Canonical 20-point Gauss–Legendre nodes on [-1, 1] (Abramowitz & Stegun
25.4), tabulated to f64 precision. Used here for the Drezner–Wesolowsky
bivariate normal CDF representation — 20 points give >30-digit accuracy for
the smooth arcsin-transformed integrand, ensuring the BVN value is exact to
f64 precision for all (h, k, ρ) — and shared with the cubic-cell B-spline
moment parity gate in [crate::gpu_kernels::cubic_bspline_moments].
384-point Gauss–Legendre nodes, re-exported for the GPU cubic-cell kernel
(src/gpu/cubic_cell/kernel_src.rs) to embed as __constant__ device
memory. Linux-only because the kernel emitter is Linux-only.
Build a partition covering (-∞, +∞) with parameter-independent outer
bounds. Interior cells use the same finite-cell polynomial algebra.
The two tail cells are guaranteed affine (c2=c3=0) because both
deviations saturate to constants outside their knot support.
Maximum moment index (i.e. max_degree passed to
evaluate_cell_moments) required to evaluate
cell_first_derivative_from_moments(derivative_coefficients, moments).
Pointwise value of the cell second-derivative integrand
(∂²/∂r∂s) exp(-q(z))/2π at a single z, evaluated from the SAME
(r, s, rs) coefficient polynomials the moment reduction
cell_second_derivative_from_moments integrates:
Derivative-moment counterpart to evaluate_cell_moments_cached. Shares
the value-moment LRU by storing both moment kinds in a single
CachedCellMoments entry keyed on the cell fingerprint — derivative
insertions preserve any pre-existing value state and vice versa, so the
two callers never evict each other’s work.