nx_gsplat4d_lib.nx

nx_gsplat4d_lib.nx -- ★4D SPATIOTEMPORAL GAUSSIAN SPLATTING: the temporal axis the estate did not have. WHAT WAS ALREADY HERE (measured 2026-08-14 three independent ways: nx_capsearch over 1033 registered tools, a coverage-complete grep of 23,023 sources, and reading the source): nx_gsplat is a real, gated, sovereign 3D splatter -- isotropic blobs AND anisotropic surfels whose 2D covariance Sigma' is computed EXACTLY by projecting the principal axes through the camera (the Jacobian, applied as a finite projection rather than a symbolic J R S S^T R^T J^T), a conic power test against an integer exp-LUT, a counting sort by depth, and front-to-back compositing C = sum c_i a_i prod(1-a_j). That is the rasterizer. It is NOT re-implemented here. WHAT WAS ABSENT: time. Every Gaussian in the estate was static. This organ adds the 4th axis and nothing else -- it deforms state and hands the result to gs_render_aniso. One rasterizer, still. THE KERNEL is the standard 4DGS temporal radial basis, integer-exact: x_t = x_0 + v * (t - t_mu) * exp(-(t - t_mu)^2 / (2 * t_sigma^2)) ★AND THE ENVELOPE COSTS NOTHING NEW: the incumbent's exp-LUT is explut[k] = GFXA*exp(-k/(2*GLUTU)), so indexing it at k = GLUTU * dt^2/sigma^2 yields exp(-dt^2/(2 sigma^2)) EXACTLY -- the Gaussian envelope, from the table the renderer already builds. A second exp table would have been a duplicate ruler. ★★NO RE-DECLARED CONSTANTS. The fixed-point scale, the LUT resolution and the record stride are FACTS OF THE RASTERIZER and are read from it (gs_fxa / gs_lutu / gs_stride_aniso). An earlier revision of this file hardcoded 256, 16 and 12; nx_magic at threshold 2 found them. They are not cosmetic: each was a SECOND RULER, and the failure mode of a second ruler is that the two disagree silently, compile cleanly, and render something plausible. There is no symptom to notice. ★★THE ARENA COUNTS ITSELF. Every allocation goes through g4_alloc, which accumulates the byte total as a side effect of allocating -- so g4_bytes() cannot drift from reality. The previous revision multiplied a HAND-WRITTEN COUNT of the field arrays; adding a field and forgetting to bump it would have silently under-reported the resource envelope, which is a fabricated number wearing the shape of a measurement. LAYOUT: Structure-of-Arrays, one array per field -- a field sweep touches contiguous memory instead of striding over a record. Arrays are lazily allocated once against a declared capacity. ⚠DELIBERATE, DECLARED DIFFERENCES FROM A LITERAL FIELD-LIST TRANSCRIPTION -- each is a defect avoided: - POSITION uses the signed ramp x envelope above (the kernel as specified). - RADIUS uses the envelope ALONE, not the ramp. A signed ramp on a radius crosses zero and inverts the splat, which the rasterizer cannot represent. A breathing surfel is the honest analogue. - NORMAL uses ramp x envelope and is then RENORMALISED, so the surfel tips without its length drifting. It is also the G-buffer normal the lighting rung will read. - QUATERNION and 3-axis SCALE are NOT stored. The incumbent surfel is a DISK (normal + in-plane