Struct PenalizedVectorGlmInputs 

pub struct PenalizedVectorGlmInputs<'a> {
    pub design: ArrayView2<'a, f64>,
    pub y: ArrayView2<'a, f64>,
    pub penalty: ArrayView2<'a, f64>,
    pub lambdas: ArrayView1<'a, f64>,
    pub fisher_w_override: Option<ArrayView3<'a, f64>>,
    pub max_iter: usize,
    pub tol: f64,
    pub class_penalty_metric: ClassPenaltyMetric,
}

Inputs to fit_penalized_vector_glm.

M (the number of active outputs / linear-predictor columns) is taken from lambdas.len(); the engine validates it against the design and override shapes. The response y is passed verbatim to the VectorLikelihood adapter, which owns its own (N, ·) shape contract (binomial columns use K = M; multinomial one-hot uses K = M + 1), so the engine does not constrain its column count beyond y.nrows() == N.

Fields

design: ArrayView2<'a, f64>

Design matrix X ∈ ℝ^{N×P} (one row per observation, shared across every output column).

y: ArrayView2<'a, f64>

Response Y ∈ ℝ^{N×·}, interpreted by the VectorLikelihood.

penalty: ArrayView2<'a, f64>

Shared smoothing penalty S ∈ ℝ^{P×P} (symmetric, PSD).

lambdas: ArrayView1<'a, f64>

Per-output smoothing parameter λ_a, length M.

fisher_w_override: Option<ArrayView3<'a, f64>>

Optional per-row Fisher-block override, shape (N, M, M). When Some, it replaces the analytic VectorLikelihood::hess_block as the Newton curvature; the gradient/residual path stays analytic (issue #349). The adapter is responsible for any family-specific structural precondition on the block (e.g. zero off-diagonals for independent columns).

max_iter: usize

Maximum Newton iterations; recommend 50.

tol: f64

Relative-step convergence tolerance; recommend 1e-7.

class_penalty_metric: ClassPenaltyMetric

Class-space metric of the replicated penalty (#1587). Diagonal preserves the historical independent-per-output penalty; Centered selects the reference-symmetric softmax penalty (requires uniform lambdas). See ClassPenaltyMetric.