nx_gsplat_perf.nx

nx_gsplat_perf.nx -- THE SPLAT RENDER BASELINE: a number for every performance claim to beat. WHY THIS ORGAN EXISTS. On 2026-08-23 the sovereign splat rasterizer's viewport stopped being a compile-time 512x384 and became a parameter. Every performance claim made after that moment needs a measured baseline, per stage, at the resolutions we actually intend to deliver. WITHOUT A BASELINE AN IMPROVEMENT CLAIM IS UNFALSIFIABLE -- so this organ ships before any optimisation does. WHAT IT REPORTS, per viewport: total frame microseconds, plus the microseconds of each of the four stages the renderer is COMPOSED FROM (project+sort, clear, blend, resolve), and the CONIC EVALUATION count -- the deterministic cost unit gs_blend_rect already returns. Two renderers can be compared in that currency without a stopwatch at all, which is what makes an optimisation provable on a loaded box. THE BLEND FIGURE IS DERIVED (total - project - clear - resolve), never measured directly: measuring it directly would require a second copy of the full-screen policy, and a second policy is a second ruler. The subtraction is stated here so nobody reads it as an independent measurement. TIMING IS THE MINIMUM OF PF_ITERS RUNS. A minimum is the least contaminated estimator of the true cost on a shared box -- the mean absorbs every interfering lane, and this box carries many. The spread (max-min) is printed beside it so the reader can see how noisy the sample was rather than trusting the minimum blindly. THE VIEWPORT LADDER IS NOT PICKED. Each rung is a resolution the estate already has a reason for: the historic cap it just escaped, the derived gallery target read from the capture wire's own recorded frame, and the 4K rung the beauty tier is aimed at.