Struct PirlsProblem 

pub struct PirlsProblem<'a, X> {
    pub x: X,
    pub offset: ArrayView1<'a, f64>,
    pub y: ArrayView1<'a, f64>,
    pub priorweights: ArrayView1<'a, f64>,
    pub covariate_se: Option<ArrayView1<'a, f64>>,
    pub gaussian_fixed_cache: Option<&'a GaussianFixedCache>,
    pub glm_first_step_gram: Option<&'a Array2<f64>>,
}

Fields

x: X

offset: ArrayView1<'a, f64>

y: ArrayView1<'a, f64>

priorweights: ArrayView1<'a, f64>

covariate_se: Option<ArrayView1<'a, f64>>

gaussian_fixed_cache: Option<&'a GaussianFixedCache>

When set, the inner PLS solver reuses the precomputed XᵀWX and XᵀW(y − offset) in original coordinates instead of streaming the O(N·p²) GEMM and the O(N·p) matvec on every outer REML iteration.

Valid only when the family is Gaussian + Identity link, prior weights are constant across outer iterations (always true in the REML outer loop), no Firth bias reduction, and no inequality / lower-bound constraints (matching the existing Identity short-circuit at pirls.rs:6237). The penalty λ·S is still added per-λ on top of the cached XᵀWX.

glm_first_step_gram: Option<&'a Array2<f64>>

Frozen-weight first-Fisher-step data-fit Gram XᵀWX for a GLM design-moving ψ-trial (#1111 / #1033 mechanism (c)), in original (conditioned x_fit) coordinates. When set, the iterative GLM P-IRLS serves its FIRST Fisher-scoring iteration’s XᵀWX from this matrix instead of streaming the O(N·p²) weighted cross-product; every later iteration restreams the true moving W, so the converged β̂ is unchanged. Mutually distinct from gaussian_fixed_cache (which is the Gaussian-identity converged-objective short-circuit); this is the GLM first-step lane and never short-circuits the iteration count.