faulted Windows Terminal five times that week) and refused. The pages were innocent and the refusal text could not say so. A refusal that does not name its cause sends the reader at the wrong subject (nx_refusal_shape_lib asks exactly this: does the refusal name its subject and its remedy?). ONE TABLE, TWO SURFACES. The browser must answer the question itself (no page can call the estate for a remedy while its GPU is refusing), so the emitted JS carries the state and remedy table -- and the SAME table is the native classifier below, so an MCP caller (nx_glprobe classify <gl2> <gl1> <gpu>) and the served page cannot disagree. nx_glprobe_gate proves the two surfaces carry identical strings and that no page emitter still ships a bare refusal, with a runtime-assembled bare refusal as the firing neg-control. WHAT THIS IS NOT: not a fallback renderer. It names a refusal; the sovereign tier that would make the refusal irrelevant is the gameengine contract (wc_mob_tex, wc_render_tiles), not this lib. license_tier: ORIGINAL No hw writes (Rule 26).