gam_models/

multinomial.rs

//! Penalized multinomial-logit (softmax) GLM driver — fixed-λ inner solve.
//!
//! This is the principled vector-response companion to the scalar PIRLS path:
//! the inner-loop Newton solver for a multi-class GAM at fixed smoothing
//! parameters λ, using the canonical multinomial-logit likelihood
//! ([`MultinomialLogitLikelihood`]) and the existing dense block-Fisher
//! assembly in [`gam_solve::pirls::dense_block_xtwx`] /
//! [`gam_solve::pirls::dense_block_xtwy`].
//!
//! # What this module does
//!
//! Solve, for the reference-coded multinomial-logit GAM with `K` classes and
//! design matrix `X ∈ ℝ^{N×P}`,
//!
//! ```text
//!     β̂ = argmin_β { − log L(β) + ½ Σ_{a=0}^{K-2} λ_a · β_a^T S β_a }
//! ```
//!
//! where `β = [β_0; β_1; …; β_{K-2}]` is the stacked coefficient vector in
//! output-major order (`β_a ∈ ℝ^P` is the coefficient block for class `a`),
//! `S ∈ ℝ^{P×P}` is the smoothing penalty matrix (shared across classes,
//! replicated as `I_{K-1} ⊗ S` over the full parameter space), and `λ_a` is
//! a per-class smoothing parameter.
//!
//! The likelihood uses class `K - 1` as the reference (`η_{K-1} ≡ 0`), so the
//! softmax gauge is fixed at the η level and no additional sum-to-zero
//! projection is required.
//!
//! # Layering
//!
//! * **Fixed-λ inner solve** — [`fit_penalized_multinomial`] is the canonical
//!   coefficient-space Newton solver at *given* smoothing parameters `λ`,
//!   built on the shared [`crate::penalized_vector_glm`] engine.
//!
//! * **REML / LAML smoothing-parameter selection** — [`fit_penalized_multinomial_formula`]
//!   routes through [`crate::custom_family::fit_custom_family_with_rho_prior`]
//!   so the per-active-class `λ_a` are selected by the outer REML/LAML loop;
//!   the caller's `init_lambda` is only a warm-start seed. The multinomial
//!   [`crate::multinomial_reml::MultinomialFamily`] `CustomFamily`
//!   impl calls the fixed-λ math above as its inner solve at each ρ trial and
//!   supplies the dense per-row Hessian block for the outer trace terms.
//!
//! * **Formula → design integration** — `build_formula_design_for_multinomial`
//!   parses the Wilkinson formula and assembles `X` and the per-term `S`
//!   blocks; the `fit_multinomial_formula_pyfunc` FFI shim wires the Python
//!   `gamfit.fit(..., family='multinomial')` entry straight to this path.
//!
//! # Convergence
//!
//! The damped-Newton-with-backtracking scaffold lives once in the shared
//! [`crate::penalized_vector_glm`] engine: at each iteration the
//! assembled penalized Hessian `H + I_{K-1} ⊗ (λ_a S)` is factored via faer's
//! symmetric-PD-with-fallback path, the full Newton step `δ = −H^{-1} ∇F` is
//! computed, and accepted with step halving if the objective fails to decrease
//! (up to a small backtracking budget). The convergence test is the relative
//! coefficient step norm `‖δ‖ / (1 + ‖β‖) ≤ tol`, matching the existing pyffi
//! reference path. This module is the softmax adapter over that engine: it
//! supplies the dense `(K-1)×(K-1)` Fisher block, the residual, and the
//! log-likelihood through [`MultinomialLogitLikelihood`], and owns the
//! class-count / simplex preconditions. The independent-binomial sibling
//! [`crate::binomial_multi`] is the same engine with a row-diagonal
//! Fisher block instead.

use crate::custom_family::{
    BlockwiseFitOptions, ParameterBlockState, PenaltyMatrix, fit_custom_family_with_rho_prior,
};
use crate::multinomial_reml::MultinomialFamily;
use crate::penalized_vector_glm::{PenalizedVectorGlmInputs, fit_penalized_vector_glm};
use crate::vector_response::{MultinomialLogitLikelihood, validate_multinomial_simplex};
use gam_terms::inference::formula_dsl::parse_formula;
use crate::model_types::EstimationError;
use crate::fit_orchestration::{
    FitConfig, build_termspec_with_geometry_and_overrides, resolved_resource_policy,
};
use gam_terms::smooth::{
    PenaltyBlockInfo, TermCollectionDesign, TermCollectionSpec, build_term_collection_design,
};
use crate::fit_orchestration::drivers::freeze_term_collection_from_design;
use gam_terms::term_builder::resolve_role_col;
use gam_problem::ResponseColumnKind;
use gam_data::ColumnKindTag;
use gam_data::EncodedDataset;
use gam_runtime::resource::ProblemHints;
use ndarray::{Array1, Array2, ArrayView1, ArrayView2, ArrayView3};
use serde::{Deserialize, Serialize};
use std::sync::Arc;

/// Solver-only numerical stabilization floor for the formula-driven
/// multinomial REML inner solve (gam#747).
///
/// Installed with [`RidgePolicy::solver_only`](gam_problem::RidgePolicy::solver_only)
/// so it stabilizes the inner joint-Newton **linear solve** but never enters
/// the REML objective, the penalty log-determinant, or the Laplace Hessian.
///
/// What it does: the multinomial smoothing penalties are rank-deficient by
/// design (each smooth carries an unpenalized polynomial null space) and the
/// formula may add a fully unpenalized parametric term (`x3` / `body_mass`). On
/// near-separable hard labels the softmax curvature is ill-conditioned along
/// those directions, so the bare Newton step `H⁻¹∇` is huge. Lifting the
/// smallest Hessian eigenvalue to `δ` bounds the step (`‖(H+δI)⁻¹∇‖ ≤ ‖∇‖/δ`),
/// keeping the screening iterates finite without poisoning the softmax with
/// `inf − inf = NaN`.
///
/// What it deliberately does NOT do: it adds no `½·δ·‖β‖²` term to the
/// objective and no `δ`-shift to the REML log-determinant. The earlier
/// `explicit_stabilization_pospart` policy folded both into the criterion,
/// which made `1e-4` a fixed-λ Gaussian prior that shrank every identified
/// coefficient off the MLE and biased smoothing-parameter selection — a value
/// that had to be tuned *between* under-stabilization (NaN seeds) and
/// over-shrinkage (lost VGAM match). As a solver-only floor that tradeoff is
/// gone: the over-shrinkage failure mode cannot occur (nothing is shrunk), the
/// optimized objective is the true penalized REML criterion, and the floor
/// only has to be large enough to keep the linear algebra finite.
///
/// The separation defect (#753) is no longer this floor's job. If the
/// multinomial MLE is genuinely at infinity for an unpenalized/null-space
/// direction (complete/quasi-complete separation), no solver floor makes that
/// direction's estimate finite. The formula REML path arms the full-span
/// Jeffreys/Firth correction CONDITIONALLY — only on separation evidence (see
/// [`multinomial_formula_separation_evidence`] and the two-attempt logic in
/// [`fit_penalized_multinomial_formula`]) — so an interior, well-identified fit
/// optimizes the unbiased penalized-REML criterion with no Firth shrinkage
/// toward the uniform simplex, while a (quasi-)separated geometry gets the
/// proper prior that is the only thing able to bound its penalty-null
/// directions (#715 real-data arm). The bare fixed-λ inner driver
/// [`fit_penalized_multinomial`] (no outer REML, no Jeffreys term) surfaces the
/// explicit `MultinomialSeparationDetected` diagnostic for the path that has no
/// proper prior to lean on.
const MULTINOMIAL_FORMULA_RIDGE_FLOOR: f64 = 1.0e-4;

/// Inner joint-Newton KKT tolerance for the multinomial formula path.
///
/// The softmax Fisher weight `W = diag(p) − ppᵀ` collapses on saturated rows,
/// so near-separable fits (penguins, #715) reach the OBJECTIVE's f64 noise
/// floor before the default `inner_tol = 1e-6` KKT target: measured on the
/// penguins arm (standardized columns), the trust region collapses to 1e-12
/// with per-attempt objective changes of ~+2e-9 on |obj| ≈ 1e2 (≈ 1e-11
/// relative — pure rounding) while the KKT residual plateaus at 2.8e-5–9.4e-5
/// against a scaled tolerance of ~1.9e-5. Demanding a residual below the
/// floating-point noise floor is certifiable-never: every eval is rejected by
/// the stall guard and the whole fit fails. `1e-5` certifies the measured
/// plateaus while still resolving β to ~1e-6 in the relevant metric — the
/// LAML criterion consumes β̂ with error O(residual²/curvature), far below
/// any quantity the outer ρ-search can read.
const MULTINOMIAL_FORMULA_INNER_TOL: f64 = 1.0e-5;

/// Formula-adapter penalty calibration for multinomial softmax REML.
///
/// The term builder's normalized penalties are calibrated on single-response
/// Gaussian-style score curvature. A reference-coded softmax class block sees
/// per-row active-class Fisher diagonal `p_a(1-p_a)` plus negative cross-class
/// coupling. At the neutral simplex (`p_k = 1/K`) the active diagonal is
/// `(K-1)/K²`, so the binary-logit calibration is `2·(K-1)/K² = 1/2` and the
/// three-class calibration is `4/9` rather than the historical hard-coded
/// `1/2`. Making the scale a function of `K` keeps the physical smoothness
/// prior tied to the likelihood curvature instead of over-penalizing every
/// class as the simplex gains categories.
fn multinomial_formula_penalty_scale(n_classes: usize) -> f64 {
    let k = n_classes.max(2) as f64;
    2.0 * (k - 1.0) / (k * k)
}

/// Largest smoothing-parameter dimension where exact dense outer curvature is
/// still worth paying for multinomial formula fits.
///
/// `D = (K - 1) * n_penalties`. Medium-size loaded models use exact curvature
/// so the optimizer does not wander into over-smoothed lambda caps on
/// near-boundary softmax surfaces. The threshold was originally calibrated at
/// `D <= 6` when each `s()` term carried ONE penalty; the double-penalty
/// migration (wiggliness + null-space shrinkage per term, mgcv `select=TRUE`
/// semantics) doubled `D` for the SAME models, silently flipping the
/// reference formula fits (2 smooths, K = 3: old `D = 4`, now `D = 8`) onto
/// the gradient-only route — where the #715 quality arm showed every
/// wiggliness ρ driven onto the ±10 box bound (smooths collapsed toward their
/// polynomial null space, truth-RMSE behind VGAM). `12 = 2 × 6` preserves the
/// original classification boundary under the doubled penalty count while
/// keeping the four-smooth penguin species quality fixture on the exact ARC
/// path: that model is `D = 16`, and first-order BFGS can cycle along the
/// near-separable lambda-to-zero ridge until the wall-clock budget expires
/// (#1082). ARC observes the same exact curvature and can halt through the
/// bound-aware cost-stall guard once the REML surface stops making useful
/// progress.
const MULTINOMIAL_EXACT_OUTER_HESSIAN_MAX_DIM: usize = 16;

fn multinomial_formula_use_outer_hessian(total_rho_dim: usize) -> bool {
    total_rho_dim <= MULTINOMIAL_EXACT_OUTER_HESSIAN_MAX_DIM
}

/// Logit magnitude beyond which fitted probabilities are saturated at ordinary
/// double precision diagnostic scale. The bare fixed-λ driver has no outer REML
/// state and still uses this threshold to reject a non-converged saturated
/// iterate as a separation artifact. The formula REML path does not use this as
/// a Firth trigger: with smoothing parameters selected, a finite saturated
/// surface can be the valid near-separated optimum that should be scored
/// directly.
const MULTINOMIAL_SEPARATION_ETA_THRESHOLD: f64 = 25.0;

/// Calibrated convergence tolerance for the OUTER REML/LAML smoothing-parameter
/// search on the formula multinomial path. Matches the primary GLM REML outer
/// (`solver::fit_orchestration::materialize` uses `tol = 1e-7`, mirrored by the
/// `LOG_LAMBDA_TOL` / `KKT_TOL_*` constants across the REML stack): tight enough
/// that the selected λ reaches the genuine REML optimum (the recovered
/// probability surface matches the mature reference), loose enough that the
/// optimizer does not grind surface-irrelevant ρ digits down to the inner KKT
/// scale (the #1082 wall-clock overrun). The caller's `tol` is floored at this
/// value for the OUTER loop, while it continues to drive the INNER joint-Newton
/// KKT target unchanged.
const MULTINOMIAL_OUTER_REML_TOL: f64 = 1e-7;

/// The first multinomial formula solve is a separation probe: it is accepted
/// when the unbiased REML criterion converges to a finite interior iterate.
/// Near-separable data such as the penguin fixture otherwise spend the caller's
/// full outer budget on an iterate that is discarded before the Firth/Jeffreys
/// refit. Keep enough iterations for ordinary interior fits to certify quickly,
/// but hand slow/non-interior probes to the proper-prior refit promptly.
const MULTINOMIAL_UNBIASED_PROBE_OUTER_MAX_ITER: usize = 20;

/// Per-observation softmax Fisher-information scale for the λ-floor units.
///
/// The penalty enters the criterion as `½ λ βᵀ S β` with a Frobenius-normalized
/// `S` (`‖S‖_F = 1`, see the term-builder calibration referenced by
/// [`multinomial_formula_penalty_scale`]), so the ridge `λ S` is directly
/// comparable to data Fisher information. One observation contributes softmax
/// information `p(1−p)` in a class's logit direction, which is bounded by the
/// logistic peak `p(1−p) ≤ ¼` at `p = ½`. Using this maximal per-observation
/// information as the unit makes the floor's strength interpretable as a count
/// of equivalent **pseudo-observations** of prior: a ridge that equals
/// `τ · ¼ · ‖S‖_F` carries the same logit-direction curvature as `τ` real rows
/// sitting at the most-informative point of the likelihood. This scale is
/// `K`-independent on purpose — the `K`-dependence of the softmax block
/// curvature already lives in the penalty matrix via
/// [`multinomial_formula_penalty_scale`], so the floor (a bound on the
/// multiplier of that already-scaled penalty) must not double-count it.
const MULTINOMIAL_FORMULA_FISHER_INFO_PER_OBS: f64 = 0.25;

/// Target prior strength of the λ-floor, in pseudo-observations, for a
/// WELL-SUPPORTED class. The floor holds the unbiased REML optimizer off the
/// zero-penalty boundary (where a boundary-overfit smooth or a Firth switch on
/// finite data would otherwise be accepted) with a prior worth a fixed small
/// fraction of one observation. `8e-4` pseudo-observations reproduces the
/// previously fixture-calibrated large-support floor `τ · ¼ = 2e-4` exactly at
/// the calibration point, now expressed as an effective-prior-strength rather
/// than a tuned λ value.
const MULTINOMIAL_FORMULA_PRIOR_PSEUDO_OBS: f64 = 8.0e-4;

/// Reference class support `n_ref`: the effective sample size per class at which
/// the data Fisher information `n_c · I₁` is large enough that the floor sits at
/// its well-supported value. Below `n_ref` the per-class data information shrinks
/// like `n_c`, so to keep the floor's prior from vanishing *relative to* that
/// shrinking data the effective pseudo-observation count is scaled up by
/// `n_ref / n_c` (the prior is held to a fixed fraction of the data information,
/// not a fixed absolute λ). At `n_c = n_ref` the scale is exactly 1.
const MULTINOMIAL_FORMULA_SPARSE_REFERENCE_SUPPORT: f64 = 50.0;

/// Cap on the floor's prior strength in the very-sparse limit, in
/// pseudo-observations. As `n_c → 0` the `n_ref / n_c` scaling diverges; the cap
/// holds the prior at `4e-3` pseudo-observations (`τ_max · ¼ = 1e-3` at the
/// calibration point, the previously-tuned strong-floor value) so the floor
/// stays a proper prior rather than a hard constraint that would dominate the
/// likelihood for a handful-of-rows class.
const MULTINOMIAL_FORMULA_SPARSE_PRIOR_PSEUDO_OBS_MAX: f64 = 4.0e-3;

/// Continuous, Fisher-information-scaled lower λ floor for the formula path,
/// derived from the minority class's effective sample size `n_c`.
///
/// # Derivation (effective-prior-strength / Fisher geometry)
///
/// The penalty `½ λ βᵀ S β` with `‖S‖_F = 1` adds curvature `λ` to the class
/// logit direction; one observation adds at most `I₁ = ¼` there. So a floor that
/// sets `λ_floor = τ_eff · I₁` gives the smooth a prior worth `τ_eff`
/// pseudo-observations. We want a fixed *absolute* prior `τ` for a well-supported
/// class, but for a minority class with only `n_c` effective observations the
/// data information in its block is `n_c · I₁`; holding the prior to a fixed
/// *fraction* of that shrinking data information requires
///
/// ```text
///     τ_eff(n_c) = τ · max(1, n_ref / n_c),   clamped to [τ, τ_max]
///     λ_floor(n_c) = τ_eff(n_c) · I₁
/// ```
///
/// This is the *same* `base · max(1, c0/c)` envelope as before — but `base`,
/// `sparse`, and `c0` are no longer fixture-tuned magic numbers: `base = τ·I₁`,
/// `sparse = τ_max·I₁`, and `c0 = n_ref` are an effective-prior-strength of
/// `τ`/`τ_max` pseudo-observations against the maximal per-observation softmax
/// information `I₁ = ¼`. Properties preserved by construction:
///   * reduces EXACTLY to `τ·I₁` for well-supported classes (`n_c ≥ n_ref`);
///   * reduces EXACTLY to `τ_max·I₁` for very sparse classes
///     (`n_c ≤ n_ref·τ/τ_max`, here `n_c ≤ 10`);
///   * interpolates monotonically and continuously between them in the middle —
///     no cliff at `n_c = n_ref`.
/// At the calibration point the endpoints equal the previous `2e-4` / `1e-3`, so
/// fixtures whose smallest class has `n_c ≥ 50` (penguins, the vgam softmax
/// arms) are unaffected — they sit at `τ·I₁ = 2e-4` exactly as before.
fn multinomial_formula_min_lambda(y_one_hot: ArrayView2<'_, f64>) -> f64 {
    let base = MULTINOMIAL_FORMULA_PRIOR_PSEUDO_OBS * MULTINOMIAL_FORMULA_FISHER_INFO_PER_OBS;
    let sparse =
        MULTINOMIAL_FORMULA_SPARSE_PRIOR_PSEUDO_OBS_MAX * MULTINOMIAL_FORMULA_FISHER_INFO_PER_OBS;
    let min_class_count = (0..y_one_hot.ncols())
        .map(|class| y_one_hot.column(class).sum())
        .fold(f64::INFINITY, f64::min);
    if !min_class_count.is_finite() || min_class_count <= 0.0 {
        return base;
    }
    // Effective pseudo-observation prior strength: held to a fixed fraction of
    // the shrinking per-class data information once n_c falls below n_ref.
    let pseudo_obs_scale =
        (MULTINOMIAL_FORMULA_SPARSE_REFERENCE_SUPPORT / min_class_count).max(1.0);
    (base * pseudo_obs_scale).clamp(base, sparse)
}

fn max_abs_eta_location(eta: ArrayView2<'_, f64>) -> (f64, usize, usize) {
    let mut best = (0.0_f64, 0usize, 0usize);
    for ((row, active_class), &value) in eta.indexed_iter() {
        let abs = value.abs();
        if abs > best.0 {
            best = (abs, row, active_class);
        }
    }
    best
}

/// Separation gate for the REML/LAML **formula** path.
///
/// Unlike the bare fixed-λ driver [`fit_penalized_multinomial`] (which has no
/// outer REML state and so must reject a saturated, non-converged iterate as a
/// separation artifact at the [`MULTINOMIAL_SEPARATION_ETA_THRESHOLD`] logit
/// magnitude), the formula path can return a finite saturated mode after the
/// coupled outer optimizer has selected smoothing parameters. A `|η| >= 25`
/// gate is therefore wrong here: the penguins arm can legitimately have large
/// fitted logits while still producing finite probabilities and a usable REML
/// mode.
///
/// Only a genuinely NON-FINITE `η` (a NaN/Inf blow-up in the inner linear
/// algebra) is a real formula-path failure. A finite, even saturated, `η` is
/// accepted so the truth-recovery / match-or-beat bars are evaluated against the
/// actual fitted surface instead of an adapter diagnostic.
fn multinomial_formula_separation_diagnostic(
    inner_cycles: usize,
    outer_iterations: usize,
    block_states: &[ParameterBlockState],
) -> Option<EstimationError> {
    let mut nonfinite: Option<(f64, usize, usize)> = None;
    for (active_class, state) in block_states.iter().enumerate() {
        for (row, &value) in state.eta.iter().enumerate() {
            if !value.is_finite() {
                nonfinite = Some((value, row, active_class));
                break;
            }
        }
        if nonfinite.is_some() {
            break;
        }
    }
    nonfinite.map(|(value, row_index, active_class_index)| {
        EstimationError::MultinomialSeparationDetected {
            iteration: inner_cycles.max(outer_iterations),
            max_abs_eta: value.abs(),
            active_class_index,
            row_index,
        }
    })
}

/// Separation EVIDENCE gate for the conditional Firth/Jeffreys engagement on
/// the formula REML path (#715 / #753).
///
/// The structural mathematics (#715 issue thread): for any coefficient
/// direction `v` with `S v = 0` (a penalty-null direction — intercept, a
/// smooth's polynomial null component, an unpenalized parametric term), the
/// penalized joint Hessian satisfies `(H + S_λ) v = H v` for EVERY smoothing
/// parameter ρ. When the data (quasi-)separate, the softmax Fisher weight
/// `W = diag(p) − p pᵀ → 0` on the saturated rows, so `H v = JᵀWJ v → 0` along
/// the penalty-null directions those rows support: `(H + S_λ) v ≈ 0` for every
/// ρ — NO λ can repair it, the inner Newton can never certify a KKT point
/// there, and every outer REML startup seed is rejected (the penguins
/// real-data arm). The only principled cure is a PROPER prior on that
/// quotient-null subspace — the Jeffreys/Firth term `Φ = ½ log|ZᵀHZ|`, whose
/// Gauss–Newton curvature supplies the missing `O(1)` bound.
///
/// But the Firth prior is not free on interior data: unconditionally armed, it
/// shrinks fitted class probabilities toward the uniform simplex `1/K`
/// (an `O(1/n)` pull that the synthetic match-or-beat arm of #715 measured as
/// a real truth-RMSE loss vs the unbiased criterion). So the formula path
/// engages it ONLY on separation evidence, mirroring the #753 "diagnose, then
/// arm" split:
///
/// * a NON-FINITE logit — the inner linear algebra blew up along an unbounded
///   direction.
///
/// Returns `Some(description)` naming the witnessing logit when evidence is
/// found, `None` for a finite fit (which is then accepted as-is, with zero
/// Firth bias). A FAILED unbiased solve (`Err` from the rho-prior driver, e.g.
/// "no startup seed passed") is the second evidence form and is handled
/// directly at the call site in [`fit_penalized_multinomial_formula`].
fn multinomial_formula_separation_evidence(block_states: &[ParameterBlockState]) -> Option<String> {
    for (active_class, state) in block_states.iter().enumerate() {
        for (row, &value) in state.eta.iter().enumerate() {
            if !value.is_finite() {
                return Some(format!(
                    "non-finite logit eta[row {row}, active class {active_class}] = {value}"
                ));
            }
        }
    }
    None
}

/// Extra evidence used only for a NON-CONVERGED capped unbiased probe.
///
/// A converged finite saturated formula fit is still a valid optimum and must be
/// scored without Firth bias. A capped probe that failed to converge while it
/// already carries separation-scale logits is different: spending the full
/// unbiased outer budget on the same lambda-to-zero surface is the #1082
/// timeout. Route that case straight to the proper-prior refit.
fn multinomial_formula_unresolved_probe_separation_evidence(
    block_states: &[ParameterBlockState],
) -> Option<String> {
    if let Some(evidence) = multinomial_formula_separation_evidence(block_states) {
        return Some(evidence);
    }

    let mut best = (0.0_f64, 0usize, 0usize);
    for (active_class, state) in block_states.iter().enumerate() {
        for (row, &value) in state.eta.iter().enumerate() {
            let abs = value.abs();
            if abs > best.0 {
                best = (abs, row, active_class);
            }
        }
    }
    if best.0 >= MULTINOMIAL_SEPARATION_ETA_THRESHOLD {
        Some(format!(
            "separation-scale finite logit |eta[row {}, active class {}]| = {:.3e} \
             after capped unbiased probe",
            best.1, best.2, best.0
        ))
    } else {
        None
    }
}

/// Inputs to [`fit_penalized_multinomial`].
///
/// The penalty matrix `S` is shared across classes; per-class smoothing
/// parameters `lambdas` (length `K - 1`) scale `S` independently for each
/// active class. The full block-replicated penalty is `diag_a(λ_a) ⊗ S`,
/// which is exactly what [`gam_solve::arrow_schur::KroneckerPenaltyOp`]
/// expresses in matrix-free form when this driver is later lifted into the
/// arrow-Schur loop.
#[derive(Debug, Clone)]
pub struct MultinomialFitInputs<'a> {
    /// Design matrix `X ∈ ℝ^{N×P}` (one row per observation).
    pub design: ArrayView2<'a, f64>,
    /// Categorical response `Y ∈ ℝ^{N×K}`. Each row must be a point on the
    /// probability simplex (`y_c ≥ 0`, `Σ_c y_c = 1`): a one-hot indicator for
    /// hard classification, or a label-smoothed probability vector. Rows whose
    /// mass departs from 1 are rejected — the softmax residual gradient and
    /// Fisher block are the derivatives of `Σ_c y_c log p_c` only under the
    /// simplex constraint (see `validate_multinomial_simplex`).
    pub y_one_hot: ArrayView2<'a, f64>,
    /// Shared smoothing penalty `S ∈ ℝ^{P×P}` (symmetric, PSD).
    pub penalty: ArrayView2<'a, f64>,
    /// Per-active-class smoothing parameter `λ_a` (length `K - 1`).
    pub lambdas: ArrayView1<'a, f64>,
    /// Optional per-row weights (length `N`); `None` ⇒ uniform 1.0.
    pub row_weights: Option<ArrayView1<'a, f64>>,
    /// Optional per-row Fisher-block override, shape `(N, K-1, K-1)` in the
    /// active-class gauge (the reference class `K-1` is dropped). When `Some`,
    /// each Newton step uses this block as the curvature `W` in place of the
    /// analytic softmax Fisher `w_n (δ_ab p_a − p_a p_b)`; the gradient/residual
    /// path stays analytic, so this is a curvature-only override (the
    /// research escape-hatch for latent multinomial fits, issue #349). Each
    /// per-row block must be symmetric, PSD, and finite — preconditions the
    /// FFI boundary discharges before constructing this view.
    pub fisher_w_override: Option<ArrayView3<'a, f64>>,
    /// Maximum Newton iterations; recommend 50.
    pub max_iter: usize,
    /// Relative-step convergence tolerance; recommend 1e-7.
    pub tol: f64,
}

/// Outputs of [`fit_penalized_multinomial`].
#[derive(Debug, Clone)]
pub struct MultinomialFitOutputs {
    /// Active-class coefficient block, shape `(P, K-1)` (column `a` is `β_a`).
    /// The reference class `K - 1` has `β_{K-1} ≡ 0` by construction and is
    /// not stored.
    pub coefficients_active: Array2<f64>,
    /// Fitted probabilities, shape `(N, K)`.
    pub fitted_probabilities: Array2<f64>,
    /// Number of Newton iterations executed (including the final step that
    /// satisfied the tolerance).
    pub iterations: usize,
    /// `true` if the relative-step test was satisfied; `false` if the
    /// solver exhausted `max_iter`. (A non-converged solve is still
    /// returned; the caller decides whether to escalate.)
    pub converged: bool,
    /// Penalized negative log-likelihood at the returned `β̂`:
    /// `−log L(β̂) + ½ Σ_a λ_a · β̂_a^T S β̂_a`.
    pub penalized_neg_log_likelihood: f64,
    /// Unpenalized deviance `−2 log L(β̂)` for diagnostic reporting.
    pub deviance: f64,
    /// Joint Laplace posterior coefficient covariance `H⁻¹` at the converged
    /// `β̂`, shape `(P·(K−1))×(P·(K−1))` (#1101). Block-ordered to match the
    /// stacked active-class coefficient vector `β = [β_0; …; β_{K-2}]`: active
    /// class `a`'s `P` coefficients occupy rows/cols `a·P .. (a+1)·P`, indexed
    /// `θ[a·P + i] = β̂[i, a]`. This is the Laplace covariance from the factored
    /// penalized Hessian `XᵀWX + diag_a(λ_a)⊗S`; it drives the delta-method
    /// per-class probability standard errors ([`Self::predict_probabilities_with_se`])
    /// on the fixed-λ inner-solve path.
    pub coefficient_covariance: Array2<f64>,
}

impl MultinomialFitOutputs {
    /// Number of active classes `M = K − 1` (columns of
    /// [`Self::coefficients_active`]).
    pub fn n_active_classes(&self) -> usize {
        self.coefficients_active.ncols()
    }

    /// Per-class coefficient dimension `P` (rows of
    /// [`Self::coefficients_active`]).
    pub fn p_per_class(&self) -> usize {
        self.coefficients_active.nrows()
    }

    /// Evaluate `softmax(X·β̂)` AND its delta-method per-class probability
    /// standard error at fresh design rows `X_new` (#1101), using the joint
    /// Laplace covariance [`Self::coefficient_covariance`].
    ///
    /// The softmax Jacobian is `∂p_c/∂η_b = p_c (δ_{cb} − p_b)` for active class
    /// `b ∈ 0..M`, and `∂η_b/∂β[i,a] = X[i]·δ_{ab}`, so the gradient of the
    /// class-`c` probability w.r.t. the block-ordered coefficient vector is
    /// `g_c[a·P + i] = X[i]·p_c (δ_{ca} − p_a)` (the reference class `M`
    /// contributes only through `−p_a` in every active block). The delta-method
    /// variance is `Var(p_c) = g_cᵀ Σ g_c` with `Σ = H⁻¹`, and
    /// `SE(p_c) = √Var(p_c)`. Returns `(probs (N,K), prob_se (N,K))`. `X_new`
    /// must have `P` columns (the same design basis used at fit time); its row
    /// count sets `N`. The SE is unclamped (the interval consumer applies the
    /// simplex `[0,1]` clamp).
    pub fn predict_probabilities_with_se(
        &self,
        x_new: ArrayView2<'_, f64>,
    ) -> Result<(Array2<f64>, Array2<f64>), EstimationError> {
        let p = self.p_per_class();
        let m = self.n_active_classes();
        let k = m + 1;
        if x_new.ncols() != p {
            crate::bail_invalid_estim!(
                "predict_probabilities_with_se: X has {} cols, expected P={p}",
                x_new.ncols()
            );
        }
        let d = p * m;
        let cov = &self.coefficient_covariance;
        if cov.dim() != (d, d) {
            crate::bail_invalid_estim!(
                "predict_probabilities_with_se: covariance shape {:?} ≠ (P·M, P·M) = ({d}, {d})",
                cov.dim()
            );
        }
        let n_new = x_new.nrows();
        let beta = &self.coefficients_active;
        let mut probs = Array2::<f64>::zeros((n_new, k));
        let mut prob_se = Array2::<f64>::zeros((n_new, k));
        let mut eta_active = vec![0.0_f64; m];
        let mut row_probs = vec![0.0_f64; k];
        let mut grad = vec![0.0_f64; d];
        for row in 0..n_new {
            for a in 0..m {
                let mut v = 0.0_f64;
                for i in 0..p {
                    v += x_new[[row, i]] * beta[[i, a]];
                }
                eta_active[a] = v;
            }
            MultinomialLogitLikelihood::softmax_with_baseline(&eta_active, &mut row_probs);
            for c in 0..k {
                probs[[row, c]] = row_probs[c];
            }
            for c in 0..k {
                let pc = row_probs[c];
                // g_c[a·P + i] = X[i] · p_c · (δ_{ca} − p_a), a active.
                for a in 0..m {
                    let pa = row_probs[a];
                    let factor = pc * (if c == a { 1.0 - pa } else { -pa });
                    let base = a * p;
                    for i in 0..p {
                        grad[base + i] = x_new[[row, i]] * factor;
                    }
                }
                // Var = gᵀ Σ g.
                let mut var = 0.0_f64;
                for r in 0..d {
                    let gr = grad[r];
                    if gr == 0.0 {
                        continue;
                    }
                    let mut acc = 0.0_f64;
                    for s in 0..d {
                        acc += cov[[r, s]] * grad[s];
                    }
                    var += gr * acc;
                }
                prob_se[[row, c]] = var.max(0.0).sqrt();
            }
        }
        Ok((probs, prob_se))
    }
}

/// Fit a penalized multinomial-logit GAM at fixed `λ`.
///
/// See the module docs for the optimization problem and conventions. This
/// function is the canonical inner solve: the outer REML/LAML loop, when
/// added, calls this at each `ρ = log λ` trial.
pub fn fit_penalized_multinomial(
    inputs: MultinomialFitInputs<'_>,
) -> Result<MultinomialFitOutputs, EstimationError> {
    let MultinomialFitInputs {
        design,
        y_one_hot,
        penalty,
        lambdas,
        row_weights,
        fisher_w_override,
        max_iter,
        tol,
    } = inputs;

    // ──────────────────────── family-specific validation ───────────────────
    // The shared engine re-validates the geometry common to every vector-GLM
    // (nonempty design, penalty shape, λ finiteness/non-negativity, override
    // `(N, M, M)` shape, finite design). The multinomial family owns the
    // class-count contract (`K ≥ 2`, λ length `K − 1`), the per-row simplex
    // precondition under which the softmax residual/Fisher are the exact
    // derivatives of `Σ_c y_c log p_c`, and the row-weight check the likelihood
    // adapter consumes.
    let n_obs = design.nrows();
    let (y_rows, k) = y_one_hot.dim();
    if y_rows != n_obs {
        crate::bail_invalid_estim!(
            "fit_penalized_multinomial: y rows {y_rows} ≠ design rows {n_obs}"
        );
    }
    if k < 2 {
        crate::bail_invalid_estim!(
            "fit_penalized_multinomial: need at least 2 classes (got K={k})"
        );
    }
    let m = k - 1;
    if lambdas.len() != m {
        crate::bail_invalid_estim!(
            "fit_penalized_multinomial: lambdas length {} ≠ K-1 = {m}",
            lambdas.len()
        );
    }
    if let Some(fw) = fisher_w_override.as_ref() {
        if fw.dim() != (n_obs, m, m) {
            crate::bail_invalid_estim!(
                "fit_penalized_multinomial: fisher_w_override shape {:?} ≠ (N, K-1, K-1) = ({n_obs}, {m}, {m})",
                fw.dim()
            );
        }
    }
    if let Some(w) = row_weights.as_ref() {
        if w.len() != n_obs {
            crate::bail_invalid_estim!(
                "fit_penalized_multinomial: row_weights length {} ≠ N = {n_obs}",
                w.len()
            );
        }
        for (i, &v) in w.iter().enumerate() {
            if !(v.is_finite() && v >= 0.0) {
                crate::bail_invalid_estim!(
                    "fit_penalized_multinomial: row_weights[{i}] must be finite and ≥ 0 (got {v})"
                );
            }
        }
    }
    validate_multinomial_simplex(y_one_hot, "fit_penalized_multinomial")?;

    // ────────────────────────── likelihood construction ───────────────────
    let mut likelihood = MultinomialLogitLikelihood::with_classes(k)?;
    if let Some(w) = row_weights.as_ref() {
        likelihood = likelihood.with_row_weights(w.to_owned())?;
    }

    // ─────────────────── shared penalized vector-GLM solve ─────────────────
    // The softmax Fisher block is dense across the `M = K − 1` active classes;
    // the engine assembles the coupled `(P·M)×(P·M)` penalized Hessian, runs
    // the damped Newton loop, and returns the converged `β̂` and `η = X β̂`.
    let fit = fit_penalized_vector_glm(
        PenalizedVectorGlmInputs {
            design,
            y: y_one_hot,
            penalty,
            lambdas,
            fisher_w_override,
            max_iter,
            tol,
            // #1587: production multinomial still uses the per-class Diagonal
            // metric pending the REML per-class→per-term λ re-key that the
            // reference-symmetric Centered metric requires (shared λ). The
            // Centered engine path + its invariance proof land first.
            class_penalty_metric: crate::penalized_vector_glm::ClassPenaltyMetric::Diagonal,
        },
        &likelihood,
        "fit_penalized_multinomial",
    )?;

    let (max_abs_eta, row_index, active_class_index) = max_abs_eta_location(fit.eta.view());
    if !fit.converged && max_abs_eta >= MULTINOMIAL_SEPARATION_ETA_THRESHOLD {
        // Perfect / quasi-perfect separation (#1854): the UNBIASED softmax MLE is
        // not finite along `active_class_index`'s saturated logit direction, so
        // the fixed-λ Newton above ran away (`|η| ≥ 25`, no convergence). A
        // penalty-null direction `v` (`S v = 0`, e.g. an unpenalized intercept /
        // linear-covariate column) under softmax saturation has
        // `(XᵀWX + λS) v → 0` for EVERY λ, so no smoothing parameter can bound it
        // — only a proper prior on that quotient-null subspace can. Rather than
        // hard-erroring, engage the Firth/Jeffreys proper prior automatically
        // (magic-by-default): the full-span `½ log|I(β)|` correction supplies the
        // `O(1)` curvature that keeps the estimate finite on exactly those
        // separated directions while leaving well-identified fits untouched. This
        // reuses the same coupled joint-Newton Jeffreys machinery the formula
        // REML path arms on separation evidence (see
        // `fit_penalized_multinomial_formula`), only here at the caller's fixed λ.
        // Engage the fallback, but never let an internal consistency panic in
        // the coupled joint-Newton assembly (e.g. the #1395 logdet-collapse
        // guard) escape as a process abort: convert any panic into the
        // documented hard separation diagnostic, exactly as if the refit had
        // returned Err. This mirrors the catch_unwind panic-to-typed-error
        // boundary already used around the faer / cudarc entry points, and keeps
        // the separation path no worse than the pre-#1854 clean error while the
        // Firth refit is still being hardened.
        let firth = std::panic::catch_unwind(std::panic::AssertUnwindSafe(|| {
            fit_penalized_multinomial_firth_fallback(
                design,
                y_one_hot,
                penalty,
                lambdas,
                row_weights,
                max_iter,
                tol,
            )
        }));
        match firth {
            Ok(Ok(out)) => return Ok(out),
            // Firth refit errored, or an internal consistency guard panicked:
            // fall back to the explicit hard separation diagnostic.
            Ok(Err(_)) | Err(_) => {
                return Err(EstimationError::MultinomialSeparationDetected {
                    iteration: fit.iterations,
                    max_abs_eta,
                    active_class_index,
                    row_index,
                });
            }
        }
    }

    let fitted_probabilities = likelihood.probabilities(fit.eta.view());

    Ok(MultinomialFitOutputs {
        coefficients_active: fit.coefficients,
        fitted_probabilities,
        iterations: fit.iterations,
        converged: fit.converged,
        penalized_neg_log_likelihood: -fit.log_likelihood + fit.penalty_term,
        deviance: -2.0 * fit.log_likelihood,
        coefficient_covariance: fit.coefficient_covariance,
    })
}

/// Firth/Jeffreys-penalized multinomial refit engaged automatically when the
/// unbiased softmax MLE separates (#1854).
///
/// The unbiased fixed-λ solve ([`fit_penalized_multinomial`]) runs away on
/// (quasi-)separated data because the softmax likelihood has no finite mode along
/// the saturated logit direction and the smoothing penalty `S` cannot bound a
/// penalty-null direction (`S v = 0` ⇒ `(XᵀWX + λS) v → 0` for every λ). This
/// refit arms the full-span Jeffreys/Firth proper prior `½ log|I(β)|` on the
/// coupled joint softmax information, which supplies the `O(1)` curvature that
/// bounds exactly those directions and keeps the estimate finite.
///
/// # The estimator
///
/// It maximizes the penalized Firth objective at the caller's *fixed* `λ`
///
/// ```text
///   ℓ*(β) = Σ_n w_n Σ_c y_{nc} log p_{nc}
///           − ½ Σ_a λ_a βₐᵀ S βₐ
///           + ½ log det I(β)
/// ```
///
/// where `I(β)` is the coupled `(P·M)×(P·M)` softmax Fisher information (block
/// `(a,b)` is `Σ_n w_n (δ_{ab} p_{na} − p_{na} p_{nb}) x_n x_nᵀ`, block-ordered so
/// `θ[a·P+i] = β[i,a]`) and `M = K−1` active classes carry the reference-coded
/// logits (`η_{ref} ≡ 0`). The Jeffreys term `½ log det I(β)` is the standard
/// Firth penalty: it diverges to `−∞` as any fitted probability approaches the
/// simplex boundary (`I → 0`), so its maximizer is interior and finite on exactly
/// the separated directions that defeat every smoothing `λ`.
///
/// # Why this fixed-λ solver rather than the outer-REML formula path
///
/// The direct entry ([`fit_penalized_multinomial`]) is a fixed-λ inner solve — it
/// carries no outer smoothing selection — so the natural Firth engagement is a
/// fixed-λ Firth Newton, not the formula path's outer-REML joint-Newton machinery
/// (which is armed instead by [`fit_penalized_multinomial_formula`] on separation
/// evidence). Solving the Firth objective directly here keeps the separation
/// contract self-contained and independent of the shared trust-region/KKT
/// certificate machinery.
///
/// # The iteration
///
/// A Fisher-scoring Newton on `ℓ*`: the ascent direction is
/// `Δ = (I + Λ⊗S)⁻¹ U*`, where `U*` is the Firth-adjusted penalized score
///
/// ```text
///   U*[(c,s)] = Σ_n w_n x_{ns} (y_{nc} − p_{nc})       (data score)
///             − λ_c (S β_c)_s                           (smoothing penalty)
///             + ½ Σ_n w_n x_{ns} h^c_n                  (Firth adjustment)
/// ```
///
/// and the Firth adjustment uses `h^c_n = Σ_{a,b} G^c_{n,ab} Q_{n,ab}` with the
/// per-row information "hat" `Q_{n,ab} = x_nᵀ [I⁻¹]_{(a,b)} x_n` and the softmax
/// third-derivative tensor
/// `G^c_{ab} = δ_{ab} p_a (δ_{ac} − p_c) − p_a p_b (δ_{ac} + δ_{bc} − 2 p_c)`.
/// This `½ Σ tr(I⁻¹ ∂I/∂β)` is exactly `∇[½ log det I]` (finite-difference
/// verified). Each step is globalized by backtracking on `ℓ*`, so a step that
/// would push a probability to the boundary (making `I` non-PD) is rejected and
/// the fit stays interior. Convergence is the Newton decrement `½ U*ᵀΔ`.
fn fit_penalized_multinomial_firth_fallback(
    design: ArrayView2<'_, f64>,
    y_one_hot: ArrayView2<'_, f64>,
    penalty: ArrayView2<'_, f64>,
    lambdas: ArrayView1<'_, f64>,
    row_weights: Option<ArrayView1<'_, f64>>,
    max_iter: usize,
    tol: f64,
) -> Result<MultinomialFitOutputs, EstimationError> {
    use faer::Side;
    use gam_linalg::faer_ndarray::{
        FaerArrayView, array1_to_col_matmut, array2_to_matmut, factorize_symmetricwith_fallback,
    };
    use gam_linalg::matrix::FactorizedSystem;

    let n_obs = design.nrows();
    let p = design.ncols();
    let k = y_one_hot.ncols();
    let m = k - 1;
    let d = p * m;

    // Local softmax likelihood mirroring the caller's row weights, used to map the
    // fitted η back to probabilities.
    let mut likelihood = MultinomialLogitLikelihood::with_classes(k)?;
    if let Some(w) = row_weights.as_ref() {
        likelihood = likelihood.with_row_weights(w.to_owned())?;
    }
    let weight = |row: usize| -> f64 { row_weights.as_ref().map_or(1.0, |w| w[row]) };

    let max_iter = max_iter.max(1);
    let tol_eff = if tol.is_finite() && tol > 0.0 { tol } else { 1e-8 };

    // Probabilities (N, K), active classes 0..M then the pinned reference at M.
    let probs_at = |beta: &Array2<f64>| -> Array2<f64> {
        let eta = design.dot(beta);
        likelihood.probabilities(eta.view())
    };

    // Coupled softmax Fisher information I (d×d), block-ordered θ[a·P+i] = β[i,a].
    let assemble_info = |probs: &Array2<f64>| -> Array2<f64> {
        let mut info = Array2::<f64>::zeros((d, d));
        for row in 0..n_obs {
            let w = weight(row);
            if w == 0.0 {
                continue;
            }
            for a in 0..m {
                let pa = probs[[row, a]];
                let ao = a * p;
                for b in 0..m {
                    let pb = probs[[row, b]];
                    let wab = w * (if a == b { pa - pa * pb } else { -pa * pb });
                    if wab == 0.0 {
                        continue;
                    }
                    let bo = b * p;
                    for i in 0..p {
                        let xi = design[[row, i]];
                        if xi == 0.0 {
                            continue;
                        }
                        let cc = wab * xi;
                        for j in 0..p {
                            info[[ao + i, bo + j]] += cc * design[[row, j]];
                        }
                    }
                }
            }
        }
        info
    };

    // Factor a symmetric matrix (with escalating ridge only if it is not SPD) and
    // return its inverse and log-determinant.
    //
    // The ridge ladder is a standard relative-jitter Cholesky recovery, not a
    // tuned knob: (a) the base jitter is scaled to the matrix by `max_diag`
    // (`max_diag · ε` with ε at the double-precision Cholesky floor ~1e-10) so it
    // is invariant to the overall scale of the Fisher information, falling back
    // to an absolute floor only when the diagonal is degenerate; (b) it is tried
    // first at ridge 0 so an already-SPD matrix is factored unperturbed; (c) it
    // grows geometrically (×4) to span the ~120 dB from the base jitter to O(1)
    // in a bounded number of steps; (d) the attempt count is capped so a
    // genuinely singular information (e.g. an exactly rank-deficient Fisher block)
    // surfaces as an explicit error rather than an unbounded loop.
    let invert_spd = |mat: &Array2<f64>, context: &str| -> Result<(Array2<f64>, f64), EstimationError> {
        let max_diag = (0..d).fold(0.0_f64, |acc, i| acc.max(mat[[i, i]].abs()));
        let base = if max_diag.is_finite() && max_diag > 0.0 {
            max_diag * 1e-10
        } else {
            1e-10
        };
        let mut ridge = 0.0_f64;
        for _ in 0..=60 {
            let mut ridged = mat.clone();
            if ridge > 0.0 {
                for i in 0..d {
                    ridged[[i, i]] += ridge;
                }
            }
            if let Ok(factor) =
                factorize_symmetricwith_fallback(FaerArrayView::new(&ridged).as_ref(), Side::Lower)
            {
                let logdet = factor.logdet();
                if logdet.is_finite() {
                    let mut rhs = Array2::<f64>::eye(d);
                    {
                        let v = array2_to_matmut(&mut rhs);
                        factor.solve_in_place(v);
                    }
                    if rhs.iter().all(|x| x.is_finite()) {
                        let mut inv = Array2::<f64>::zeros((d, d));
                        for i in 0..d {
                            for j in 0..d {
                                inv[[i, j]] = 0.5 * (rhs[[i, j]] + rhs[[j, i]]);
                            }
                        }
                        return Ok((inv, logdet));
                    }
                }
            }
            ridge = if ridge > 0.0 { ridge * 4.0 } else { base };
        }
        Err(EstimationError::InvalidInput(format!(
            "multinomial Firth fallback: {context} not invertible (max_diag={max_diag:.3e})"
        )))
    };

    // SPD log-determinant only (no ridge): used by the backtracking line search to
    // reject any candidate that pushes a fitted probability to the simplex
    // boundary (where I loses positive-definiteness and the Firth term → −∞).
    let spd_logdet = |mat: &Array2<f64>| -> Option<f64> {
        factorize_symmetricwith_fallback(FaerArrayView::new(mat).as_ref(), Side::Lower)
            .ok()
            .map(|factor| factor.logdet())
            .filter(|ld| ld.is_finite())
    };

    // Penalized Firth objective ℓ* (MAXIMIZED), given probabilities, β, and the
    // precomputed log det I(β).
    let objective = |probs: &Array2<f64>, beta: &Array2<f64>, logdet_info: f64| -> f64 {
        let mut ll = 0.0_f64;
        for row in 0..n_obs {
            let w = weight(row);
            if w == 0.0 {
                continue;
            }
            for c in 0..k {
                let ycn = y_one_hot[[row, c]];
                if ycn != 0.0 {
                    ll += w * ycn * probs[[row, c]].max(f64::MIN_POSITIVE).ln();
                }
            }
        }
        let mut pen = 0.0_f64;
        for a in 0..m {
            let la = lambdas[a];
            if la != 0.0 {
                let bcol = beta.column(a);
                let sbeta = penalty.dot(&bcol);
                pen += 0.5 * la * bcol.dot(&sbeta);
            }
        }
        ll - pen + 0.5 * logdet_info
    };

    // Firth-adjusted penalized score U* (length d, block-ordered).
    let firth_score = |probs: &Array2<f64>, beta: &Array2<f64>, iinv: &Array2<f64>| -> Array1<f64> {
        let mut u = Array1::<f64>::zeros(d);
        let mut xn = vec![0.0_f64; p];
        let mut pa = vec![0.0_f64; m];
        let mut q = vec![0.0_f64; m * m];
        for row in 0..n_obs {
            let w = weight(row);
            if w == 0.0 {
                continue;
            }
            for i in 0..p {
                xn[i] = design[[row, i]];
            }
            for a in 0..m {
                pa[a] = probs[[row, a]];
            }
            // Data score: U[(a,i)] += w x_{ni} (y_{na} − p_{na}).
            for a in 0..m {
                let resid = y_one_hot[[row, a]] - pa[a];
                let ao = a * p;
                for i in 0..p {
                    u[ao + i] += w * xn[i] * resid;
                }
            }
            // Per-row information hat Q_{ab} = x_nᵀ [I⁻¹]_{(a,b)} x_n.
            for a in 0..m {
                let ao = a * p;
                for b in 0..m {
                    let bo = b * p;
                    let mut s = 0.0_f64;
                    for i in 0..p {
                        let xi = xn[i];
                        if xi == 0.0 {
                            continue;
                        }
                        let mut inner = 0.0_f64;
                        for j in 0..p {
                            inner += iinv[[ao + i, bo + j]] * xn[j];
                        }
                        s += xi * inner;
                    }
                    q[a * m + b] = s;
                }
            }
            // Firth adjustment: U[(c,s)] += ½ w x_{ns} h^c_n.
            for c in 0..m {
                let pc = pa[c];
                let mut h = 0.0_f64;
                for a in 0..m {
                    for b in 0..m {
                        let dab = if a == b { 1.0 } else { 0.0 };
                        let dac = if a == c { 1.0 } else { 0.0 };
                        let dbc = if b == c { 1.0 } else { 0.0 };
                        let g = dab * pa[a] * (dac - pc)
                            - pa[a] * pa[b] * (dac + dbc - 2.0 * pc);
                        h += g * q[a * m + b];
                    }
                }
                let co = c * p;
                for s in 0..p {
                    u[co + s] += 0.5 * w * h * xn[s];
                }
            }
        }
        // Smoothing penalty gradient: U[(a,i)] −= λ_a (S β_a)_i.
        for a in 0..m {
            let la = lambdas[a];
            if la != 0.0 {
                let sbeta = penalty.dot(&beta.column(a));
                let ao = a * p;
                for i in 0..p {
                    u[ao + i] -= la * sbeta[i];
                }
            }
        }
        u
    };

    // Penalized Hessian H = I + blockdiag_a(λ_a S) (positive definite).
    let penalized_hessian = |info: &Array2<f64>| -> Array2<f64> {
        let mut h = info.clone();
        for a in 0..m {
            let la = lambdas[a];
            if la != 0.0 {
                let ao = a * p;
                for i in 0..p {
                    for j in 0..p {
                        h[[ao + i, ao + j]] += la * penalty[[i, j]];
                    }
                }
            }
        }
        h
    };

    // Solve H Δ = U* for the SPD penalized Hessian, ridge-escalating only on
    // factorization failure. Same relative-jitter Cholesky-recovery ladder as
    // `invert_spd` above (see its comment for the rationale); the base jitter is
    // one decade tighter (`max_diag · 1e-12`) because the penalized Hessian
    // solved here is better conditioned than the Fisher information inverted
    // there, so a smaller perturbation suffices before escalating.
    let solve_spd = |mat: &Array2<f64>, rhs: &Array1<f64>| -> Result<Array1<f64>, EstimationError> {
        let max_diag = (0..d).fold(0.0_f64, |acc, i| acc.max(mat[[i, i]].abs()));
        let base = if max_diag.is_finite() && max_diag > 0.0 {
            max_diag * 1e-12
        } else {
            1e-12
        };
        let mut ridge = 0.0_f64;
        for _ in 0..=60 {
            let mut ridged = mat.clone();
            if ridge > 0.0 {
                for i in 0..d {
                    ridged[[i, i]] += ridge;
                }
            }
            if let Ok(factor) =
                factorize_symmetricwith_fallback(FaerArrayView::new(&ridged).as_ref(), Side::Lower)
            {
                let mut sol = rhs.clone();
                {
                    let v = array1_to_col_matmut(&mut sol);
                    factor.solve_in_place(v);
                }
                if sol.iter().all(|x| x.is_finite()) {
                    return Ok(sol);
                }
            }
            ridge = if ridge > 0.0 { ridge * 4.0 } else { base };
        }
        Err(EstimationError::InvalidInput(
            "multinomial Firth fallback: penalized Hessian solve failed".to_string(),
        ))
    };

    // ─────────────────────────── Firth Newton loop ────────────────────────────
    let mut beta = Array2::<f64>::zeros((p, m));
    let mut converged = false;
    let mut iterations = 0_usize;
    for it in 0..max_iter {
        iterations = it + 1;
        let probs = probs_at(&beta);
        let info = assemble_info(&probs);
        let (iinv, logdet_info) = invert_spd(&info, "Fisher information")?;
        let u = firth_score(&probs, &beta, &iinv);
        let hmat = penalized_hessian(&info);
        let step_vec = solve_spd(&hmat, &u)?;

        // Newton decrement ½ U*ᵀ H⁻¹ U* = ½ U*ᵀ Δ (≥ 0, scale-aware stop).
        let decrement = u.dot(&step_vec);
        if 0.5 * decrement.abs() < tol_eff {
            converged = true;
            break;
        }

        // Δ as (P, M): delta[i, a] = step_vec[a·P + i].
        let mut delta = Array2::<f64>::zeros((p, m));
        for a in 0..m {
            let ao = a * p;
            for i in 0..p {
                delta[[i, a]] = step_vec[ao + i];
            }
        }

        // Backtracking line search on ℓ* (ascent). Reject any candidate whose I is
        // not SPD (boundary), so the iterate stays interior.
        let o0 = objective(&probs, &beta, logdet_info);
        let mut step = 1.0_f64;
        let mut accepted = false;
        for _ in 0..60 {
            let cand = &beta + &(&delta * step);
            let cand_probs = probs_at(&cand);
            let cand_info = assemble_info(&cand_probs);
            if let Some(cand_logdet) = spd_logdet(&cand_info) {
                let o1 = objective(&cand_probs, &cand, cand_logdet);
                if o1 >= o0 - 1e-12 {
                    beta = cand;
                    accepted = true;
                    break;
                }
            }
            step *= 0.5;
        }
        if !accepted {
            // Backtracking exhausted 60 halvings without an admissible ascent
            // step. This is convergence ONLY if the iterate is already first-order
            // stationary; a line-search stall at a non-stationary point is a
            // solver failure and must be reported as such, never papered over as
            // `converged = true` (#2066 — SPEC: do not report a non-converged
            // iterate as success).
            //
            // The verdict is the loop's OWN stationarity test — the Newton
            // decrement `½·Uᵀ H⁻¹ U` against `tol_eff`, the same criterion the top
            // of the loop uses to break as converged. A true interior mode never
            // reaches this branch: an infinitesimal step (`step → 0`) leaves the
            // iterate SPD with `o1 ≈ o0`, so it is accepted; a numerically flat
            // mode is caught by the `max_step` test below after that accepted
            // tiny step. Reaching here therefore means Newton still sees a
            // meaningful ascent direction it cannot realize (boundary / near-
            // singular Fisher information), i.e. a genuine stall → not converged.
            converged = 0.5 * decrement.abs() < tol_eff;
            break;
        }

        let max_step = step * delta.iter().fold(0.0_f64, |acc, &v| acc.max(v.abs()));
        let scale = 1.0 + beta.iter().fold(0.0_f64, |acc, &v| acc.max(v.abs()));
        if max_step < tol_eff * scale {
            converged = true;
            break;
        }
    }

    // ─────────────────────────── final quantities ─────────────────────────────
    for (idx, &v) in beta.iter().enumerate() {
        if !v.is_finite() {
            crate::bail_invalid_estim!(
                "multinomial Firth fallback: non-finite coefficient at flat index {idx} = {v}"
            );
        }
    }
    let coefficients_active = beta;

    let probs = probs_at(&coefficients_active);
    let info = assemble_info(&probs);
    // Laplace covariance H⁻¹ at the converged mode (block-ordered θ[a·P+i]).
    let hmat = penalized_hessian(&info);
    let coefficient_covariance = match invert_spd(&hmat, "penalized Hessian covariance") {
        Ok((cov, _)) => cov,
        Err(_) => Array2::<f64>::zeros((d, d)),
    };

    let fitted_probabilities = probs;
    let mut log_likelihood = 0.0_f64;
    for row in 0..n_obs {
        let w = weight(row);
        for c in 0..k {
            let ycn = y_one_hot[[row, c]];
            if ycn != 0.0 {
                log_likelihood +=
                    w * ycn * fitted_probabilities[[row, c]].max(f64::MIN_POSITIVE).ln();
            }
        }
    }

    let mut penalty_term = 0.0_f64;
    for a in 0..m {
        let beta_col = coefficients_active.column(a);
        let sbeta = penalty.dot(&beta_col);
        penalty_term += 0.5 * lambdas[a] * beta_col.dot(&sbeta);
    }

    Ok(MultinomialFitOutputs {
        coefficients_active,
        fitted_probabilities,
        iterations,
        converged,
        penalized_neg_log_likelihood: -log_likelihood + penalty_term,
        deviance: -2.0 * log_likelihood,
        coefficient_covariance,
    })
}

// ---------------------------------------------------------------------------
// Formula-driven multinomial pipeline
// ---------------------------------------------------------------------------
//
// Slice A of the multinomial integration: a single public entry that takes
// a parsed `EncodedDataset`, a Wilkinson-style formula, and a uniform initial
// smoothing parameter, then runs the full
//
//     parse → termspec → design (X, S blocks) → one-hot Y → REML λ-selection
//
// pipeline. `fit_penalized_multinomial_formula` drives the outer REML/LAML
// loop (via the custom-family path) to select an independent λ per (class,
// term); `init_lambda` (default 1.0) is only the warm-start seed for every
// block. The reference class is the last level of the categorical response
// column as recorded in the dataset schema.

/// Saved-model payload for a multinomial fit driven by a Wilkinson formula.
///
/// This is what the FFI returns to Python. It carries everything the Python
/// `MultinomialModel.predict` path needs to evaluate `softmax(X_new · β)` on
/// fresh data using the *training* basis / penalty structure (no refit on
/// predict, no re-derivation of class levels).
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct MultinomialSavedModel {
    /// The training formula, verbatim. Stored so Python's `summary()` and
    /// any round-trip persistence path can echo what was fit.
    pub formula: String,
    /// Names of the *training* response levels in canonical order. The last
    /// entry is the reference class (η = 0); the first `K - 1` carry the
    /// active linear-predictor blocks. Class permutations are forbidden:
    /// this list is fixed at fit time and predictions emit columns in the
    /// same order.
    pub class_levels: Vec<String>,
    /// Index of the reference class within `class_levels` — currently always
    /// `class_levels.len() - 1`, exposed as a field so future "user-pinned
    /// reference" gauges (e.g. `family='multinomial', reference='setosa'`)
    /// can land without changing the on-disk shape.
    pub reference_class_index: usize,
    /// Resolved term-collection spec used to build `X` at fit time. Replayed
    /// on predict via [`gam_terms::smooth::build_term_collection_design`].
    pub resolved_termspec: TermCollectionSpec,
    /// Active-class coefficient block, shape `(P, K-1)`. Column `a` is the
    /// coefficient vector for class `class_levels[a]`. Stored flat in
    /// row-major order to keep the serde payload self-describing.
    pub coefficients_flat: Vec<f64>,
    /// `P` — coefficient count per active class. Matches the column count of
    /// the design matrix the saved `resolved_termspec` produces.
    pub p_per_class: usize,
    /// Number of active classes (`K - 1`).
    pub n_active_classes: usize,
    /// Original training column headers, in dataset-column order. Needed at
    /// predict time so the FFI can align a fresh `Dataset` to the training
    /// schema before evaluating the basis.
    pub training_headers: Vec<String>,
    /// REML/LAML-selected smoothing parameters, one per `(active class, smooth
    /// term)`, flattened in block-major order: all of class 0's per-term λ,
    /// then class 1's, and so on. Per-term penalties (#561) mean each active
    /// class block selects an *independent* λ for every smooth term, so this
    /// vector has length `Σ_a (#terms in class a)` = `(K − 1) · #terms`. Use
    /// [`MultinomialSavedModel::lambdas_per_block`] to segment it by class. An
    /// unpenalized model (no smooth terms) yields an empty vector.
    pub lambdas: Vec<f64>,
    /// Number of smoothing parameters (smooth terms) in each active class
    /// block, parallel to `class_levels[0..K-1]`. Segments the flat `lambdas`
    /// vector: class `a`'s λ are `lambdas[Σ_{b<a} lambdas_per_block[b] ..][..
    /// lambdas_per_block[a]]`. Every entry is identical in the shared-design
    /// architecture (all classes share the same term structure), but it is
    /// stored explicitly so consumers never have to assume that.
    pub lambdas_per_block: Vec<usize>,
    /// Newton iterations executed; recorded for the summary report.
    pub iterations: usize,
    /// `true` if the inner Newton solver hit the relative-step tolerance.
    pub converged: bool,
    /// Penalized negative log-likelihood at the returned `β̂`.
    pub penalized_neg_log_likelihood: f64,
    /// Unpenalized deviance `−2 log L(β̂)`.
    pub deviance: f64,
    /// Per-active-class effective degrees of freedom (hat-matrix trace),
    /// length `K - 1`. Populated when the REML driver reports an
    /// inference block; falls back to `None` for the legacy fixed-λ path.
    #[serde(default)]
    pub edf_per_class: Option<Vec<f64>>,
    /// Per-PENALTY effective degrees of freedom, one entry per smoothing
    /// parameter (length `== lambdas.len()`), aligned block-major with the flat
    /// [`Self::lambdas`] / [`Self::lambdas_per_block`] layout. Each entry is the
    /// penalty-block trace EDF `rank(S_k) − λ_k·tr(H⁻¹ S_k)`, clamped to
    /// `[0, rank(S_k)]`. This is the per-(class, term, penalty) resolution that
    /// the per-class [`Self::edf_per_class`] SUM deliberately hides: only the
    /// per-penalty vector reveals whether an individual smooth collapsed onto its
    /// polynomial null space (its wiggliness λ driven to the λ-cap), which a
    /// per-class total cannot show. Populated whenever the REML driver reports an
    /// inference block; `None` on the legacy fixed-λ path or when the trace
    /// channel is mis-shaped. Unlike `edf_per_class`, the entries do NOT sum to
    /// the model EDF when several penalties share one coefficient range (a
    /// double-penalty smooth has `Σ_k rank(S_k) > p_per_class`).
    #[serde(default)]
    pub edf_per_penalty: Option<Vec<f64>>,
    /// Joint posterior coefficient covariance `H⁻¹` (#1101), block-ordered to
    /// match the stacked active-class coefficient vector `β = [β_0; …; β_{K-2}]`
    /// (class `a`'s `P` coefficients occupy rows/cols `a·P .. (a+1)·P`). This is
    /// the Laplace covariance the REML driver already computes from the factored
    /// penalized Hessian; storing it gives the predict path delta-method
    /// per-class probability standard errors and the summary its Wald
    /// smooth-term tests. Flattened row-major over the `(P·M)×(P·M)` matrix.
    /// `None` for a model fitted before covariance was surfaced.
    #[serde(default)]
    pub coefficient_covariance_flat: Option<Vec<f64>>,
    /// Joint coefficient-space influence matrix `F = H⁻¹ X'WX` (#1101),
    /// block-ordered identically to [`Self::coefficient_covariance_flat`].
    /// Its per-term diagonal block trace is the term's effective degrees of
    /// freedom and its `tr(F_jj)²/tr(F_jj²)` the Wood reference d.f., feeding
    /// the rank-truncated Wald smooth-term test in `summary()`. Flattened
    /// row-major over the `(P·M)×(P·M)` matrix. `None` when unavailable.
    #[serde(default)]
    pub coefficient_influence_flat: Option<Vec<f64>>,
    /// Per-(active class, smooth term) coefficient column range and unpenalized
    /// nullspace dimension within the `P`-wide class block (#1101). Parallel to
    /// the smooth terms the design produced; replicated across classes by the
    /// shared-design architecture. Drives the Wald smooth-term table in
    /// `summary()`. Empty for a wholly parametric (no-smooth) model.
    #[serde(default)]
    pub smooth_term_spans: Vec<MultinomialSmoothTermSpan>,
    /// One descriptive label per *penalty component* within a single active-class
    /// block, parallel to that block's λ slice (i.e. length
    /// `lambdas_per_block[0]`). The Marra–Wood double penalty (and tensor /
    /// operator smooths) emit **more than one** penalty component — hence more
    /// than one λ — per smooth term, so this is NOT 1:1 with
    /// [`Self::smooth_term_spans`]: a single `s(x)` term contributes a primary
    /// wiggliness λ labelled `s(x)` and a null-space shrinkage λ labelled
    /// `s(x) [null space]`. The summary renderer pairs `lambdas` with these
    /// labels component-for-component so no λ is ever dropped (#1544). Built from
    /// the per-component term name + penalty role at fit time; empty for a
    /// wholly parametric model or a model serialized before this field existed.
    #[serde(default)]
    pub lambda_labels: Vec<String>,
}

/// One smooth term's coefficient span within a class block, plus its
/// unpenalized nullspace dimension and a display label (#1101). The Wald
/// smooth-significance test in `summary()` slices the joint covariance /
/// influence at `a·P + col_start .. a·P + col_end` for active class `a`.
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct MultinomialSmoothTermSpan {
    /// Human-readable term label (the smooth's formula token), for the table.
    pub label: String,
    /// Start column of the term within the per-class `P`-wide coefficient block.
    pub col_start: usize,
    /// End column (exclusive) of the term within the per-class block.
    pub col_end: usize,
    /// Leading unpenalized (polynomial nullspace) dimension within the term.
    pub nullspace_dim: usize,
}

/// Descriptive label for one penalty *component* (one λ) within a class block,
/// for the `summary()` per-class λ rollup (#1544). A smooth term can emit
/// several penalty components — the Marra–Wood double penalty splits `s(x)`
/// into a primary wiggliness penalty and a null-space shrinkage penalty, and
/// tensor / operator smooths emit a component per margin / differential
/// operator — each with its own independently-selected λ. The label is the
/// term name (from `PenaltyBlockInfo::termname`) plus a role suffix derived
/// from the penalty's [`PenaltySource`], so each λ in the summary names both
/// the term it smooths and the role it plays. `pen_idx` is the global penalty
/// index, used only as a last-resort fallback label.
fn penalty_component_label(info: Option<&PenaltyBlockInfo>, pen_idx: usize) -> String {
    use gam_terms::basis::PenaltySource;
    let term = info
        .and_then(|i| i.termname.clone())
        .unwrap_or_else(|| format!("s{pen_idx}"));
    let role = match info.map(|i| &i.penalty.source) {
        // The primary wiggliness penalty is the term's "main" λ; show the bare
        // term name so the common single-penalty case reads cleanly.
        Some(PenaltySource::Primary) | None => None,
        Some(PenaltySource::DoublePenaltyNullspace) => Some("null space".to_string()),
        Some(PenaltySource::OperatorMass) => Some("mass".to_string()),
        Some(PenaltySource::OperatorTension) => Some("tension".to_string()),
        Some(PenaltySource::OperatorStiffness) => Some("stiffness".to_string()),
        Some(PenaltySource::OperatorRelevance { axis }) => Some(format!("axis {axis}")),
        Some(PenaltySource::TensorMarginal { dim }) => Some(format!("margin {dim}")),
        Some(PenaltySource::TensorSeparable { penalized_margins }) => {
            Some(format!("separable {penalized_margins:?}"))
        }
        Some(PenaltySource::TensorGlobalRidge) => Some("ridge".to_string()),
        Some(PenaltySource::Other(s)) => Some(s.clone()),
    };
    match role {
        Some(role) => format!("{term} [{role}]"),
        None => term,
    }
}

impl MultinomialSavedModel {
    /// Active-class coefficient block as an `(P, K-1)` `ndarray` view.
    pub fn coefficients_active(&self) -> Array2<f64> {
        Array2::from_shape_vec(
            (self.p_per_class, self.n_active_classes),
            self.coefficients_flat.clone(),
        )
        .expect(
            "MultinomialSavedModel.coefficients_flat length must equal p_per_class * n_active_classes",
        )
    }

    /// Evaluate `softmax(X · β)` at fresh data rows. `X_new` must have
    /// `self.p_per_class` columns (i.e. it was built from the same
    /// `resolved_termspec` as fit time). Returns an `(N_new, K)` matrix
    /// with rows summing to 1; column order matches `self.class_levels`.
    pub fn predict_probabilities(&self, x_new: ArrayView2<'_, f64>) -> Array2<f64> {
        let n_new = x_new.nrows();
        let p = self.p_per_class;
        let m = self.n_active_classes;
        let k = m + 1;
        assert_eq!(
            x_new.ncols(),
            p,
            "MultinomialSavedModel.predict_probabilities: X has {} cols, expected {p}",
            x_new.ncols()
        );
        let beta = self.coefficients_active();
        let mut probs = Array2::<f64>::zeros((n_new, k));
        let mut eta_active = vec![0.0_f64; m];
        let mut row_probs = vec![0.0_f64; k];
        for row in 0..n_new {
            for a in 0..m {
                let mut v = 0.0_f64;
                for i in 0..p {
                    v += x_new[[row, i]] * beta[[i, a]];
                }
                eta_active[a] = v;
            }
            MultinomialLogitLikelihood::softmax_with_baseline(&eta_active, &mut row_probs);
            for c in 0..k {
                probs[[row, c]] = row_probs[c];
            }
        }
        probs
    }

    /// Reconstruct the joint posterior covariance `H⁻¹` as a `(P·M)×(P·M)`
    /// `ndarray`, block-ordered to match the stacked coefficient vector
    /// `θ[a·P + i] = β[i, a]` (#1101). `None` when the model was fitted before
    /// covariance was surfaced (legacy payload).
    pub fn coefficient_covariance(&self) -> Option<Array2<f64>> {
        let d = self.p_per_class.checked_mul(self.n_active_classes)?;
        let flat = self.coefficient_covariance_flat.as_ref()?;
        Array2::from_shape_vec((d, d), flat.clone()).ok()
    }

    /// Reconstruct the joint influence matrix `F = H⁻¹ X'WX` as a
    /// `(P·M)×(P·M)` `ndarray`, block-ordered like
    /// [`Self::coefficient_covariance`] (#1101). `None` when unavailable.
    pub fn coefficient_influence(&self) -> Option<Array2<f64>> {
        let d = self.p_per_class.checked_mul(self.n_active_classes)?;
        let flat = self.coefficient_influence_flat.as_ref()?;
        Array2::from_shape_vec((d, d), flat.clone()).ok()
    }

    /// Evaluate `softmax(X·β)` AND its delta-method per-class probability
    /// standard error at fresh data rows (#1101).
    ///
    /// For active classes `b ∈ 0..M` the softmax Jacobian is
    /// `∂p_c/∂η_b = p_c (δ_{cb} − p_b)`, and `∂η_b/∂β[i,a] = X[i]·δ_{ab}`, so the
    /// gradient of class-`c` probability w.r.t. the block-ordered coefficient
    /// vector is `g_c[a·P + i] = X[i]·p_c (δ_{ca} − p_a)` (active `a`; the
    /// reference class `M` contributes `p_c(0 − p_a)` via every active block).
    /// The delta-method variance is `Var(p_c) = g_cᵀ Σ g_c` with `Σ = H⁻¹` the
    /// joint posterior covariance, and `SE(p_c) = √Var(p_c)`. Returns
    /// `(probs (N,K), prob_se (N,K))`; `prob_se` is `None` when no covariance is
    /// stored. The simplex `[0,1]` clamp is applied by the interval consumer, not
    /// here (the SE itself is unclamped).
    pub fn predict_probabilities_with_se(
        &self,
        x_new: ArrayView2<'_, f64>,
    ) -> (Array2<f64>, Option<Array2<f64>>) {
        let probs = self.predict_probabilities(x_new);
        let Some(cov) = self.coefficient_covariance() else {
            return (probs, None);
        };
        let n_new = x_new.nrows();
        let p = self.p_per_class;
        let m = self.n_active_classes;
        let k = m + 1;
        let d = p * m;
        let mut prob_se = Array2::<f64>::zeros((n_new, k));
        let mut grad = vec![0.0_f64; d];
        for row in 0..n_new {
            let prow = probs.row(row);
            for c in 0..k {
                let pc = prow[c];
                // g_c[a·P + i] = X[i] · p_c · (δ_{ca} − p_a), a active.
                for a in 0..m {
                    let pa = prow[a];
                    let factor = pc * (if c == a { 1.0 - pa } else { -pa });
                    let base = a * p;
                    for i in 0..p {
                        grad[base + i] = x_new[[row, i]] * factor;
                    }
                }
                // Var = gᵀ Σ g.
                let mut var = 0.0_f64;
                for r in 0..d {
                    let gr = grad[r];
                    if gr == 0.0 {
                        continue;
                    }
                    let mut acc = 0.0_f64;
                    for s in 0..d {
                        acc += cov[[r, s]] * grad[s];
                    }
                    var += gr * acc;
                }
                prob_se[[row, c]] = var.max(0.0).sqrt();
            }
        }
        (probs, Some(prob_se))
    }

    /// Wood (2013) rank-truncated Wald smooth-significance test per
    /// `(active class, smooth term)` (#1101), reusing the exact scalar-summary
    /// kernel [`gam_terms::inference::smooth_test::wood_smooth_test`]. For active
    /// class `a` and term span `[c0, c1)` within the class block, the global
    /// coefficient range is `a·P + c0 .. a·P + c1`; the joint covariance and
    /// influence are sliced there. The term EDF is the influence-block trace
    /// `tr(F_jj)` (when present) and the reference d.f. uses `tr(F_jj)²/tr(F_jj²)`,
    /// exactly as the scalar path. The multinomial softmax is a known-dispersion
    /// family, so the χ²_{ref_df} branch applies. Returns one row per
    /// `(class label, term label, edf, ref_df, statistic, p_value)`; empty when
    /// no covariance/smooth terms are available.
    pub fn smooth_significance(&self) -> Vec<MultinomialSmoothSignificance> {
        let mut out = Vec::new();
        let p = self.p_per_class;
        let m = self.n_active_classes;
        let Some(cov) = self.coefficient_covariance() else {
            return out;
        };
        if self.smooth_term_spans.is_empty() {
            return out;
        }
        let beta = self.coefficients_active();
        // Block-ordered θ = [β_0; …; β_{M-1}], θ[a·P + i] = β[i, a].
        let d = p * m;
        let mut theta = Array1::<f64>::zeros(d);
        for a in 0..m {
            for i in 0..p {
                theta[a * p + i] = beta[[i, a]];
            }
        }
        let influence = self.coefficient_influence();
        for a in 0..m {
            let class_label = self
                .class_levels
                .get(a)
                .cloned()
                .unwrap_or_else(|| format!("class{a}"));
            let base = a * p;
            for span in &self.smooth_term_spans {
                if span.col_end > p {
                    continue;
                }
                let start = base + span.col_start;
                let end = base + span.col_end;
                // Term EDF = tr(F_jj); without an influence matrix fall back to
                // the block coefficient count (full-rank Wald on the span).
                let block_len = (span.col_end - span.col_start) as f64;
                let edf = influence
                    .as_ref()
                    .map(|f| (start..end).map(|i| f[[i, i]]).sum::<f64>())
                    .filter(|v| v.is_finite() && *v > 0.0)
                    .unwrap_or(block_len);
                let result = gam_terms::inference::smooth_test::wood_smooth_test(
                    gam_terms::inference::smooth_test::SmoothTestInput {
                        beta: theta.view(),
                        covariance: &cov,
                        influence_matrix: influence.as_ref(),
                        whitening_gram: None,
                        coeff_range: start..end,
                        edf,
                        nullspace_dim: span.nullspace_dim,
                        residual_df: f64::INFINITY,
                        scale: gam_terms::inference::smooth_test::SmoothTestScale::Known,
                    },
                );
                if let Some(res) = result {
                    out.push(MultinomialSmoothSignificance {
                        class_label: class_label.clone(),
                        term_label: span.label.clone(),
                        edf,
                        ref_df: res.ref_df,
                        statistic: res.statistic,
                        p_value: res.p_value,
                    });
                }
            }
        }
        out
    }

    /// Draw `n_draws` posterior-predictive replicate class assignments at fresh
    /// rows (#1101). Each draw independently samples every row's class from
    /// `Categorical(p_row)` with `p = softmax(X·β̂)` — the plug-in predictive
    /// distribution, i.e. the multinomial observation noise wrapped around the
    /// fitted mean (the categorical analogue of the scalar families'
    /// `sample_replicates`). The returned `(n_draws, N)` matrix holds class
    /// INDICES `0..K`, aligned to [`Self::class_levels`]. The draw stream is a
    /// `StdRng` seeded by `seed`, so `(x_new, n_draws, seed)` reproduce
    /// bit-identically — the engine for posterior-predictive checks and
    /// simulation-based calibration. `x_new` must have `self.p_per_class`
    /// columns (built from the same `resolved_termspec` as fit time).
    pub fn sample_replicate_classes(
        &self,
        x_new: ArrayView2<'_, f64>,
        n_draws: usize,
        seed: u64,
    ) -> Array2<u32> {
        use rand::{RngExt, SeedableRng};
        let probs = self.predict_probabilities(x_new);
        let n = probs.nrows();
        let k = probs.ncols();
        let mut out = Array2::<u32>::zeros((n_draws, n));
        let mut rng = rand::rngs::StdRng::seed_from_u64(seed);
        for d in 0..n_draws {
            for row in 0..n {
                let u: f64 = rng.random::<f64>();
                // Inverse-CDF categorical draw over the K simplex weights.
                let mut acc = 0.0_f64;
                let mut chosen = k - 1; // numerical fallback = reference class
                for c in 0..k {
                    acc += probs[[row, c]];
                    if u < acc {
                        chosen = c;
                        break;
                    }
                }
                out[[d, row]] = chosen as u32;
            }
        }
        out
    }
}

/// One row of the multinomial smooth-significance table (#1101): the Wood
/// rank-truncated Wald test for one `(active class, smooth term)` pair.
#[derive(Debug, Clone)]
pub struct MultinomialSmoothSignificance {
    pub class_label: String,
    pub term_label: String,
    pub edf: f64,
    pub ref_df: f64,
    pub statistic: f64,
    pub p_value: f64,
}

/// One-hot-encode the categorical response column and return both the
/// encoding and the captured level names. The level order matches the order
/// recorded in the dataset schema, which is the canonical (lexicographically
/// sorted) factor order produced by inferred-schema construction (#1319) — so
/// it is a deterministic function of the label *set*, independent of training
/// row order (no silent class permutation under a row shuffle), and matches the
/// R `factor()` / pandas `Categorical` convention.
fn one_hot_categorical_response(
    data: &EncodedDataset,
    y_col: usize,
    response_name: &str,
) -> Result<(Array2<f64>, Vec<String>), EstimationError> {
    let levels: Vec<String> = data
        .schema
        .columns
        .get(y_col)
        .map(|sc| sc.levels.clone())
        .unwrap_or_default();
    if levels.len() < 2 {
        crate::bail_invalid_estim!(
            "multinomial response '{response_name}' must have at least 2 categorical levels (got {})",
            levels.len()
        );
    }
    let n = data.values.nrows();
    let k = levels.len();
    let mut y_one_hot = Array2::<f64>::zeros((n, k));
    for row in 0..n {
        let encoded = data.values[[row, y_col]];
        if !encoded.is_finite() {
            crate::bail_invalid_estim!(
                "multinomial response '{response_name}' row {row} is non-finite ({encoded})"
            );
        }
        let class_idx = encoded.round() as i64;
        if class_idx < 0 || (class_idx as usize) >= k {
            crate::bail_invalid_estim!(
                "multinomial response '{response_name}' row {row} encoded as {encoded} \
                 is outside the level range 0..{k}"
            );
        }
        y_one_hot[[row, class_idx as usize]] = 1.0;
    }
    Ok((y_one_hot, levels))
}

/// Build `(TermCollectionSpec, TermCollectionDesign)` from a formula against
/// a categorical-response dataset. Mirrors the early scaffolding inside
/// `materialize_standard` (response role resolution, geometry-aware spec
/// build) without touching the scalar-family resolution path — multinomial
/// owns its own response kind check.
fn build_formula_design_for_multinomial(
    formula: &str,
    data: &EncodedDataset,
    config: &FitConfig,
) -> Result<
    (
        TermCollectionSpec,
        TermCollectionDesign,
        usize,
        String,
        ResponseColumnKind,
    ),
    EstimationError,
> {
    let parsed = parse_formula(formula).map_err(|err| {
        EstimationError::InvalidInput(format!(
            "multinomial fit: failed to parse formula {formula:?}: {err}"
        ))
    })?;
    let col_map = data.column_map();
    let y_col = resolve_role_col(&col_map, &parsed.response, "response")
        .map_err(|err| EstimationError::InvalidInput(format!("multinomial fit: {err}")))?;
    let y_kind = crate::fit_orchestration::response_column_kind(data, y_col);
    let policy = resolved_resource_policy(config, data, ProblemHints::default());
    let mut inference_notes: Vec<String> = Vec::new();
    let spec = build_termspec_with_geometry_and_overrides(
        &parsed.terms,
        data,
        &col_map,
        &mut inference_notes,
        config.scale_dimensions,
        &policy,
        config.smooth_overrides.as_ref(),
    )
    .map_err(|err| {
        EstimationError::InvalidInput(format!("multinomial fit: build termspec: {err}"))
    })?;
    let design = build_term_collection_design(data.values.view(), &spec).map_err(|err| {
        EstimationError::InvalidInput(format!("multinomial fit: build design: {err}"))
    })?;
    Ok((spec, design, y_col, parsed.response, y_kind))
}

fn scale_multinomial_formula_penalty(penalty: PenaltyMatrix, scale: f64) -> PenaltyMatrix {
    match penalty {
        PenaltyMatrix::Dense(matrix) => PenaltyMatrix::Dense(matrix.mapv(|v| v * scale)),
        PenaltyMatrix::KroneckerFactored { left, right } => PenaltyMatrix::KroneckerFactored {
            left: left.mapv(|v| v * scale),
            right,
        },
        PenaltyMatrix::Blockwise {
            local,
            col_range,
            total_dim,
        } => PenaltyMatrix::Blockwise {
            local: local.mapv(|v| v * scale),
            col_range,
            total_dim,
        },
        PenaltyMatrix::Labeled { label, inner } => PenaltyMatrix::Labeled {
            label,
            inner: Box::new(scale_multinomial_formula_penalty(*inner, scale)),
        },
        PenaltyMatrix::Fixed { log_lambda, inner } => PenaltyMatrix::Fixed {
            log_lambda,
            inner: Box::new(scale_multinomial_formula_penalty(*inner, scale)),
        },
    }
}

/// Build a warm-started copy of `blocks` whose per-block `initial_log_lambdas`
/// are seeded from a previously-selected flat `log_lambdas` vector (#1082).
///
/// The flat `log_lambdas` returned by [`fit_custom_family_with_rho_prior`]
/// concatenates each block's penalty log-λ in block order — the same order
/// `build_block_specs()` emits the blocks and the same per-block penalty order
/// the spec carries — so it splits back across blocks by each block's penalty
/// count. Warm-starting the OUTER ρ-search from a prior iterate changes only the
/// optimizer's starting point, never the penalized objective or its optimum, so
/// the converged fit is identical; it just resumes near the prior iterate
/// instead of restarting from the cold `init_lambda` seed.
///
/// Returns `None` (caller falls back to the cold blocks) if the flat vector does
/// not have exactly one entry per penalty across all blocks, or carries a
/// non-finite value — i.e. anything that would make the seed unsafe.
fn warm_start_blocks_from_log_lambdas(
    blocks: &[crate::custom_family::ParameterBlockSpec],
    log_lambdas: &[f64],
) -> Option<Vec<crate::custom_family::ParameterBlockSpec>> {
    let total: usize = blocks.iter().map(|b| b.initial_log_lambdas.len()).sum();
    if total == 0 || log_lambdas.len() != total {
        return None;
    }
    if log_lambdas.iter().any(|v| !v.is_finite()) {
        return None;
    }
    let mut warm = blocks.to_vec();
    let mut offset = 0usize;
    for block in warm.iter_mut() {
        let k = block.initial_log_lambdas.len();
        for slot in 0..k {
            block.initial_log_lambdas[slot] = log_lambdas[offset + slot];
        }
        offset += k;
    }
    Some(warm)
}

/// Top-level formula-driven multinomial fit.
///
/// Routes through [`fit_custom_family_with_rho_prior`] so the per-active-class
/// smoothing parameters `λ_a` (one per class block, shared-penalty
/// architecture) are selected by the outer REML/LAML loop rather than pinned
/// by the caller. `init_lambda` survives as a warm-start hint that seeds
/// every block's `initial_log_lambdas`. `max_iter` / `tol` drive the OUTER
/// REML/LAML smoothing-parameter search (`outer_max_iter` / `outer_tol`); the
/// inner joint-Newton solve runs on the framework's principled production cycle
/// budget at the default KKT tolerance so an ill-conditioned, LM-damped
/// near-simplex-boundary solve can certify a stationary point instead of being
/// declared non-converged after only `max_iter` cycles (#715).
///
/// The Jeffreys/Firth proper prior is engaged CONDITIONALLY: attempt 1 runs
/// the unbiased penalized-REML criterion; only on separation evidence (a failed
/// solve or a non-finite logit; see [`multinomial_formula_separation_evidence`])
/// is the fit re-solved once with the full-span Firth prior armed, which bounds
/// the penalty-null directions no smoothing parameter can (`S v = 0` ⇒
/// `(H + S_λ) v = H v → 0` when the softmax likelihood has no finite mode).
///
/// The categorical response column is recognised via the dataset schema
/// (`ColumnKindTag::Categorical`); reference class = last level. Returns a
/// [`MultinomialSavedModel`] that can be serialised to bytes for the Python
/// wrapper or used in-process for `predict_probabilities`.
pub fn fit_penalized_multinomial_formula(
    data: &EncodedDataset,
    formula: &str,
    config: &FitConfig,
    init_lambda: f64,
    max_iter: usize,
    tol: f64,
) -> Result<MultinomialSavedModel, EstimationError> {
    if !(init_lambda.is_finite() && init_lambda > 0.0) {
        crate::bail_invalid_estim!(
            "multinomial fit: init_lambda must be finite and > 0 (got {init_lambda})"
        );
    }
    let (raw_spec, design, y_col, response_name, y_kind) =
        build_formula_design_for_multinomial(formula, data, config)?;
    // Freeze the data-derived basis state (B-spline knot vectors, by-factor
    // level sets, spatial centers, joint-null rotations, residualization
    // charts) from the fit design back onto the spec. The raw geometry spec
    // records only *which* columns and *what kind* of basis each smooth uses;
    // the actual column count and basis evaluation depend on quantities the
    // builder derives from the training data (knot placement, the distinct
    // by-factor levels, etc.). Saving the raw spec made predict re-derive those
    // from the (smaller, differently-distributed) predict frame, so the rebuilt
    // design had a different column count than the fitted one — the panic
    // "predict design has 42 cols, saved model expects 191" for an `s(x,
    // by=group)` smooth-by-factor model. Every other family's persistence path
    // freezes the spec the same way (see `freeze_term_collection_from_design`
    // call sites in `main_parts`); multinomial was the lone exception.
    let spec = freeze_term_collection_from_design(&raw_spec, &design)?;
    let class_levels = match y_kind {
        ResponseColumnKind::Categorical { levels } => levels,
        ResponseColumnKind::Binary => vec!["0".to_string(), "1".to_string()],
        ResponseColumnKind::Numeric => {
            crate::bail_invalid_estim!(
                "multinomial fit: response '{response_name}' is numeric, not categorical; \
                 use family='gaussian'/'binomial'/... or convert the column to a categorical type"
            );
        }
    };
    if data.column_kinds.get(y_col) == Some(&ColumnKindTag::Binary) {
        // Promote to a 2-level categorical for the multinomial driver; the
        // caller explicitly asked for multinomial, so we route through the
        // K-1 = 1 active-class softmax (equivalent math to logistic).
    } else if data.column_kinds.get(y_col) != Some(&ColumnKindTag::Categorical) {
        crate::bail_invalid_estim!(
            "multinomial fit: response '{response_name}' must be a categorical column \
             (got column kind {:?})",
            data.column_kinds.get(y_col)
        );
    }
    let (y_one_hot, _) = one_hot_categorical_response(data, y_col, &response_name)?;
    // Build the global X dense (the design is a DesignMatrix abstraction).
    let mut x_dense = design
        .design
        .try_to_dense_by_chunks("multinomial fit design")
        .map_err(EstimationError::InvalidInput)?;

    // ── #715 real-data conditioning: standardize unpenalized parametric
    // columns. Raw-unit linear covariates (penguins `body_mass_g` ~ 4e3 grams)
    // inflate the joint Newton information by the squared column scale (a κ(H)
    // multiplier of ~s² ≈ 1e7 against the intercept), which is what turns the
    // near-separable LM-damped inner solve into a geometric grind that
    // exhausts its cycle budgets — the adapter-level face of "all REML startup
    // seeds rejected". Because these columns are UNPENALIZED (parametric terms
    // carry no default ridge, #749), the affine reparameterization
    // `x_j ↦ (x_j − m_j)/s_j` is EXACT for the whole criterion: the optimized
    // REML/LAML objective, the fitted η, the selected λ, and the separation
    // diagnostics are all invariant — only the conditioning of `H` changes.
    // Fitted coefficients are mapped back to raw units at repack below, so the
    // saved model and the (raw-design) predict path are untouched. Penalized
    // columns are left alone (a penalty makes the rescaling non-equivalent),
    // and nothing is touched when explicit coefficient bounds/constraints
    // exist (those are stated in raw units).
    let parametric_standardization: Vec<(usize, f64, f64)> =
        if design.coefficient_lower_bounds.is_some() || design.linear_constraints.is_some() {
            Vec::new()
        } else {
            let p_total = x_dense.ncols();
            let mut penalized = vec![false; p_total];
            for bp in &design.penalties {
                for col in bp.col_range.clone() {
                    if col < p_total {
                        penalized[col] = true;
                    }
                }
            }
            let has_intercept = !design.intercept_range.is_empty();
            let n_rows = x_dense.nrows().max(1) as f64;
            let mut standardized = Vec::new();
            for (_, range) in &design.linear_ranges {
                for col in range.clone() {
                    if col >= p_total || penalized[col] {
                        continue;
                    }
                    let column = x_dense.column(col);
                    let mean = column.sum() / n_rows;
                    let var = column.iter().map(|v| (v - mean) * (v - mean)).sum::<f64>() / n_rows;
                    let scale = var.sqrt();
                    // Skip near-constant or degenerate columns: no conditioning to
                    // be gained and the back-map would divide by ~0.
                    if !(scale.is_finite() && scale > 1e-8 * (mean.abs() + 1.0)) {
                        continue;
                    }
                    // Centering shifts mass onto the intercept; without one the
                    // shift is not representable, so scale only.
                    let center = if has_intercept { mean } else { 0.0 };
                    for v in x_dense.column_mut(col).iter_mut() {
                        *v = (*v - center) / scale;
                    }
                    standardized.push((col, center, scale));
                }
            }
            standardized
        };
    // Preserve the per-smooth-term penalty block structure (#561): each smooth
    // term `t` contributes its own `P × P` penalty component (`Blockwise` with
    // `total_dim = P`, the term's local `S_t` embedded at its `col_range`), and
    // every active class block receives the FULL list. The outer REML/LAML loop
    // then selects an independent smoothing parameter λ_{a,t} per (class, term),
    // matching mgcv/VGAM. Pre-summing the terms into one fused `S` (the prior
    // behaviour) forced a single λ per class that scales `Σ_t S_t`, so one
    // shared λ had to over-smooth a rough term while under-smoothing a smooth
    // one — biasing any multi-term class-probability surface.
    let k = y_one_hot.ncols();
    let m = k - 1;
    let n_obs = y_one_hot.nrows();
    let penalty_scale = multinomial_formula_penalty_scale(k);
    let per_term_penalties: Vec<PenaltyMatrix> = design
        .penalties_as_penalty_matrix()
        .into_iter()
        .map(|penalty| scale_multinomial_formula_penalty(penalty, penalty_scale))
        .collect();
    let per_term_nullspace_dims = design.nullspace_dims.clone();

    // ── Custom-family driven REML/LAML path ───────────────────────────────
    // Each active class becomes one ParameterBlockSpec, all sharing X and the
    // per-term penalty list. `initial_log_lambdas` is seeded from the caller's
    // `init_lambda` (one entry per term).
    let design_arc = Arc::new(x_dense);
    let penalties_arc = Arc::new(per_term_penalties);
    let nullspace_dims_arc = Arc::new(per_term_nullspace_dims);
    let weights = Array1::<f64>::ones(n_obs);
    // First attempt runs the UNBIASED penalized-REML criterion (no Firth
    // shrinkage toward the uniform simplex); the Jeffreys/Firth proper prior is
    // armed conditionally below, only on separation evidence (#715/#753 — see
    // `multinomial_formula_separation_evidence`).
    let log_init = init_lambda.ln();
    let family = MultinomialFamily::new(
        y_one_hot.clone(),
        weights,
        k,
        design_arc.clone(),
        penalties_arc.clone(),
        nullspace_dims_arc.clone(),
    )
    .map_err(EstimationError::InvalidInput)?
    .with_joint_jeffreys_term(false)
    // gam#1587: the per-block smooth penalties are emptied (the centered `M⊗S_t`
    // joint penalty is the sole smoothing carrier), so the `init_lambda` warm
    // start must seed the JOINT penalty's `initial_log_lambda` — the per-block
    // `initial_log_lambdas` loop below is now a no-op (empty per-block list).
    .with_initial_log_lambda(log_init);
    let mut blocks = family.build_block_specs();
    for spec_block in blocks.iter_mut() {
        for v in spec_block.initial_log_lambdas.iter_mut() {
            *v = log_init;
        }
    }

    // ── Outer-derivative policy: dimension-gated exact curvature ────────────
    // The total smoothing-parameter dimension is `D = (K−1) · n_terms`.
    // Medium-D formula fits need exact curvature to keep lambda selection away
    // from over-smoothed caps, while smooth-by-factor `D = 8` models still avoid
    // the O(D²) dense Hessian path.
    let total_rho_dim = m.saturating_mul(penalties_arc.len());
    let use_outer_hessian = multinomial_formula_use_outer_hessian(total_rho_dim);

    // ── Inner-vs-outer control split (#715 non-convergence root cause) ────────
    // The legacy `max_iter` / `tol` parameters are the *outer* REML/LAML
    // smoothing-parameter optimization controls — "how hard to search λ". The
    // earlier wiring routed them straight into `inner_max_cycles` / `inner_tol`,
    // capping the joint-Newton inner solve at `max_iter` (=50 in the quality
    // suite) cycles with a `tol`-tight (=1e-8) KKT target. That is the #715
    // hang: near the simplex boundary the softmax Fisher weight
    // `W = diag(p) − p pᵀ` collapses, so `H = JᵀWJ + S_λ` is full-rank but
    // ILL-CONDITIONED. The self-vanishing Levenberg–Marquardt damping
    // (`levenberg_on_ill_conditioning()`) that keeps the inner solve from
    // oscillating on those near-singular modes makes it converge only
    // GEOMETRICALLY (linearly), not quadratically. Reaching a 1e-8 relative KKT
    // residual under geometric descent needs FAR more than 50 cycles, so the
    // inner returned `converged = false` on every outer ρ-evaluation; with the
    // exact-Hessian outer optimizer on `FallbackPolicy::Disabled` that rejects
    // every ρ-step — each rejected eval still paying a near-full 50-cycle inner
    // solve plus the O(D²) pairwise outer-Hessian directional work — so the
    // outer never certifies and the fit runs unbounded (the observed >8-minute
    // non-termination). The certificate cannot be reached, not merely slow.
    //
    // Fix: give the INNER joint-Newton the framework's principled production
    // budget (`DEFAULT_CUSTOM_FAMILY_INNER_MAX_CYCLES` cycles at the default
    // `inner_tol`), which exists precisely so an ill-conditioned LM-damped solve
    // can certify a stationary KKT point instead of being declared non-converged
    // prematurely — and the KKT/objective certificates still exit in a handful
    // of cycles on the well-conditioned interior fits, so this is free there.
    // The caller's `max_iter` / `tol` become the OUTER controls they were always
    // meant to be (smoothing-parameter search depth / accuracy). The inner KKT
    // target is kept no tighter than the outer accuracy can consume — and no
    // tighter than the softmax objective's f64 noise floor on near-separable
    // fits (see `MULTINOMIAL_FORMULA_INNER_TOL`).
    let outer_max_iter = max_iter.max(1);
    // The OUTER REML/LAML smoothing-parameter search must converge to a
    // well-calibrated ρ-gradient tolerance, NOT to the caller's (typically very
    // tight) INNER KKT tolerance. The #715 control-split repurposed the caller's
    // `tol` as the outer control, but feeding an inner-scale `tol = 1e-8`
    // straight into `outer_tol` makes REML grind dozens of extra exact-gradient
    // outer iterations (each an O(D·p³) Laplace-derivative assembly over the full
    // P·M joint design) to squeeze ρ digits that no longer move the fitted
    // surface — the smooth-by-factor 269s wall-clock overrun (#1082).
    //
    // The right target is the framework's CALIBRATED REML convergence tolerance,
    // `MULTINOMIAL_OUTER_REML_TOL = 1e-7` — the same value the primary GLM REML
    // outer uses (`solver::fit_orchestration::materialize` `tol: 1e-7`, mirrored by the
    // `LOG_LAMBDA_TOL`/`KKT_TOL_*` constants across the REML stack). At 1e-7 the
    // λ-search reaches the genuine REML optimum (so the recovered probability
    // surface matches the mature reference), but it does NOT chase the last
    // surface-irrelevant ρ digits down to 1e-8. The earlier 1e-5 floor (the
    // generic `BlockwiseFitOptions` default) was too LOOSE: the optimizer halted
    // in a low-curvature region with λ still well above its optimum, UNDER-fitting
    // the smooth-by-factor surface (truth-RMSE 0.164 vs VGAM's 0.061). So the
    // outer tolerance is floored at the calibrated REML tol — never tighter than
    // it (perf), never looser (accuracy) — while the caller's `tol` continues to
    // drive the INNER joint-Newton KKT target (`inner_tol` below), where its
    // precision actually matters.
    let outer_tol = if tol.is_finite() && tol > 0.0 {
        tol.max(MULTINOMIAL_OUTER_REML_TOL)
    } else {
        MULTINOMIAL_OUTER_REML_TOL
    };
    // #1082 root cause: the outer convergence test derives BOTH the absolute
    // projected-gradient floor (`max(outer_tol, n·1e-9)`) AND the relative-cost
    // stop (`rel_cost = outer_tol`) from the single `outer_tol`. The accuracy of
    // the smooth-by-factor surface is governed by the ABSOLUTE floor reaching the
    // n-scaled REML resolution `n·1e-9` (≈ 1.8e-6 at n = 1800) — that is why the
    // earlier 1e-5 floor UNDER-fit (its absolute floor was pinned at 1e-5, well
    // above the genuine optimum's gradient) and why 1e-7 recovered accuracy (it
    // unpins the floor down to the n-scaled 1.8e-6). But tightening `outer_tol`
    // to 1e-7 ALSO tightened the rel-cost stop to 1e-7, which on this family's
    // dead-flat REML ridge NEVER trips — so the optimizer no longer converges and
    // grinds all the way to `outer_max_iter`, each surplus step an O(D·p³) Laplace-
    // derivative assembly over the 382-dim joint design (the >600s wall-clock
    // overrun; tightening tol REINTRODUCED the crawl the 1e-5 floor had removed).
    //
    // The two requirements live on two different criteria, so they must be set
    // independently. Keep `outer_tol = 1e-7` (drives the accurate absolute floor)
    // but FLOOR the relative-cost stop at the framework default 1e-5 (the loose,
    // fast value that resolves the cost-decrease plateau without chasing the flat
    // tail). The absolute n·1e-9 floor still gates final λ accuracy; the rel-cost
    // stop just lets the optimizer DECLARE convergence on the flat ridge instead
    // of crawling to the iteration cap.
    let outer_rel_cost_tol = Some(BlockwiseFitOptions::default().outer_tol);
    let inner_tol = MULTINOMIAL_FORMULA_INNER_TOL.max(tol.max(0.0));

    let options = BlockwiseFitOptions {
        inner_max_cycles: crate::custom_family::DEFAULT_CUSTOM_FAMILY_INNER_MAX_CYCLES,
        inner_tol,
        outer_max_iter,
        outer_tol,
        outer_rel_cost_tol,
        rho_lower_bound: multinomial_formula_min_lambda(y_one_hot.view()).ln(),
        ridge_floor: MULTINOMIAL_FORMULA_RIDGE_FLOOR,
        // #747: the stabilization floor is SOLVER-ONLY — it keeps the inner
        // joint-Newton linear solve finite during screening (bounding the step
        // `(H+δI)⁻¹∇` away from a near-separable, rank-deficient curvature) but
        // is excluded from the REML objective, the penalty log-determinant, and
        // the Laplace Hessian. The earlier default (`explicit_stabilization_pospart`)
        // folded `½·δ·‖β‖²` and a `δ`-shift of the log-determinant into the
        // criterion, shrinking every identified coefficient off the MLE and
        // perturbing smoothing-parameter selection — a fixed-λ prior masking
        // separation, not a numerical stabilizer. With the floor solver-only the
        // optimized objective is the true penalized REML criterion (value tracks
        // its analytic gradient), and the smooth directions remain governed
        // solely by their own REML-selected `λ`.
        ridge_policy: gam_problem::RidgePolicy::solver_only(),
        use_outer_hessian,
        // #715 real-data arm ("canonical-gauge null direction rejects all REML
        // seeds"): skip the multi-seed outer screening cascade and let the
        // pinned `init_lambda` ρ flow straight to the outer optimizer.
        //
        // The multinomial family declares `levenberg_on_ill_conditioning() ->
        // true`: near the simplex boundary (the near-separable penguins regime)
        // the softmax Fisher weight `W = diag(p) − p pᵀ → 0`, so the joint
        // information `H = JᵀWJ + S_λ` can become full-rank but
        // ILL-CONDITIONED. The self-vanishing LM damping that keeps the inner
        // joint-Newton from oscillating on those near-singular modes converges
        // only GEOMETRICALLY. The default screening policy ranks candidate seeds
        // with a 2-cycle inner cap (`outer_seed_config`); under geometric
        // LM-damped descent two cycles never reach a finite, meaningful proxy
        // objective, so EVERY capped seed can collapse to non-finite cost and
        // the cascade escalates to ×4, ×16, then an UNCAPPED full inner solve
        // PER SEED on the near-singular Hessian. That is the adapter-level face
        // of "all REML startup seeds rejected" and the multi-minute timeout.
        //
        // The pinned seed is already principled here: `init_lambda` gives every
        // (class, term) ρ a sensible moderate warm start, and the per-term
        // effective-df-floor upper bounds (`effective_df_floor_rho_upper_bounds`,
        // #715 arm (a)) keep any λ from collapsing the smooth onto its polynomial
        // null space. So the outer ARC/BFGS optimizer performs the real REML ρ
        // search from this seed; screening only adds the cascade cost and, on the
        // near-separable arm, the rejection stall.
        screen_initial_rho: false,
        // #1101: compute the joint Laplace posterior covariance `H⁻¹` (and the
        // influence matrix `F = H⁻¹ X'WX`) at the converged mode so the saved
        // model can surface delta-method per-class probability standard errors
        // and Wald smooth-term p-values. The driver factorizes the penalized
        // Hessian during the inner solve regardless; this only asks it to keep
        // and invert the factor instead of discarding it.
        compute_covariance: true,
        ..BlockwiseFitOptions::default()
    };
    // ── Conditional Firth/Jeffreys engagement (#715 arm (b) / #753) ──────────
    // Attempt 1: the unbiased criterion (Jeffreys disarmed above). If the
    // returned mode is converged, finite, and interior, it is the exact penalized-REML
    // optimum with zero Firth bias — accept it (this is the synthetic-arm /
    // interior-data path, #715 arm (a)). If the solve FAILS (e.g. the
    // (quasi-)separated penguins geometry where `(H + S_λ)v ≈ 0` along
    // penalty-null directions for EVERY ρ rejects every REML startup seed) or
    // returns a non-finite artifact, that is direct separation evidence:
    // re-solve once with the full-span Jeffreys/Firth proper prior armed, which
    // supplies the O(1) curvature on the quotient-null subspace that smoothing
    // parameters mathematically cannot (`Sv = 0` ⇒ λ never touches `v`). The
    // Firth refit is the accepted result only when the unbiased formula solve
    // failed, did not converge on its full budget, or blew up; finite
    // formula-path logits can be large on valid near-separated optima and
    // should not be shrunk toward the uniform simplex once the unbiased outer
    // solve has actually certified.
    let mut unbiased_probe_options = options.clone();
    unbiased_probe_options.outer_max_iter = unbiased_probe_options
        .outer_max_iter
        .min(MULTINOMIAL_UNBIASED_PROBE_OUTER_MAX_ITER);
    // The FINAL accepted Firth/Jeffreys refit runs to the caller's full outer
    // budget: it is the result we ship, so it must reach the genuine REML
    // optimum, not a truncated iterate. The near-separable penguin refit that
    // motivated #1082's wall-clock concern is now halted honestly at its true
    // bound optimum by the KKT-stationary-at-bound guard
    // (`CostStallGuard`, #1082 / 64711ed82) and the Newton-decrement residual
    // certificate (363af9b56 / 2c9580b1f): on separable data the outer ARC
    // certifies and stops early on its own, so no artificial iteration cap is
    // needed to land in budget. On non-separable data (e.g. the
    // `vgam_smooth_by_factor` double-penalty arm) the refit needs the caller's
    // full budget to converge, which a `.min(20)` cap would cut off — accepting
    // a non-converged fit, which is dishonest. So the refit keeps `options`
    // unchanged. Only the discarded unbiased separation probe above is capped.
    let firth_refit_options = &options;

    let run_firth_refit = |evidence: String| {
        let firth_family = family.clone().with_joint_jeffreys_term(true);
        fit_custom_family_with_rho_prior(
            &firth_family,
            &blocks,
            firth_refit_options,
            gam_problem::RhoPrior::Flat,
        )
        .map_err(|err| {
            EstimationError::InvalidInput(format!(
                "multinomial REML: Firth/Jeffreys-armed refit (separation evidence: \
                 {evidence}) failed: {err}"
            ))
        })
    };

    // #1082: the capped unbiased probe and the (separable-path) Firth decision
    // are driven by separation scans over the full P×M logit block. The previous
    // match recomputed `multinomial_formula_separation_evidence` /
    // `..._unresolved_probe_separation_evidence` in BOTH the match guard AND the
    // arm body — three to four full logit walks per fit, paid on the hot
    // near-separable penguin path where this branch fires every iterate. Run the
    // probe once, evaluate each scan once into a binding, and branch on the
    // precomputed results. Behaviour is identical (same scans, same order of
    // precedence: converged-interior, unresolved-probe-separation,
    // no-separation-needs-full-solve, otherwise-Firth); only the duplicate
    // O(n·classes) scans are removed.
    let probe_attempt = fit_custom_family_with_rho_prior(
        &family,
        &blocks,
        &unbiased_probe_options,
        gam_problem::RhoPrior::Flat,
    );
    let fit = match probe_attempt {
        Ok(probe_fit) => {
            let separation = multinomial_formula_separation_evidence(&probe_fit.block_states);
            if probe_fit.outer_converged && separation.is_none() {
                // Interior, converged, no separation: accept the probe directly.
                probe_fit
            } else if let Some(evidence) =
                multinomial_formula_unresolved_probe_separation_evidence(&probe_fit.block_states)
            {
                // Non-converged probe already carrying separation-scale logits:
                // hand straight to the proper-prior Firth refit (do not spend the
                // full unbiased budget grinding the λ→0 separable ridge).
                run_firth_refit(format!(
                    "unbiased-criterion REML probe did not converge after {} outer iterations; {evidence}",
                    probe_fit.outer_iterations
                ))?
            } else if separation.is_none() {
                // Interior but the capped probe ran out of iterations without
                // certifying: re-solve at the caller's full outer budget.
                //
                // #1082 wall-clock: the capped probe is a strict prefix of this
                // solve from the same family/seed, so a COLD restart repeats the
                // probe's outer iterations. WARM-START the re-solve from the ρ the
                // probe already reached — seed each block's `initial_log_lambdas`
                // from the probe's selected `log_lambdas` (same block/penalty
                // order: the flat vector concatenates per-block penalties in block
                // order, exactly the order `build_block_specs()` emits them). This
                // changes only the optimizer's STARTING point, never the objective
                // or its optimum, but lets the full solve resume near the probe's
                // last iterate instead of crawling up from `init_lambda` again —
                // removing the probe-iterations double-pay on the non-separable
                // (e.g. `vgam_smooth_by_factor`) arm. If the probe's λ vector does
                // not line up with the block layout (it always should), fall back
                // to the cold `blocks` seed.
                let warm_blocks = warm_start_blocks_from_log_lambdas(
                    &blocks,
                    probe_fit.log_lambdas.as_slice().unwrap_or(&[]),
                );
                let resolve_blocks = warm_blocks.as_deref().unwrap_or(&blocks);
                match fit_custom_family_with_rho_prior(
                    &family,
                    resolve_blocks,
                    &options,
                    gam_problem::RhoPrior::Flat,
                ) {
                    Ok(full_unbiased_fit) => {
                        let full_separation = multinomial_formula_separation_evidence(
                            &full_unbiased_fit.block_states,
                        );
                        if full_unbiased_fit.outer_converged && full_separation.is_none() {
                            full_unbiased_fit
                        } else {
                            let evidence = full_separation.unwrap_or_else(|| {
                                format!(
                                    "full unbiased-criterion REML solve did not converge after {} outer iterations",
                                    full_unbiased_fit.outer_iterations
                                )
                            });
                            run_firth_refit(evidence)?
                        }
                    }
                    Err(err) => run_firth_refit(format!(
                        "full unbiased-criterion REML solve failed: {err}"
                    ))?,
                }
            } else {
                // Probe converged (or capped) but shows interior separation
                // evidence: Firth refit using the already-computed scan.
                let evidence = separation.unwrap_or_else(|| {
                    format!(
                        "unbiased-criterion REML probe did not converge after {} outer iterations",
                        probe_fit.outer_iterations
                    )
                });
                run_firth_refit(evidence)?
            }
        }
        Err(err) => run_firth_refit(format!("unbiased-criterion REML solve failed: {err}"))?,
    };
    if let Some(err) = multinomial_formula_separation_diagnostic(
        fit.inner_cycles,
        fit.outer_iterations,
        &fit.block_states,
    ) {
        return Err(err);
    }

    // ── Repack coefficients (P, K-1) from per-block β vectors ─────────────
    if fit.blocks.len() != m {
        crate::bail_invalid_estim!(
            "multinomial REML: expected {m} fitted blocks (K-1), got {}",
            fit.blocks.len()
        );
    }
    let p_per_class = fit.blocks[0].beta.len();
    let mut coefficients_active = Array2::<f64>::zeros((p_per_class, m));
    for (a, block) in fit.blocks.iter().enumerate() {
        if block.beta.len() != p_per_class {
            crate::bail_invalid_estim!(
                "multinomial REML: block {a} has {} coefs, expected {p_per_class}",
                block.beta.len()
            );
        }
        for i in 0..p_per_class {
            coefficients_active[[i, a]] = block.beta[i];
        }
    }
    // Map the standardized-column coefficients back to raw units (the exact
    // inverse of the conditioning reparameterization above): β_raw = b/s, with
    // the centering mass `Σ_j b_j·m_j/s_j` returned to the intercept.
    if !parametric_standardization.is_empty() {
        let intercept_col = design.intercept_range.clone().next();
        for a in 0..m {
            let mut intercept_adjust = 0.0;
            for &(col, center, scale) in &parametric_standardization {
                if col < p_per_class {
                    let raw = coefficients_active[[col, a]] / scale;
                    coefficients_active[[col, a]] = raw;
                    intercept_adjust += raw * center;
                }
            }
            if let Some(i0) = intercept_col
                && i0 < p_per_class
            {
                coefficients_active[[i0, a]] -= intercept_adjust;
            }
        }
    }
    // Flatten every (class, term) smoothing parameter in block-major order
    // (class 0's terms, then class 1's, …). With per-term penalties each block
    // now carries one λ per smooth term, so a single λ per class would discard
    // the independent per-term selection that fixes #561. `lambdas_per_block`
    // segments the flat vector by class so callers can recover per-term λ.
    // ── gam#1587/#561 joint-penalty reconstruction ───────────────────────────
    // Under the #1587 centered-metric architecture every active class block
    // leaves its per-block penalty list EMPTY — the entire fit's smoothing rides
    // on a single full-width JOINT penalty `S_λ = Σ_t λ_t (M ⊗ S_t)` whose one
    // shared `λ_t` per smooth component is selected by the outer REML loop and
    // surfaced on `fit.artifacts.joint_log_lambdas`. So `fit.blocks[a].lambdas`
    // is `[]`, the inference layer's per-block trace channel is empty, and the
    // older per-block reporting (`lambdas_per_block = [0, 0]`, `edf_per_class =
    // None`, …) collapsed (#561 reopen).
    //
    // Reconstruct the per-(class, component) λ and the influence-matrix EDF
    // directly from the selected joint `λ_t` and the COUPLED penalty
    // `S_λ = Σ_t λ_t (M ⊗ S_t)` (NOT a block-diagonal `Σ_t λ_{a,t} S_t`: the
    // centered metric `M` couples classes off the block diagonal, so a
    // block-diagonal `S_λ` would mis-state both the influence matrix and every
    // trace). With `H⁻¹ = fit.covariance_conditional` now assembled WITH the
    // joint penalty (the `compute_joint_covariance` fix), the influence matrix is
    // exactly `F = I − H⁻¹ S_λ`, its per-class diagonal-block trace is the honest
    // per-class EDF, and `Σ_a edf_a = tr(F) = edf_total`.
    let joint_recon = fit.artifacts.joint_log_lambdas.as_ref().and_then(|jll| {
        let n_components = penalties_arc.len();
        if jll.len() != n_components || n_components == 0 {
            return None;
        }
        let expected_joint = p_per_class.saturating_mul(m);
        let hinv = fit
            .covariance_conditional
            .as_ref()
            .filter(|c| c.nrows() == expected_joint && c.ncols() == expected_joint)?;
        // The coupled joint penalty components `M ⊗ S_t` at the selected `λ_t`,
        // in raw stacked (class-major) coordinates — exactly the operator the
        // inner solve and the now-fixed covariance path penalize with.
        let joint_specs = family.centered_joint_penalty_specs();
        if joint_specs.len() != n_components {
            return None;
        }
        let lam: Vec<f64> = jll.iter().map(|&l| l.exp()).collect();
        // Per-component `H⁻¹ (M ⊗ S_t)` (full mp×mp), reused for both the joint
        // influence matrix and the per-(class, component) trace decomposition.
        let mut hinv_st: Vec<Array2<f64>> = Vec::with_capacity(n_components);
        for spec in &joint_specs {
            if spec.matrix.nrows() != expected_joint || spec.matrix.ncols() != expected_joint {
                return None;
            }
            hinv_st.push(hinv.dot(&spec.matrix));
        }
        // F = I − H⁻¹ S_λ = I − Σ_t λ_t H⁻¹ (M ⊗ S_t).
        let mut f = Array2::<f64>::eye(expected_joint);
        for (t, hs) in hinv_st.iter().enumerate() {
            f.scaled_add(-lam[t], hs);
        }
        // Per-class diagonal-block trace of F (the honest per-class EDF), and the
        // per-(class, component) penalty trace `tr_{a,t} = λ_t · Σ_{i∈class a}
        // (H⁻¹ (M⊗S_t))[i,i]` for the per-penalty EDF rollup.
        let mut edf_per_class = Vec::with_capacity(m);
        // class-major per-penalty EDF (class 0's components, then class 1's, …),
        // aligned 1:1 with the flat per-component λ replicated per class.
        let mut edf_per_penalty = Vec::with_capacity(m * n_components);
        for a in 0..m {
            let base = a * p_per_class;
            let mut class_trace = 0.0_f64;
            for t in 0..n_components {
                let mut tr_at = 0.0_f64;
                for i in 0..p_per_class {
                    tr_at += hinv_st[t][[base + i, base + i]];
                }
                tr_at *= lam[t];
                class_trace += tr_at;
                // A single component's per-class trace EDF `rank(S_t) − tr_{a,t}`,
                // bounded by its local rank (≤ p_per_class).
                let ns_t = nullspace_dims_arc.get(t).copied().unwrap_or(0);
                let rank_t = (p_per_class as f64 - ns_t as f64).max(0.0);
                edf_per_penalty.push((rank_t - tr_at).clamp(0.0, p_per_class as f64));
            }
            edf_per_class
                .push((p_per_class as f64 - class_trace).clamp(0.0, p_per_class as f64));
        }
        Some((f, edf_per_class, edf_per_penalty, n_components, lam))
    });

    // Flatten every (class, component) smoothing parameter in class-major order.
    // Under the joint-penalty architecture each active class carries the SAME
    // per-component λ set (the centered metric ties `λ_t` across classes for
    // reference-class invariance), so the flat vector is the selected `λ_t`
    // replicated `K-1` times and `lambdas_per_block = [n_components; K-1]`. When
    // the joint reconstruction is unavailable (legacy fixed-λ path or absent
    // covariance) fall back to the raw — now empty — per-block λ lists.
    let (lambdas_per_block, lambdas_flat): (Vec<usize>, Vec<f64>) = match joint_recon.as_ref() {
        Some((_, _, _, n_components, lam)) => {
            let per_block = vec![*n_components; m];
            let mut flat = Vec::with_capacity(m * n_components);
            for _ in 0..m {
                flat.extend(lam.iter().copied());
            }
            (per_block, flat)
        }
        None => {
            let per_block: Vec<usize> = fit.blocks.iter().map(|b| b.lambdas.len()).collect();
            let flat: Vec<f64> = fit
                .blocks
                .iter()
                .flat_map(|b| b.lambdas.iter().copied())
                .collect();
            (per_block, flat)
        }
    };
    // Per-active-class effective degrees of freedom, length `K-1`, summing to
    // the model `edf_total`. The REML inference block reports `edf_by_block` as
    // ONE entry per *penalty block* (per (class, term, penalty)), each computed
    // as `rank(S_kk) − tr(H⁻¹ λ_kk S_kk)`. That per-block sum OVER-COUNTS the
    // model EDF whenever several penalties share one coefficient range — a
    // double-penalty / te / ti / adaptive smooth has ≥2 penalty blocks over the
    // same columns, so `Σ_kk rank(S_kk) > p` and `Σ_kk edf_by_block > edf_total`
    // (the observed ~79 for a ~24-coefficient model). Handing that raw per-block
    // vector out as the documented length-(K-1) per-class EDF is therefore both
    // the wrong LENGTH (it is `Σ_a n_blocks_a`, not `K-1`) and an over-count.
    //
    // The honest per-class EDF is the influence-matrix trace over each class's
    // coefficient block. Classes occupy DISJOINT `p_per_class`-wide coefficient
    // ranges, and the per-block traces `tr_kk = tr(H⁻¹ λ_kk S_kk)` are additive
    // (no rank double-counting), so class `a`'s EDF is
    // `p_per_class − Σ_{kk ∈ class a} tr_kk`, and `Σ_a edf_a = m·p_per_class −
    // Σ_kk tr_kk = p − Σ tr_kk = edf_total` exactly. Segment the block-major
    // `penalty_block_trace` by `lambdas_per_block` (the same per-class λ-count
    // segmentation `lambdas_flat` uses). Fall back to `None` when the trace
    // channel is unavailable or mis-shaped (legacy fixed-λ path), exactly as the
    // raw `edf_by_block` map did before.
    let edf_per_class = joint_recon
        .as_ref()
        .map(|(_, epc, _, _, _)| epc.clone())
        .or_else(|| {
            // Legacy per-block trace path (fixed-λ / pre-#1587 fits whose
            // smoothing is still carried per block). Segment the block-major
            // `penalty_block_trace` by `lambdas_per_block`, exactly as before.
            fit.inference.as_ref().and_then(|info| {
                let traces = &info.penalty_block_trace;
                if traces.len() != lambdas_per_block.iter().sum::<usize>() {
                    return None;
                }
                let mut per_class = Vec::with_capacity(m);
                let mut cursor = 0usize;
                for &n_blocks in &lambdas_per_block {
                    let class_trace: f64 = traces[cursor..cursor + n_blocks].iter().sum();
                    per_class
                        .push((p_per_class as f64 - class_trace).clamp(0.0, p_per_class as f64));
                    cursor += n_blocks;
                }
                Some(per_class)
            })
        });
    // Per-PENALTY EDF: the inference layer's `edf_by_block` is already the
    // clamped per-penalty-block trace EDF `rank(S_k) − λ_k·tr(H⁻¹ S_k)`, one
    // entry per smoothing parameter and block-major aligned 1:1 with the flat
    // `lambdas`. Surface it verbatim (guarding only on the length contract) so
    // consumers can inspect per-(class, term, penalty) collapse onto the null
    // space — a signal the per-class EDF SUM hides. This is NOT a per-class
    // total: with double-penalty smooths `Σ_k rank(S_k) > p_per_class`, so the
    // entries deliberately need not sum to the model EDF (the per-class field
    // carries that contract instead).
    let edf_per_penalty = joint_recon
        .as_ref()
        .map(|(_, _, epp, _, _)| epp.clone())
        .or_else(|| {
            // Legacy per-block path: the inference layer's `edf_by_block` is
            // already the clamped per-penalty-block trace EDF, aligned 1:1 with
            // the flat `lambdas`.
            fit.inference.as_ref().and_then(|info| {
                if info.edf_by_block.len() != lambdas_flat.len() {
                    return None;
                }
                Some(
                    info.edf_by_block
                        .iter()
                        .map(|&e| e.max(0.0))
                        .collect::<Vec<f64>>(),
                )
            })
        });
    let coefficients_flat: Vec<f64> = coefficients_active.iter().copied().collect();

    // #1101: surface the joint Laplace posterior covariance `H⁻¹` (block-ordered
    // [β_0; …; β_{K-2}]) and the influence matrix `F = H⁻¹ X'WX` the REML driver
    // computed at the converged mode. These power the predict path's delta-method
    // per-class probability standard errors and the summary's Wald smooth-term
    // tests. The joint matrices are `(P·M)×(P·M)`. The covariance is mapped back
    // to RAW units (see below) so it pairs with the raw predict design; the
    // influence is kept in the fitted basis (the Wald table only slices penalized
    // columns, which the standardization affine leaves identity-mapped).
    let expected_joint = p_per_class.saturating_mul(m);
    // The joint Hessian (and thus `H⁻¹`) was assembled in the STANDARDIZED
    // parametric basis used during fitting, while the saved coefficients and the
    // raw predict design are in raw units. Map the covariance to raw units with
    // the same exact affine reparameterization `β_raw = A β_std`: for each
    // standardized parametric column `col`, `β_raw[col] = β_std[col]/scale` and
    // the intercept absorbs `−Σ_col (center/scale)·β_std[col]`. So `A = I` except
    // `A[col,col] = 1/scale` and `A[i0,col] = −center/scale`, replicated
    // block-diagonally per active class, and `Cov_raw = A Cov_std Aᵀ`. With no
    // standardization (`parametric_standardization` empty) `A = I` and this is a
    // no-op. The smooth-term (penalized) columns are untouched by `A`, so the
    // Wald table's per-term blocks are identical in both bases.
    let intercept_col0 = design.intercept_range.clone().next();
    let build_per_class_affine = |amat: &mut Array2<f64>| {
        for &(col, center, scale) in &parametric_standardization {
            if col >= p_per_class {
                continue;
            }
            amat[[col, col]] = 1.0 / scale;
            if let Some(i0) = intercept_col0
                && i0 < p_per_class
            {
                amat[[i0, col]] = -center / scale;
            }
        }
    };
    let coefficient_covariance_flat = fit
        .covariance_conditional
        .as_ref()
        .filter(|c| c.nrows() == expected_joint && c.ncols() == expected_joint)
        .map(|cov_std| {
            if parametric_standardization.is_empty() {
                return cov_std.iter().copied().collect::<Vec<f64>>();
            }
            // Block-diagonal joint A (same per active class).
            let mut a_joint = Array2::<f64>::eye(expected_joint);
            let mut a_class = Array2::<f64>::eye(p_per_class);
            build_per_class_affine(&mut a_class);
            for a in 0..m {
                let base = a * p_per_class;
                for i in 0..p_per_class {
                    for j in 0..p_per_class {
                        a_joint[[base + i, base + j]] = a_class[[i, j]];
                    }
                }
            }
            let cov_raw = a_joint.dot(cov_std).dot(&a_joint.t());
            cov_raw.iter().copied().collect::<Vec<f64>>()
        });
    // The influence matrix `F = H⁻¹ X'WX = H⁻¹(H − S_λ) = I − H⁻¹ S_λ`. The
    // exact-Newton multinomial blocks carry no IRLS pseudo-data, so the generic
    // inference path does not export `coefficient_influence`; reconstruct it
    // exactly here. Under the #1587 joint-penalty architecture the penalty is the
    // COUPLED centered metric `S_λ = Σ_t λ_t (M ⊗ S_t)` (off the class-block
    // diagonal), already assembled in `joint_recon` above, so reuse that exact
    // `F`. Only fall back to the legacy block-diagonal `Σ_t λ_{a,t} S_t`
    // reconstruction when the joint reconstruction is unavailable (pre-#1587
    // per-block fits whose class blocks still carry their own penalties).
    let coefficient_influence_flat = match joint_recon.as_ref() {
        Some((f, _, _, _, _)) => Some(f.iter().copied().collect::<Vec<f64>>()),
        None => fit
            .covariance_conditional
            .as_ref()
            .filter(|c| c.nrows() == expected_joint && c.ncols() == expected_joint)
            .and_then(|hinv| {
                if fit.blocks.len() != m {
                    return None;
                }
                // Joint S_λ (block-diagonal across active classes).
                let mut s_lambda = Array2::<f64>::zeros((expected_joint, expected_joint));
                for (a, block) in fit.blocks.iter().enumerate() {
                    if block.lambdas.len() != penalties_arc.len() {
                        return None;
                    }
                    let base = a * p_per_class;
                    for (t, pen) in penalties_arc.iter().enumerate() {
                        let lam = block.lambdas[t];
                        if lam == 0.0 {
                            continue;
                        }
                        let dense = pen.to_dense();
                        if dense.nrows() != p_per_class || dense.ncols() != p_per_class {
                            return None;
                        }
                        for i in 0..p_per_class {
                            for j in 0..p_per_class {
                                s_lambda[[base + i, base + j]] += lam * dense[[i, j]];
                            }
                        }
                    }
                }
                // F = I − H⁻¹ S_λ.
                let hinv_s = hinv.dot(&s_lambda);
                let mut f = Array2::<f64>::eye(expected_joint);
                f -= &hinv_s;
                Some(f.iter().copied().collect::<Vec<f64>>())
            }),
    };

    // Per-(smooth term) coefficient span within a single class block, deduped by
    // col_range (the #561 double-penalty migration emits two penalty blocks per
    // term sharing one col_range; the Wald test covers the whole term block once).
    let mut smooth_term_spans: Vec<MultinomialSmoothTermSpan> = Vec::new();
    for (pen_idx, bp) in design.penalties.iter().enumerate() {
        let col_start = bp.col_range.start;
        let col_end = bp.col_range.end;
        if col_start >= col_end || col_end > p_per_class {
            continue;
        }
        if smooth_term_spans
            .iter()
            .any(|s| s.col_start == col_start && s.col_end == col_end)
        {
            continue;
        }
        let label = design
            .penaltyinfo
            .get(pen_idx)
            .and_then(|info| info.termname.clone())
            .unwrap_or_else(|| format!("s{pen_idx}"));
        let nullspace_dim = design
            .nullspace_dims
            .get(pen_idx)
            .copied()
            .unwrap_or(0)
            .min(col_end - col_start);
        smooth_term_spans.push(MultinomialSmoothTermSpan {
            label,
            col_start,
            col_end,
            nullspace_dim,
        });
    }

    // One descriptive label per penalty *component* within a single class block,
    // parallel to that block's λ slice (#1544). `design.penalties` is index-
    // parallel to every active class's `block.lambdas` (each block carries the
    // full per-component penalty list, validated above by
    // `block.lambdas.len() == penalties_arc.len()`), so iterating it in order
    // yields exactly `lambdas_per_block[0]` labels aligned with the per-block λ.
    // This is deliberately NOT deduped by col_range (unlike `smooth_term_spans`):
    // the double penalty's primary and null-space components share one col_range
    // but select independent λ, and each must keep its own label so the summary
    // renderer never collapses or drops a λ.
    let lambda_labels: Vec<String> = design
        .penalties
        .iter()
        .enumerate()
        .map(|(pen_idx, _)| penalty_component_label(design.penaltyinfo.get(pen_idx), pen_idx))
        .collect();

    // Unpenalized deviance read directly from the converged unpenalized
    // log-likelihood the rho-prior driver already computed (issue #348):
    // MultinomialFamily::evaluate sets FamilyEvaluation.log_likelihood =
    // log_lik(η, y) with no penalty term, and that value flows unchanged into
    // UnifiedFitResult.log_likelihood. This reproduces the legacy fixed-λ
    // path's `deviance = -2 · log_lik` contract bit-for-bit, so the previous
    // row-by-row η = Xβ rebuild and softmax recompute were pure dead work.
    let deviance = -2.0 * fit.log_likelihood;

    Ok(MultinomialSavedModel {
        formula: formula.to_string(),
        class_levels: class_levels.clone(),
        reference_class_index: class_levels.len() - 1,
        resolved_termspec: spec,
        coefficients_flat,
        p_per_class,
        n_active_classes: m,
        training_headers: data.headers.clone(),
        lambdas: lambdas_flat,
        lambdas_per_block,
        iterations: fit.inner_cycles,
        converged: fit.outer_converged,
        penalized_neg_log_likelihood: -fit.log_likelihood + 0.5 * fit.stable_penalty_term,
        deviance,
        edf_per_class,
        edf_per_penalty,
        coefficient_covariance_flat,
        coefficient_influence_flat,
        smooth_term_spans,
        lambda_labels,
    })
}

/// Replay the saved termspec to build the predict-time dense design `X` on a
/// fresh dataset, realigning feature columns **by name** so the predict frame
/// need not reproduce the training column order or carry the response column.
/// Shared by every multinomial predict path (probabilities, SE bands, and the
/// posterior-predictive replicate draws).
fn build_multinomial_predict_design(
    model: &MultinomialSavedModel,
    data: &EncodedDataset,
) -> Result<Array2<f64>, EstimationError> {
    // The saved termspec stores feature columns as absolute indices into the
    // *training* table `[response, features...]`. Realign them onto this
    // dataset's columns by name, so prediction works on label-free new data
    // (the response column is never referenced by any term; issue #803).
    let predict_columns = data.column_map();
    let realigned = model.resolved_termspec.remap_feature_columns(
        |index| -> Result<usize, EstimationError> {
            let name = model.training_headers.get(index).ok_or_else(|| {
                EstimationError::InvalidInput(format!(
                    "multinomial predict: saved training column index {index} is out of bounds \
                     for {} training headers",
                    model.training_headers.len()
                ))
            })?;
            resolve_role_col(&predict_columns, name, "feature")
                .map_err(|err| EstimationError::InvalidInput(err.to_string()))
        },
    )?;
    let design = build_term_collection_design(data.values.view(), &realigned).map_err(|err| {
        EstimationError::InvalidInput(format!(
            "multinomial predict: rebuild design from saved termspec: {err}"
        ))
    })?;
    let x_dense = design
        .design
        .try_to_dense_by_chunks("multinomial predict design")
        .map_err(EstimationError::InvalidInput)?;
    if x_dense.ncols() != model.p_per_class {
        crate::bail_invalid_estim!(
            "multinomial predict: predict design has {} cols, saved model expects {}",
            x_dense.ncols(),
            model.p_per_class
        );
    }
    Ok(x_dense)
}

/// Replay the saved termspec to build the predict-time design on a fresh
/// dataset, then evaluate softmax probabilities. The predict dataset must carry
/// the same feature columns the training data did, matched **by name** — it need
/// not reproduce the training column order, and in particular need not carry the
/// response column (prediction is for label-free new data).
pub fn predict_multinomial_formula(
    model: &MultinomialSavedModel,
    data: &EncodedDataset,
) -> Result<Array2<f64>, EstimationError> {
    let x_dense = build_multinomial_predict_design(model, data)?;
    Ok(model.predict_probabilities(x_dense.view()))
}

/// Draw `n_draws` posterior-predictive replicate class-label assignments for a
/// saved multinomial model on fresh data (#1101). Rebuilds the predict design
/// exactly as [`predict_multinomial_formula`], then samples each row's class
/// from `Categorical(softmax(X·β̂))` (see
/// [`MultinomialSavedModel::sample_replicate_classes`]). Returns an
/// `(n_draws, N)` matrix of class INDICES `0..K` aligned to `model.class_levels`,
/// deterministic in `seed`.
pub fn posterior_predict_multinomial_formula(
    model: &MultinomialSavedModel,
    data: &EncodedDataset,
    n_draws: usize,
    seed: u64,
) -> Result<Array2<u32>, EstimationError> {
    if n_draws == 0 {
        crate::bail_invalid_estim!("multinomial posterior_predict: n_draws must be >= 1");
    }
    let x_dense = build_multinomial_predict_design(model, data)?;
    Ok(model.sample_replicate_classes(x_dense.view(), n_draws, seed))
}

/// Predict class probabilities AND delta-method per-class probability standard
/// errors for a saved multinomial model on fresh data (#1101). Replays the
/// saved termspec to build the predict design exactly as
/// [`predict_multinomial_formula`], then applies the softmax-Jacobian delta
/// method against the stored joint posterior covariance. Returns
/// `(probs (N,K), prob_se (N,K) | None)`; `prob_se` is `None` for a legacy
/// model fitted before covariance was surfaced.
pub fn predict_multinomial_formula_with_se(
    model: &MultinomialSavedModel,
    data: &EncodedDataset,
) -> Result<(Array2<f64>, Option<Array2<f64>>), EstimationError> {
    let x_dense = build_multinomial_predict_design(model, data)?;
    Ok(model.predict_probabilities_with_se(x_dense.view()))
}

#[cfg(test)]
mod fisher_override_tests {
    use super::*;
    use ndarray::Array3;

    fn toy() -> (Array2<f64>, Array2<f64>, Array2<f64>, Array1<f64>) {
        let n = 15;
        let p = 2;
        let k = 3;
        let design =
            Array2::<f64>::from_shape_fn(
                (n, p),
                |(i, j)| {
                    if j == 0 { 1.0 } else { ((i + 2) as f64).cos() }
                },
            );
        let mut y = Array2::<f64>::zeros((n, k));
        for i in 0..n {
            y[[i, i % k]] = 1.0;
        }
        let penalty = Array2::<f64>::eye(p);
        let lambdas = Array1::<f64>::from_elem(k - 1, 0.5);
        (design, y, penalty, lambdas)
    }

    #[test]
    fn fisher_override_none_reproduces_analytic() {
        // Issue #349: None override is exactly the analytic fit.
        let (design, y, penalty, lambdas) = toy();
        let mk = |over: Option<ndarray::ArrayView3<'_, f64>>| {
            fit_penalized_multinomial(MultinomialFitInputs {
                design: design.view(),
                y_one_hot: y.view(),
                penalty: penalty.view(),
                lambdas: lambdas.view(),
                row_weights: None,
                fisher_w_override: over,
                max_iter: 50,
                tol: 1.0e-9,
            })
            .expect("fit must succeed")
        };
        let a = mk(None);
        let b = mk(None);
        for (x, z) in a
            .coefficients_active
            .iter()
            .zip(b.coefficients_active.iter())
        {
            assert_eq!(x, z);
        }
    }

    #[test]
    fn fisher_override_wrong_shape_is_rejected() {
        let (design, y, penalty, lambdas) = toy();
        let n = design.nrows();
        let m = y.ncols(); // K, not K-1 — deliberately wrong
        let bad = Array3::<f64>::zeros((n, m, m));
        let err = fit_penalized_multinomial(MultinomialFitInputs {
            design: design.view(),
            y_one_hot: y.view(),
            penalty: penalty.view(),
            lambdas: lambdas.view(),
            row_weights: None,
            fisher_w_override: Some(bad.view()),
            max_iter: 50,
            tol: 1.0e-9,
        })
        .expect_err("wrong active-block shape must error");
        assert!(format!("{err}").contains("fisher_w_override shape"));
    }

    /// #1101 regression: the fixed-λ inner solve now surfaces the joint Laplace
    /// coefficient covariance `H⁻¹`, and the multinomial predictor derives
    /// finite delta-method per-class probability standard errors from it. Before
    /// this change `MultinomialFitOutputs` carried NO covariance at all, so the
    /// covariance-dimension / predictor assertions below could not even compile
    /// (fail-before). Asserts, with un-weakened bounds:
    ///   1. covariance is `(P·(K−1))²`, all-finite, symmetric, and PSD (every
    ///      diagonal ≥ 0 and `vᵀΣv ≥ 0` on probe vectors);
    ///   2. the delta-method per-class probability SEs are finite and within
    ///      `[0, 1]` (a probability SE can never exceed the unit interval);
    ///   3. predicted probabilities are finite, in `[0, 1]`, and each row sums
    ///      to 1 (simplex).
    #[test]
    fn covariance_and_delta_method_se_are_finite_and_wellformed_1101() {
        let (design, y, penalty, lambdas) = toy();
        let p = design.ncols();
        let k = y.ncols();
        let m = k - 1;
        let d = p * m;

        let fit = fit_penalized_multinomial(MultinomialFitInputs {
            design: design.view(),
            y_one_hot: y.view(),
            penalty: penalty.view(),
            lambdas: lambdas.view(),
            row_weights: None,
            fisher_w_override: None,
            max_iter: 50,
            tol: 1.0e-9,
        })
        .expect("fit must succeed");
        assert!(fit.converged, "toy multinomial fit must converge");

        // (1) Covariance shape, finiteness, symmetry.
        let cov = &fit.coefficient_covariance;
        assert_eq!(
            cov.dim(),
            (d, d),
            "covariance must be (P·(K−1))² = ({d},{d})"
        );
        for &v in cov.iter() {
            assert!(v.is_finite(), "covariance entry must be finite (got {v})");
        }
        for i in 0..d {
            for j in 0..d {
                let asym = (cov[[i, j]] - cov[[j, i]]).abs();
                assert!(
                    asym <= 1e-9 * (1.0 + cov[[i, j]].abs()),
                    "covariance must be symmetric at ({i},{j}): |Σ_ij − Σ_ji| = {asym:.3e}"
                );
            }
        }
        // PSD: diagonal ≥ 0 and quadratic forms on deterministic probe vectors
        // (unit axes and the all-ones vector) are non-negative. `H = XᵀWX + λS`
        // with W PSD (softmax Fisher) and S PSD (identity here) is positive
        // definite, so its inverse is PD; these probes must all be positive.
        for i in 0..d {
            assert!(
                cov[[i, i]] >= 0.0,
                "covariance diagonal[{i}] must be ≥ 0 (got {})",
                cov[[i, i]]
            );
        }
        let mut probes: Vec<Vec<f64>> = Vec::new();
        for i in 0..d {
            let mut e = vec![0.0_f64; d];
            e[i] = 1.0;
            probes.push(e);
        }
        probes.push(vec![1.0_f64; d]);
        for v in &probes {
            let mut q = 0.0_f64;
            for i in 0..d {
                for j in 0..d {
                    q += v[i] * cov[[i, j]] * v[j];
                }
            }
            assert!(
                q >= -1e-9,
                "covariance must be PSD: vᵀΣv = {q:.3e} < 0"
            );
        }

        // (2) & (3) Delta-method SEs and simplex probabilities on the training
        // design (any P-column matrix in the fitted basis works).
        let (probs, prob_se) = fit
            .predict_probabilities_with_se(design.view())
            .expect("delta-method SE must succeed");
        let n = design.nrows();
        assert_eq!(probs.dim(), (n, k));
        assert_eq!(prob_se.dim(), (n, k));
        for row in 0..n {
            let mut rowsum = 0.0_f64;
            for c in 0..k {
                let pc = probs[[row, c]];
                assert!(pc.is_finite() && (0.0..=1.0).contains(&pc), "prob[{row},{c}]={pc}");
                rowsum += pc;
                let se = prob_se[[row, c]];
                assert!(se.is_finite(), "prob_se[{row},{c}] must be finite (got {se})");
                assert!(
                    (0.0..=1.0).contains(&se),
                    "prob_se[{row},{c}] must be in [0,1] (got {se})"
                );
            }
            assert!(
                (rowsum - 1.0).abs() < 1e-9,
                "row {row} probabilities must sum to 1 (got {rowsum})"
            );
        }
    }

    #[test]
    fn formula_outer_route_uses_exact_curvature_for_medium_d() {
        // The 2-smooth reference formula fit (K = 3, double-penalty terms) is
        // D = (K-1) * 2 terms * 2 penalties = 8 and needs exact curvature to
        // avoid over-smoothed lambda caps (#715 arm (a)).
        assert!(
            multinomial_formula_use_outer_hessian(8),
            "D=8 loaded multinomial fits need exact curvature to avoid over-smoothed lambda caps"
        );
        assert!(
            multinomial_formula_use_outer_hessian(12),
            "D=12 (3 double-penalty smooth terms, K=3) stays on exact curvature"
        );
    }

    #[test]
    fn formula_outer_route_uses_exact_curvature_for_d16_penguin_fixture() {
        // Four k=10 penguin smooths (K = 3) are D = 16 under double-penalty
        // terms. They must reach the exact ARC route so the #1082 cost-stall
        // halt is available on the near-separable lambda-to-zero ridge.
        assert!(
            multinomial_formula_use_outer_hessian(16),
            "D=16 multinomial fits need exact ARC curvature for the #1082 stall halt"
        );
    }

    #[test]
    fn formula_min_lambda_floor_is_continuous_and_information_scaled() {
        // Build a one-hot label matrix whose smallest class carries `count` rows.
        fn floor_for_min_count(count: usize) -> f64 {
            // Two classes: a large one (1000 rows) and a minority one (`count`).
            let n = 1000 + count;
            let mut y = Array2::<f64>::zeros((n, 2));
            for r in 0..1000 {
                y[[r, 0]] = 1.0;
            }
            for r in 1000..n {
                y[[r, 1]] = 1.0;
            }
            multinomial_formula_min_lambda(y.view())
        }

        // The floor's endpoints are now DERIVED from a target prior strength in
        // pseudo-observations against the maximal per-observation softmax Fisher
        // information I₁ = ¼ (base = τ·I₁, sparse = τ_max·I₁). Pin them to the
        // previously fixture-calibrated values so the near-separable quality arms
        // (penguins, vgam softmax) — whose smallest class has n_c ≥ 50 — are
        // byte-for-byte unaffected: the derivation REDUCES TO the old constants
        // at the calibration point.
        let base = MULTINOMIAL_FORMULA_PRIOR_PSEUDO_OBS * MULTINOMIAL_FORMULA_FISHER_INFO_PER_OBS;
        let sparse = MULTINOMIAL_FORMULA_SPARSE_PRIOR_PSEUDO_OBS_MAX
            * MULTINOMIAL_FORMULA_FISHER_INFO_PER_OBS;
        assert!(
            (base - 2.0e-4).abs() < 1e-18,
            "derived base floor must equal the calibrated 2e-4"
        );
        assert!(
            (sparse - 1.0e-3).abs() < 1e-18,
            "derived sparse floor must equal the calibrated 1e-3"
        );

        // Well-supported (n_c >= n_ref=50) sits exactly at the base floor.
        assert!((floor_for_min_count(50) - base).abs() < 1e-18);
        assert!((floor_for_min_count(200) - base).abs() < 1e-18);
        // Very sparse (n_c <= n_ref·base/sparse = 10) clamps to the strong floor.
        assert!((floor_for_min_count(10) - sparse).abs() < 1e-18);
        assert!((floor_for_min_count(5) - sparse).abs() < 1e-18);
        // No cliff at the old hard threshold: 49 vs 50 differ by < 5% (the old
        // step jumped 5x). Floor is monotone non-increasing in support.
        let f49 = floor_for_min_count(49);
        let f50 = floor_for_min_count(50);
        assert!(
            f49 >= f50 && f49 <= f50 * 1.05,
            "floor must be continuous across c0, got {f49} vs {f50}"
        );
        let f25 = floor_for_min_count(25);
        assert!(
            f25 > f50 && f25 < floor_for_min_count(10),
            "mid-support floor must interpolate strictly between the two endpoints"
        );

        // FIRST-PRINCIPLES SCALING: in the interpolating regime the floor equals
        // exactly τ·I₁·(n_ref/n_c) — the effective-pseudo-observation prior held
        // to a fixed fraction of the per-class data information n_c·I₁. Halving
        // the effective sample size doubles the floor (until the cap), and the
        // absolute value matches the closed-form n_c-scaled prior.
        for &n_c in &[12usize, 16, 20, 30, 40] {
            let expected = base * (MULTINOMIAL_FORMULA_SPARSE_REFERENCE_SUPPORT / n_c as f64);
            assert!(
                (floor_for_min_count(n_c) - expected).abs() < 1e-15,
                "floor at n_c={n_c} must be τ·I₁·n_ref/n_c = {expected}, got {}",
                floor_for_min_count(n_c)
            );
        }
        // Inverse scaling with effective sample size: n_c -> n_c/2 doubles the
        // floor inside the unclamped band (20 and 40 are both interior; 40 < 50
        // so it is scaled, 20 > 10 so it is not capped).
        assert!(
            (floor_for_min_count(20) - 2.0 * floor_for_min_count(40)).abs() < 1e-15,
            "floor must scale like 1/n_c (effective Fisher information) in the interior band"
        );
    }

    #[test]
    fn formula_penalty_scale_tracks_softmax_fisher_curvature() {
        assert!(
            (multinomial_formula_penalty_scale(2) - 0.5).abs() < 1.0e-12,
            "binary-logit neutral-simplex curvature scale should remain at 1/2"
        );
        assert!(
            (multinomial_formula_penalty_scale(3) - 4.0 / 9.0).abs() < 1.0e-12,
            "three-class softmax penalties should be calibrated to 2*(K-1)/K^2"
        );
        assert!(
            multinomial_formula_penalty_scale(5) < multinomial_formula_penalty_scale(3),
            "active-class Fisher curvature decreases as the simplex gains classes"
        );
    }

    #[test]
    fn fixed_lambda_multinomial_firth_keeps_complete_separation_finite() {
        // #1854: complete softmax separation used to be a HARD diagnostic
        // (`MultinomialSeparationDetected`). It now automatically engages the
        // Firth/Jeffreys proper prior (`½ log|I(β)|`, magic-by-default) so the fit
        // stays finite instead of running away — the same guarantee the formula
        // REML path already provided. The class regions are cleanly separated by
        // `x`, so the unbiased MLE is at infinity; the Firth-penalized fit must
        // still converge to a finite mode and recover the region structure.
        let n = 90;
        let design = Array2::<f64>::from_shape_fn((n, 2), |(row, col)| match col {
            0 => 1.0,
            _ => -3.0 + 6.0 * (row as f64) / ((n - 1) as f64),
        });
        let mut y = Array2::<f64>::zeros((n, 3));
        for row in 0..n {
            let x = design[[row, 1]];
            let class = if x < -1.0 {
                0
            } else if x > 1.0 {
                1
            } else {
                2
            };
            y[[row, class]] = 1.0;
        }
        let penalty = Array2::<f64>::zeros((2, 2));
        let lambdas = Array1::<f64>::zeros(2);
        let out = fit_penalized_multinomial(MultinomialFitInputs {
            design: design.view(),
            y_one_hot: y.view(),
            penalty: penalty.view(),
            lambdas: lambdas.view(),
            row_weights: None,
            fisher_w_override: None,
            max_iter: 80,
            tol: 1.0e-12,
        })
        .expect("Firth/Jeffreys prior keeps the separated multinomial fit finite (#1854)");
        assert!(
            out.converged,
            "the Firth-penalized separation refit must report convergence"
        );
        // Every coefficient is finite — the whole point of the Firth prior on the
        // separated (unpenalized) logit directions.
        for &b in out.coefficients_active.iter() {
            assert!(
                b.is_finite(),
                "Firth-penalized coefficients must be finite, got {b}"
            );
        }
        // Fitted probabilities remain a valid simplex per row.
        for row in 0..n {
            let mut mass = 0.0_f64;
            for c in 0..3 {
                let p = out.fitted_probabilities[[row, c]];
                assert!(
                    p.is_finite() && (0.0..=1.0 + 1e-9).contains(&p),
                    "row {row} class {c} probability {p} out of [0,1]"
                );
                mass += p;
            }
            assert!(
                (mass - 1.0).abs() < 1e-6,
                "row {row} probabilities must sum to 1, got {mass}"
            );
        }
        // The finite fit still recovers the separated structure: on a clearly
        // interior representative of each region the predicted class is correct.
        let predict = |x: f64| -> usize {
            let mut eta = [0.0_f64; 3];
            for a in 0..2 {
                eta[a] = out.coefficients_active[[0, a]] + out.coefficients_active[[1, a]] * x;
            }
            let mut best = 0usize;
            for c in 1..3 {
                if eta[c] > eta[best] {
                    best = c;
                }
            }
            best
        };
        assert_eq!(predict(-2.5), 0, "deep-left region should predict class 0");
        assert_eq!(predict(2.5), 1, "deep-right region should predict class 1");
        assert_eq!(predict(0.0), 2, "central region should predict class 2");
    }

    #[test]
    fn formula_multinomial_accepts_finite_saturated_logits() {
        // A saturated-but-FINITE logit surface can be a valid formula REML mode
        // (the #715 penguins regime: bill/flipper cleanly separate the species,
        // so fitted logits can legitimately exceed ±25). `outer_converged ==
        // false` then signals only that the driver auto-escalated to never-fail
        // posterior sampling about that finite mode (gam#860), NOT a separation
        // artifact — the adapter must accept it, never raise
        // `MultinomialSeparationDetected`.
        let saturated_states = vec![
            ParameterBlockState {
                beta: Array1::from_vec(vec![1.0, 2.0]),
                eta: Array1::from_vec(vec![0.2, 4.0, -7.0]),
            },
            ParameterBlockState {
                beta: Array1::from_vec(vec![-1.0, 3.0]),
                eta: Array1::from_vec(vec![1.0, 25.5, -0.1]),
            },
        ];
        assert!(
            multinomial_formula_separation_diagnostic(17, 9, &saturated_states).is_none(),
            "a finite (even saturated, |eta|>25) formula optimum is a valid fit, \
             not a separation diagnostic"
        );

        // Only a genuinely NON-FINITE logit — a NaN/Inf blow-up in the inner
        // linear algebra with no finite mode to sample about — is a real
        // formula-path failure.
        let blown_up = vec![
            ParameterBlockState {
                beta: Array1::from_vec(vec![1.0, 2.0]),
                eta: Array1::from_vec(vec![0.2, 4.0, -7.0]),
            },
            ParameterBlockState {
                beta: Array1::from_vec(vec![-1.0, 3.0]),
                eta: Array1::from_vec(vec![1.0, f64::INFINITY, -0.1]),
            },
        ];
        let err = multinomial_formula_separation_diagnostic(17, 9, &blown_up)
            .expect("a non-finite formula logit must raise the separation diagnostic");
        assert!(
            matches!(
                err,
                EstimationError::MultinomialSeparationDetected {
                    iteration: 17,
                    max_abs_eta,
                    active_class_index: 1,
                    row_index: 1,
                } if !max_abs_eta.is_finite()
            ),
            "expected typed multinomial separation diagnostic at the non-finite channel, got {err:?}"
        );
    }

    #[test]
    fn separation_evidence_gate_arms_firth_only_on_blowup() {
        // Interior fit: finite logits well inside the saturation threshold ⇒ NO
        // separation evidence ⇒ the unbiased criterion's mode is accepted as-is
        // and the Firth/Jeffreys prior stays disarmed (#715 arm (a): no 1/K
        // shrinkage on well-identified data).
        let interior = vec![
            ParameterBlockState {
                beta: Array1::from_vec(vec![1.0, 2.0]),
                eta: Array1::from_vec(vec![0.2, 4.0, -7.0]),
            },
            ParameterBlockState {
                beta: Array1::from_vec(vec![-1.0, 3.0]),
                eta: Array1::from_vec(vec![1.0, -3.5, -0.1]),
            },
        ];
        assert!(
            multinomial_formula_separation_evidence(&interior).is_none(),
            "an interior finite mode must not arm the Firth refit"
        );

        // Saturated but finite logits are valid formula-path modes on
        // near-separated real data. They must not arm the Firth refit because
        // the Jeffreys pull can over-regularize the held-out probabilities.
        let saturated = vec![
            ParameterBlockState {
                beta: Array1::from_vec(vec![1.0, 2.0]),
                eta: Array1::from_vec(vec![0.2, 4.0, -7.0]),
            },
            ParameterBlockState {
                beta: Array1::from_vec(vec![-1.0, 3.0]),
                eta: Array1::from_vec(vec![1.0, 25.5, -0.1]),
            },
        ];
        assert!(
            multinomial_formula_separation_evidence(&saturated).is_none(),
            "a finite saturated formula-mode logit must not arm the Firth refit"
        );

        // Non-finite logit ⇒ inner blow-up along an unbounded direction ⇒
        // separation evidence.
        let blown_up = vec![ParameterBlockState {
            beta: Array1::from_vec(vec![1.0, 2.0]),
            eta: Array1::from_vec(vec![0.2, f64::NAN, -7.0]),
        }];
        let evidence = multinomial_formula_separation_evidence(&blown_up)
            .expect("a non-finite logit is separation evidence");
        assert!(
            evidence.contains("non-finite logit") && evidence.contains("row 1"),
            "evidence must name the non-finite logit, got {evidence}"
        );

        // Large finite logits below the fixed-lambda diagnostic threshold are
        // likewise accepted on the formula path.
        let near = vec![ParameterBlockState {
            beta: Array1::from_vec(vec![1.0, 2.0]),
            eta: Array1::from_vec(vec![0.2, 24.9, -24.9]),
        }];
        assert!(
            multinomial_formula_separation_evidence(&near).is_none(),
            "logits below the saturation threshold must not arm the Firth refit"
        );
    }

    #[test]
    fn unresolved_probe_evidence_arms_firth_on_saturated_finite_logits() {
        let saturated = vec![
            ParameterBlockState {
                beta: Array1::from_vec(vec![1.0, 2.0]),
                eta: Array1::from_vec(vec![0.2, 4.0, -7.0]),
            },
            ParameterBlockState {
                beta: Array1::from_vec(vec![-1.0, 3.0]),
                eta: Array1::from_vec(vec![1.0, 25.5, -0.1]),
            },
        ];

        assert!(
            multinomial_formula_separation_evidence(&saturated).is_none(),
            "a converged finite saturated formula optimum remains unbiased"
        );
        let evidence = multinomial_formula_unresolved_probe_separation_evidence(&saturated)
            .expect("a non-converged saturated probe should arm the Firth refit");
        assert!(
            evidence.contains("separation-scale finite logit")
                && evidence.contains("row 1")
                && evidence.contains("active class 1"),
            "unresolved-probe evidence should name the saturated channel, got {evidence}"
        );

        let near = vec![ParameterBlockState {
            beta: Array1::from_vec(vec![1.0, 2.0]),
            eta: Array1::from_vec(vec![0.2, 24.9, -24.9]),
        }];
        assert!(
            multinomial_formula_unresolved_probe_separation_evidence(&near).is_none(),
            "finite logits below the separation threshold still get the full unbiased retry"
        );
    }

    #[test]
    fn scaled_fisher_override_changes_first_step() {
        // Curvature scaled by 4× shrinks the first Newton step relative to the
        // analytic fit, so a single-iteration fit must differ.
        let (design, y, penalty, lambdas) = toy();
        let n = design.nrows();
        let m = y.ncols() - 1;
        // Analytic block at β = 0: p_a = 1/K = 1/3, so diag = p_a(1−p_a),
        // off-diag = −p_a p_b. Scale that exact block by 4.
        let pk = 1.0 / (y.ncols() as f64);
        let mut over = Array3::<f64>::zeros((n, m, m));
        for row in 0..n {
            for a in 0..m {
                for b in 0..m {
                    let analytic = if a == b { pk * (1.0 - pk) } else { -pk * pk };
                    over[[row, a, b]] = 4.0 * analytic;
                }
            }
        }
        let scaled = fit_penalized_multinomial(MultinomialFitInputs {
            design: design.view(),
            y_one_hot: y.view(),
            penalty: penalty.view(),
            lambdas: lambdas.view(),
            row_weights: None,
            fisher_w_override: Some(over.view()),
            max_iter: 1,
            tol: 1.0e-9,
        })
        .expect("override fit must succeed");
        let analytic = fit_penalized_multinomial(MultinomialFitInputs {
            design: design.view(),
            y_one_hot: y.view(),
            penalty: penalty.view(),
            lambdas: lambdas.view(),
            row_weights: None,
            fisher_w_override: None,
            max_iter: 1,
            tol: 1.0e-9,
        })
        .expect("analytic fit must succeed");
        let differs = scaled
            .coefficients_active
            .iter()
            .zip(analytic.coefficients_active.iter())
            .any(|(a, b)| (a - b).abs() > 1.0e-6);
        assert!(differs, "scaled curvature must change the first step");
    }
}

#[cfg(test)]
mod separation_firth_tests {
    //! Regression for #1854: on (quasi-)perfect separation the fixed-λ direct
    //! multinomial solve must engage the Firth/Jeffreys penalty and return a
    //! finite, converged, well-behaved fit instead of hard-erroring with
    //! `MultinomialSeparationDetected`.
    use super::*;

    /// A perfectly linearly separable 3-class problem with an UNPENALIZED design
    /// (`S = 0`), so no smoothing `λ` can bound the saturated logits — only the
    /// Firth prior `½ log det I(β)` keeps the estimate finite. The unbiased MLE
    /// here runs `|η| → ∞` (separation), which is exactly the #1854 trigger.
    fn separated_three_class() -> (Array2<f64>, Array2<f64>, Array2<f64>, Array1<f64>) {
        let n = 21;
        let p = 2; // intercept + ordering covariate x
        let k = 3;
        let mut design = Array2::<f64>::zeros((n, p));
        let mut y = Array2::<f64>::zeros((n, k));
        for i in 0..n {
            let x = -3.0 + 6.0 * (i as f64) / ((n - 1) as f64);
            design[[i, 0]] = 1.0;
            design[[i, 1]] = x;
            let cls = if x < -1.0 {
                0
            } else if x < 1.0 {
                1
            } else {
                2
            };
            y[[i, cls]] = 1.0;
        }
        // S = 0: no smoothing direction can bound the separated logits.
        let penalty = Array2::<f64>::zeros((p, p));
        let lambdas = Array1::<f64>::from_elem(k - 1, 1.0);
        (design, y, penalty, lambdas)
    }

    #[test]
    fn separation_engages_firth_finite_converged_fit() {
        let (design, y, penalty, lambdas) = separated_three_class();
        let out = fit_penalized_multinomial(MultinomialFitInputs {
            design: design.view(),
            y_one_hot: y.view(),
            penalty: penalty.view(),
            lambdas: lambdas.view(),
            row_weights: None,
            fisher_w_override: None,
            max_iter: 300,
            tol: 1e-10,
        })
        .expect("separated multinomial must engage Firth and return a fit, not error");

        assert!(out.converged, "Firth-engaged separation fit must converge");
        assert!(
            out.coefficients_active.iter().all(|v| v.is_finite()),
            "all coefficients must be finite under the Firth prior"
        );
        assert!(out.deviance.is_finite(), "deviance must be finite");

        // The runaway MLE would drive fitted probabilities to the {0,1} boundary;
        // the Firth prior keeps them strictly interior.
        for v in out.fitted_probabilities.iter() {
            assert!(
                *v > 0.0 && *v < 1.0,
                "Firth fit must stay interior, got p={v}"
            );
        }

        // Perfect separation ⇒ every training row classified to its true class.
        let n = design.nrows();
        let k = y.ncols();
        for i in 0..n {
            let mut best = 0usize;
            for c in 1..k {
                if out.fitted_probabilities[[i, c]] > out.fitted_probabilities[[i, best]] {
                    best = c;
                }
            }
            let truth = (0..k)
                .find(|&c| y[[i, c]] == 1.0)
                .expect("one-hot truth class");
            assert_eq!(best, truth, "row {i} misclassified under separation");
        }
    }

    #[test]
    fn separation_firth_returns_finite_wellshaped_covariance() {
        // Distinct angle: the Firth separation path must also expose a finite,
        // correctly-shaped (P·M × P·M) Laplace coefficient covariance — the
        // downstream SE machinery consumes it. A runaway MLE would have a
        // singular (non-invertible) information here.
        let (design, y, penalty, lambdas) = separated_three_class();
        let p = design.ncols();
        let k = y.ncols();
        let m = k - 1;
        let out = fit_penalized_multinomial(MultinomialFitInputs {
            design: design.view(),
            y_one_hot: y.view(),
            penalty: penalty.view(),
            lambdas: lambdas.view(),
            row_weights: None,
            fisher_w_override: None,
            max_iter: 300,
            tol: 1e-10,
        })
        .expect("separated multinomial must return a Firth fit");

        assert_eq!(
            out.coefficient_covariance.dim(),
            (p * m, p * m),
            "covariance must be P·M square"
        );
        assert!(
            out.coefficient_covariance.iter().all(|v| v.is_finite()),
            "Firth covariance entries must be finite"
        );
        // A genuine Laplace covariance is PSD ⇒ non-negative diagonal.
        for i in 0..(p * m) {
            assert!(
                out.coefficient_covariance[[i, i]] >= -1e-9,
                "covariance diagonal must be non-negative, got {}",
                out.coefficient_covariance[[i, i]]
            );
        }
    }

    #[test]
    fn firth_solver_does_not_over_report_convergence_when_truncated() {
        // #2066 (convergence honesty): the Firth Newton loop must report
        // `converged` according to a genuine stationarity criterion — never as a
        // side effect of simply stopping. Before the fix, a line-search stall set
        // `converged = true` unconditionally; more broadly, the flag must be
        // false whenever the solve is stopped short of stationarity.
        //
        // Angle: run the SAME separated problem that converges under a full
        // budget (`separation_engages_firth_finite_converged_fit`) but starve the
        // iteration budget so it provably cannot reach the interior Firth mode.
        // The honest report is `converged = false`; the coefficients must still be
        // finite (no NaN leak from the truncated iterate).
        let (design, y, penalty, lambdas) = separated_three_class();

        let truncated = fit_penalized_multinomial_firth_fallback(
            design.view(),
            y.view(),
            penalty.view(),
            lambdas.view(),
            None,
            1, // one Newton iteration — far from the separated mode
            1e-12,
        )
        .expect("Firth fallback must return a (non-converged) fit, not error");
        assert!(
            !truncated.converged,
            "a Firth solve stopped after one iteration on a separated problem \
             must report converged=false, not paper over non-convergence"
        );
        assert!(
            truncated.coefficients_active.iter().all(|v| v.is_finite()),
            "truncated Firth iterate must remain finite"
        );

        // Contrast: with a full budget the same problem does reach stationarity
        // and is honestly reported as converged — so the flag tracks the solve,
        // not the exit.
        let full = fit_penalized_multinomial_firth_fallback(
            design.view(),
            y.view(),
            penalty.view(),
            lambdas.view(),
            None,
            300,
            1e-10,
        )
        .expect("Firth fallback must converge under a full budget");
        assert!(
            full.converged,
            "full-budget Firth solve on the separated problem must converge"
        );
    }
}

#[cfg(test)]
mod reference_class_invariance_tests {
    //! Regression for #1587: a penalized multinomial-logit GAM fit must be
    //! invariant to which class is the (arbitrary) softmax reference/baseline.
    //!
    //! The production REML path (`fit_penalized_multinomial_formula`) reference-
    //! codes the `K` classes (the last sorted label is the baseline) and, with
    //! the legacy `Diagonal` penalty metric, penalizes only the `K−1`
    //! reference-anchored ALR contrasts `½ Σ_a λ_a β_aᵀ S β_a`. Relabeling the
    //! response so a *different* class sorts last penalizes a different frame of
    //! log-odds contrasts, so the predicted probabilities drift (~1e-2 absolute)
    //! even though they are mathematically independent of the reference choice.
    //!
    //! This test fits the SAME 3-class softmax sample under three cyclic
    //! relabelings — each making a different original class the baseline —
    //! realigns the predicted probability columns back to the original class
    //! identities, and asserts the cross-labeling drift is below `1e-3`
    //! (the defect is ~1e-2; refitting the same labeling twice agrees to
    //! ~1e-12). It is the Rust-level sibling of
    //! `tests/bug_hunt_multinomial_fit_depends_on_reference_class_test.py`.

    use super::*;
    use gam_data::load_dataset_projected;
    use std::fmt::Write as _;
    use std::fs;
    use tempfile::tempdir;

    /// Deterministic `splitmix64` → `[0,1)` uniform stream (no external RNG dep;
    /// the only requirement is a well-distributed, reproducible draw).
    struct SplitMix64(u64);
    impl SplitMix64 {
        fn next_u64(&mut self) -> u64 {
            self.0 = self.0.wrapping_add(0x9E37_79B9_7F4A_7C15);
            let mut z = self.0;
            z = (z ^ (z >> 30)).wrapping_mul(0xBF58_476D_1CE4_E5B9);
            z = (z ^ (z >> 27)).wrapping_mul(0x94D0_49BB_1331_11EB);
            z ^ (z >> 31)
        }
        fn unit(&mut self) -> f64 {
            // 53-bit mantissa uniform in [0, 1).
            (self.next_u64() >> 11) as f64 / (1u64 << 53) as f64
        }
    }

    /// Draw a clean 3-class softmax regression sample (the issue's generator).
    /// Returns `(x, class)` with integer classes `0/1/2`.
    fn sample_classes(seed: u64, n: usize) -> (Vec<f64>, Vec<usize>) {
        let mut rng = SplitMix64(seed.wrapping_add(0x1234_5678));
        let mut x = Vec::with_capacity(n);
        let mut cls = Vec::with_capacity(n);
        for _ in 0..n {
            let xi = -2.0 + 4.0 * rng.unit();
            let eta = [0.5 + 0.8 * xi, -0.3 - 0.5 * xi, 0.0];
            let mut p = [eta[0].exp(), eta[1].exp(), eta[2].exp()];
            let s: f64 = p.iter().sum();
            for v in &mut p {
                *v /= s;
            }
            // Inverse-CDF draw into one of the 3 classes.
            let u = rng.unit();
            let c = if u < p[0] {
                0
            } else if u < p[0] + p[1] {
                1
            } else {
                2
            };
            x.push(xi);
            cls.push(c);
        }
        (x, cls)
    }

    /// Build an `EncodedDataset` with columns `x` (numeric) and `y`
    /// (categorical, from the given string labels) by round-tripping a CSV.
    fn dataset_xy(dir: &std::path::Path, tag: &str, x: &[f64], y: &[String]) -> gam_data::EncodedDataset {
        let path = dir.join(format!("data_{tag}.csv"));
        let mut csv = String::from("x,y\n");
        for (xi, yi) in x.iter().zip(y.iter()) {
            writeln!(csv, "{xi},{yi}").unwrap();
        }
        fs::write(&path, csv).expect("write training csv");
        load_dataset_projected(&path, &["x".to_string(), "y".to_string()])
            .expect("load training dataset")
    }

    /// Fit `y ~ s(x)` under the relabeling `name_map` (original class `c` gets
    /// label `name_map[c]`), predict on `grid`, and return the predicted
    /// probabilities **realigned to the original class order** 0/1/2, shape
    /// `(grid.len(), 3)`.
    fn fit_predict_aligned(
        dir: &std::path::Path,
        tag: &str,
        x: &[f64],
        cls: &[usize],
        name_map: [&str; 3],
        grid: &[f64],
    ) -> Array2<f64> {
        let labels: Vec<String> = cls.iter().map(|&c| name_map[c].to_string()).collect();
        let train = dataset_xy(dir, tag, x, &labels);
        let config = FitConfig::default();
        let model = fit_penalized_multinomial_formula(&train, "y ~ s(x)", &config, 1.0, 60, 1e-6)
            .expect("multinomial formula fit must succeed");

        // Predict on the grid. The categorical `y` column is not needed for
        // prediction, but the schema is simplest if we supply a dummy.
        let grid_y: Vec<String> = grid.iter().map(|_| name_map[0].to_string()).collect();
        let grid_ds = dataset_xy(dir, &format!("{tag}_grid"), grid, &grid_y);
        let probs = predict_multinomial_formula(&model, &grid_ds)
            .expect("multinomial predict must succeed");

        // `model.class_levels` is the sorted label order; the column for original
        // class `c` is at the rank of `name_map[c]` among the sorted labels.
        let mut sorted: Vec<&str> = name_map.to_vec();
        sorted.sort_unstable();
        let col_of_orig: Vec<usize> = (0..3)
            .map(|c| sorted.iter().position(|l| *l == name_map[c]).unwrap())
            .collect();
        // Sanity: the model's class_levels must match the sorted labels.
        assert_eq!(
            model.class_levels,
            sorted.iter().map(|s| s.to_string()).collect::<Vec<_>>(),
            "class_levels must be the sorted label order"
        );
        let n = grid.len();
        let mut aligned = Array2::<f64>::zeros((n, 3));
        for r in 0..n {
            for c in 0..3 {
                aligned[[r, c]] = probs[[r, col_of_orig[c]]];
            }
        }
        aligned
    }

    fn max_abs_diff(a: &Array2<f64>, b: &Array2<f64>) -> f64 {
        a.iter()
            .zip(b.iter())
            .map(|(p, q)| (p - q).abs())
            .fold(0.0_f64, f64::max)
    }

    // gam#1587: now that the reference-symmetric centered `M⊗S_t` joint penalty
    // is wired through the custom-family outer REML loop (per-eval
    // `JointPenaltyBundle` + outer penalty_coords/logdet/operator), the
    // production multinomial fit is invariant to the arbitrary reference class,
    // so this guard runs by default (the opt-in skip attribute it carried while
    // the fix was pending is also forbidden by the build.rs ban-scanner). It is
    // an end-to-end fit guard (a handful of full softmax `y ~ s(x)` fits) —
    // slower than a unit test but a true production-path regression.
    #[test]
    fn multinomial_fit_is_invariant_to_reference_class_1587() {
        let td = tempdir().expect("tempdir");
        let dir = td.path();
        // The reference-class drift is STRUCTURAL (it does not shrink with n, see
        // the issue table), so a modest n exposes it just as cleanly as n=900
        // while keeping this an affordable CI guard.
        let (x, cls) = sample_classes(0, 300);
        let grid: Vec<f64> = (0..7).map(|i| -1.5 + 3.0 * (i as f64) / 6.0).collect();

        // Three labelings that each make a DIFFERENT original class the baseline
        // (the class whose label sorts LAST is the reference K−1):
        //   ["A","B","C"] → ref = class 2
        //   ["B","C","A"] → ref = class 1
        //   ["C","A","B"] → ref = class 0
        let a = fit_predict_aligned(dir, "abc", &x, &cls, ["A", "B", "C"], &grid);
        let b = fit_predict_aligned(dir, "bca", &x, &cls, ["B", "C", "A"], &grid);
        let c = fit_predict_aligned(dir, "cab", &x, &cls, ["C", "A", "B"], &grid);

        // Refitting the SAME labeling twice must agree to ~machine precision —
        // this isolates optimizer noise from the structural reference drift.
        let a2 = fit_predict_aligned(dir, "abc2", &x, &cls, ["A", "B", "C"], &grid);
        let refit_noise = max_abs_diff(&a, &a2);
        assert!(
            refit_noise < 1e-6,
            "refitting the same labeling must be deterministic (got {refit_noise:.3e})"
        );

        let drift = max_abs_diff(&a, &b)
            .max(max_abs_diff(&a, &c))
            .max(max_abs_diff(&b, &c));
        assert!(
            drift < 1e-3,
            "predicted probabilities must be invariant to the reference class; \
             cross-labeling drift = {drift:.3e} (refit noise = {refit_noise:.3e})"
        );
    }
}