Module cubic_cell 

GPU substrate for de-nested cubic-cell derivative moments.

This module is the shared GPU evaluator for the de-nested cubic transport kernel that currently lives in src/families/cubic_cell_kernel.rs. For each partition cell (left, right, c_0, c_1, c_2, c_3) it computes the derivative-moment vector

  M_k = ∫_{left}^{right} z^k · exp(-q(z)) dz,   k = 0..=max_degree,
  q(z) = 0.5 · (z² + η(z)²),
  η(z) = c_0 + c_1·z + c_2·z² + c_3·z³.

Three branches feed into the same device API:

• Affine (c_2 = c_3 = 0, finite interval): closed-form via the T_n(a,b) recurrence used by affine_anchor_moment_vector_into.
• Non-affine finite: fixed 384-point Gauss–Legendre on the cell.
• Affine tail: closed-form on a semi-infinite (or whole-line) interval.

This is distinct from src/gpu/cubic_bspline_moments.rs, which computes tensor B-spline cell moments. The two modules share neither math nor data layout: do not conflate them.

Layout

• [branch] — host-side branch classifier; mirrors cubic_cell_kernel::branch_cell + the semi-infinite tail logic of evaluate_cell_state_dispatched.
• [host_substrate] — CPU-resident implementation. Works on every platform and is the parity reference for the device kernel.
• [kernel_src] — NVRTC-compilable CUDA C++ source as Rust string constants (D9 / D15 / D21 specializations).
• [device] — Linux+CUDA dispatcher that compiles, launches, and gathers the NVRTC kernel for the NonAffineFinite bucket; Affine / AffineTail buckets stay on CPU until Stage-2.