nx_gsplat_tile_lib.nx

nx_gsplat_tile_lib.nx -- ★TILED RASTERIZATION: a second POLICY over the SAME three rasterizer stages, not a second rasterizer. It calls gs_project_sort, gs_blend_rect and gs_resolve -- the identical projection, the identical conic blend, the identical resolve that the full-screen path uses. Nothing about the maths is restated here, so the two renderers CANNOT drift. ★★THE ACCEPT RULE, DECLARED BEFORE THE EXPERIMENT (and enforced by nx_gsplat_tile_gate): CORRECTNESS -- the tiled frame must be BYTE-IDENTICAL to the full-screen frame. Not "close", not "within tolerance": identical. This optimisation is only allowed to change COST. COST -- it must perform STRICTLY FEWER conic evaluations on a scene with occlusion. If either fails, it does not ship. A tiled renderer that is merely as fast, or that differs by a rounding nobody can see, is not an improvement -- it is a second ruler. WHY BYTE-IDENTITY IS ACHIEVABLE RATHER THAN LUCKY: every pixel belongs to exactly one tile; within a tile the splats are visited in the SAME global depth order (the list is built by walking the sorted order and appending); and the early-out skips a splat only when EVERY pixel of the tile is already saturated -- pixels the blend would have skipped anyway on its own `tr > 1` test. So no contribution is dropped and none is reordered. The saving is work that provably could not have changed a pixel. ⚠WHAT IS AND IS NOT CLAIMED: this is a serial CPU renderer. Tiling here buys OCCLUSION SKIPPING and memory locality; it does NOT buy parallelism, because there is no thread or GPU to spread tiles across. The parallel dispatch that tiling exists for on a GPU is a NAMED GAP with a named blocker. license_tier: ORIGINAL No hw writes (Rule 26).