gam_terms/

latent.rs

//! `LatentCoord` — per-row latent coordinates as a first-class gamfit parameter.
//!
//! The Riemannian update path follows manifold GPLVM practice (mGPLVM;
//! Jensen/Kao/Tran/Stevenson 2020 and related head-direction / population
//! manifold work): angular, spherical, and product-topology latents are
//! updated on their natural manifold instead of as Euclidean coordinates
//! with basis-side periodic hacks. Retractions and Euclidean-to-Riemannian
//! Hessian conversion follow Absil/Mahony/Sepulchre (2008) and the Manopt /
//! Pymanopt implementation pattern. In the audit-revised gauge framing, the
//! Riemannian update is itself a gauge restriction: Circle/Sphere/Torus
//! structure identifies the latent up to the corresponding global isometry
//! (for example one rotation per cycle), not up to the full diffeomorphism
//! group of an unconstrained Euclidean latent chart.
//!
//! ## Summary
//!
//! `LatentCoordValues` is the structural sibling of [`SpatialLogKappaCoords`]
//! (see [`crate::smooth`]). Both store a flat `Array1<f64>` that the
//! REML/IFT outer loop treats as *design-moving, non-penalty-like*
//! hyper-coordinates. `SpatialLogKappaCoords` holds one or more kernel-shape
//! coordinates per spatial term. `LatentCoordValues`
//! holds an `N × d` matrix of per-row latent coordinates `t_n ∈ ℝ^d`.
//!
//! For a Duchon (or any radial) basis:
//!
//! ```text
//! Φ_{n,k} = φ(‖t_n − c_k‖),
//! ∂Φ_{n,k}/∂t_n = φ'(r_{nk}) · (t_n − c_k) / r_{nk}.
//! ```
//!
//! The radial-gradient `φ'(r)` is the same scalar the kernel-shape machinery already
//! computes via [`crate::basis::duchon_radial_jets`]; the chain rule
//! `(t_n − c_k)/r_{nk}` is what differs between "differentiate against the
//! kernel scale" and "differentiate against the first kernel argument t".
//! Everything downstream of `HyperDesignDerivative::from_implicit` (matrix-free
//! Newton, IFT cache, persistent warm-start, REML/LAML evaluation) is reused
//! verbatim.
//!
//! ## Gauge fixing
//!
//! The bare data-fit `½‖y − Φ(t)β‖²` is invariant under any diffeomorphism
//! `t ↦ φ(t)` (absorb into a re-fit β), so the inner Hessian in the latent
//! block is singular and IFT breaks. [`LatentIdMode`] enumerates the
//! gauge-fix penalties exposed at the configuration layer:
//!
//! * [`LatentIdMode::AuxPrior`] — iVAE-style auxiliary-conditional prior
//!   `R_id(t,u) = ½ μ ‖t − ĥ(u)‖²` where `ĥ` is a small ridge / linear map
//!   fit internally against the auxiliary `u`. `μ` is REML-selectable like a
//!   smoothing parameter only when the marginal likelihood includes the
//!   log-`μ` normalizer, `ĥ` is at least C¹, and the conditional precision is
//!   positive-definite on the anchored subspace. Under those regularity
//!   conditions this is the principled identifiability fix (Khemakhem et al.
//!   2020).
//! * [`LatentIdMode::DimSelection`] — ARD on each latent axis. One ridge
//!   penalty per axis; REML drives unused axes' precision to infinity only
//!   after `AuxPrior` or a future isometry prior fixes the gauge.
//! * [`LatentIdMode::None`] — no gauge fix. Useful only as an explicit
//!   opt-out; the caller is responsible for separately providing a unique
//!   inner minimum (e.g. via a custom penalty).
//!
//! [`LatentIdMode::IsometryToReference`] (proposal §4(b)) anchors the latent to
//! a caller-supplied reference configuration via `½ μ ‖t − reference‖²` with a
//! REML-selectable `μ`, fixing the gauge without an auxiliary signal `u`.

use crate::basis::{BasisError, RadialScalarKind};
use gam_problem::LatentRetractionRegistry;
use ndarray::{Array1, Array2, Array3, ArrayView1, ArrayView2, ArrayView3};
use std::sync::atomic::{AtomicU64, Ordering};
const SPHERE_NORMAL_PIN: f64 = 1.0;
static NEXT_LATENT_COORD_ID: AtomicU64 = AtomicU64::new(1);

fn next_latent_coord_id() -> u64 {
    NEXT_LATENT_COORD_ID.fetch_add(1, Ordering::Relaxed)
}

/// Choice of auxiliary-prior conditional mean estimator `ĥ(u)`.
///
/// `Ridge` is the cheap default that closes form (one `K_u × K_u` solve);
/// `Linear` is equivalent to `Ridge` with zero ridge and is intended for
/// auxiliaries `u` that are already low-dimensional and well-conditioned.
#[derive(Debug, Clone, Copy)]
pub enum AuxPriorFamily {
    /// Ridge regression `t ≈ U · A` with a small diagonal regularizer.
    /// The default ridge strength is `1e-6 · trace(UᵀU)/p`, which is
    /// numerically benign and never under-constrains the fit when
    /// `n_obs > p`.
    Ridge,
    /// Plain linear projection (no ridge). Errors out at construction if
    /// `UᵀU` is singular.
    Linear,
}

/// Strength of the auxiliary-prior identifiability penalty.
///
/// `Auto` defers the choice to REML — the strength is added to the outer
/// vector as one extra `ρ`-axis (one log-precision per `LatentCoord`). When
/// the caller supplies an explicit `Fixed(μ)` the strength is held constant
/// throughout the fit; useful for warm-starts and reproducibility. The REML
/// path is valid only with the prior normalizer included, a C¹ conditional
/// mean map, and positive-definite precision on the anchored subspace.
#[derive(Debug, Clone, Copy)]
pub enum AuxPriorStrength {
    Auto,
    Fixed(f64),
}

/// Identifiability / gauge-fix mode for a [`LatentCoordValues`] block.
///
/// `AuxPrior` is currently the only standalone gauge-fixing mode; see the
/// module docstring. `DimSelection` must be paired with `AuxPrior` (or a
/// future isometry mode) by higher-level assembly before fitting.
#[derive(Debug, Clone)]
pub enum LatentIdMode {
    /// Conditional Gaussian prior `p(t | u)` with mean `ĥ(u)` fit by
    /// `family`. The penalty contribution is
    /// `R_id = ½ μ · ‖t − ĥ(u)‖²`. `u` has shape `(n_obs, p)`. If
    /// `strength == Auto`, REML selection of `μ` requires the log-`μ`
    /// normalizer, C¹ regularity of `ĥ`, and positive-definiteness on the
    /// subspace anchored by `u`.
    AuxPrior {
        u: Array2<f64>,
        family: AuxPriorFamily,
        strength: AuxPriorStrength,
    },
    /// Auxiliary prior plus ARD over latent axes. `AuxPrior` supplies the
    /// identifiability anchor; `init_log_precision` seeds the per-axis ARD
    /// coordinates.
    AuxPriorDimSelection {
        u: Array2<f64>,
        family: AuxPriorFamily,
        strength: AuxPriorStrength,
        init_log_precision: Option<Array1<f64>>,
    },
    /// ARD over latent axes. One ridge penalty per latent axis; the per-axis
    /// log-precision joins the outer ρ vector. `init_log_precision` seeds
    /// the per-axis ρ — a vector of length `d`. `None` defaults to a flat
    /// zero seed (precision = 1 on every axis).
    DimSelection {
        init_log_precision: Option<Array1<f64>>,
    },
    /// Behaviorally-anchored head (issue #912). The auxiliary signal is
    /// promoted from a fixed-covariate *prior* to a modeled *outcome*: a GLM
    /// behavioral head `g(E[y|t]) = a + t·w` whose design columns are the
    /// latent codes contributes a *likelihood* term to the joint objective,
    /// so REML balances reconstruction vs. behavioral fit with no trade-off
    /// scalar (magic by default).
    ///
    /// The head's coefficients are direct hyperparameters appended to θ (one
    /// `(1 + d)` block per η-channel), like the AuxPrior log-`μ`. Because a
    /// single binary label pins ~1 gauge dimension, `AuxOutcome` *composes*
    /// with `DimSelection` ARD (the `init_log_precision` seed) and the
    /// isometry pin rather than replacing them; the validator requires that
    /// composition and rejects a head with no labels.
    AuxOutcome {
        head: crate::decoders::behavioral_head::BehavioralHead,
        /// ARD seed composed with the head, one log-precision per latent axis
        /// (length `d`). `AuxOutcome` always carries the ARD axis-selection
        /// alongside the behavioral anchor, since the label alone under-pins
        /// the gauge. `None` defaults to a flat zero seed.
        init_log_precision: Option<Array1<f64>>,
    },
    /// Anchor the latent configuration to a caller-supplied reference up to
    /// the global isometry the chosen manifold already quotients out: penalty
    /// `R_id = ½ μ · ‖t − reference‖²` with REML-selectable `μ` (the log-`μ`
    /// normalizer enters the marginal likelihood exactly as in `AuxPrior`).
    /// Unlike `AuxPrior`, the target is a fixed reference configuration (e.g. a
    /// pilot embedding that fixes the isometry representative) rather than an
    /// auxiliary-conditional mean `ĥ(u)`, so it pins the gauge with no
    /// auxiliary signal `u`. `reference` has shape `(n_obs, d)`. As a standalone
    /// anchor it is a valid gauge fix; it also composes with `DimSelection` ARD.
    IsometryToReference {
        reference: Array2<f64>,
        strength: AuxPriorStrength,
    },
    /// No gauge fix. Inner Hessian is rank-deficient; results are not
    /// uniquely defined. Intended only for the explicit "I supply my own
    /// gauge constraint via the smoothing penalty" pathway.
    None,
}

/// Natural manifold for per-row latent-coordinate updates.
///
/// `Euclidean` preserves the original additive update. `Circle` is a scalar
/// angular coordinate wrapped modulo `2π`. `Sphere { dim }` is the embedded
/// unit sphere in `R^dim`, with retraction `(t + ξ) / ||t + ξ||`. `Product`
/// composes these blockwise; inside a product, `Euclidean` denotes one
/// unconstrained scalar axis.
#[derive(Debug, Clone, PartialEq, Default)]
pub enum LatentManifold {
    /// Unconstrained `R^d` — the current default.
    #[default]
    Euclidean,
    /// Scalar periodic coordinate on `S^1` with caller-supplied period.
    ///
    /// Wraps modulo `period`; pass `period = 2π` for radian conventions and
    /// `period = 1.0` for basis evaluators that interpret the latent as a
    /// fraction of one period. The metric weight uses `1/period²` so the
    /// trust-region radius respects the chosen unit.
    Circle { period: f64 },
    /// Embedded unit sphere `S^(dim-1)`.
    Sphere { dim: usize },
    /// Closed interval in `R`; the retraction clamps to the boundary.
    Interval { lo: f64, hi: f64 },
    /// Product manifold, split block-by-block in row-major ambient storage.
    Product(Vec<LatentManifold>),
    /// Product manifold with explicit per-axis trust-region metric weights.
    ///
    /// Without per-axis weighting, a Product of Circle + Interval treats
    /// 1 radian as commensurate with the entire bounded range. With weights
    /// = 1/scale², the trust-region radius respects each axis's natural unit.
    ProductWithMetric {
        manifolds: Vec<LatentManifold>,
        weights: Vec<f64>,
    },
}

impl LatentManifold {
    pub fn is_euclidean(&self) -> bool {
        matches!(self, Self::Euclidean)
    }

    /// Whether the Euclidean→Riemannian geometry transform applied by
    /// `crate::solver::arrow_schur::ArrowSchurSystem::apply_riemannian_latent_geometry`
    /// is the **identity** on the per-row gradient, `H_tt`, and `H_tβ` blocks
    /// for *every* coordinate `t` on this chart.
    ///
    /// This is the exact condition under which a coupled Gauss-Newton block
    /// `μ AᵀA = [[htt, cross],[crossᵀ, hbb]]` assembled from one residual
    /// Jacobian survives the geometry pass with its PSD coherence intact: if
    /// the transform leaves `htt` and the `htbeta` cross-block untouched, the
    /// whole block is still `μ AᵀA` (PSD) and its Schur complement is PSD, so
    /// the isometry cross-coupling can be kept (faster, exact Newton).
    ///
    /// A chart that rewrites `htt` with a curvature/connection term or
    /// column-projects the cross-block (`Sphere`, an active `Interval`
    /// boundary, any curved `Product` factor) breaks that pairing — the
    /// cross-block is then no longer matched to diagonals from the same
    /// Jacobian and the Schur complement can go indefinite (the #681
    /// circle/sphere failure mode). Such charts must drop the cross-block.
    ///
    /// Flat charts (`Euclidean`, `Circle`, and `Product`s built only from
    /// these) transform as the identity unconditionally — their tangent
    /// projection is the identity, they carry no connection term, and they add
    /// no normal pinning — so coherence is preserved and the cross-block is
    /// kept. `Interval` is excluded: its tangent projection masks coordinates
    /// at an active boundary (a `t`-dependent projection), which breaks the
    /// pairing exactly like a curved chart.
    pub fn preserves_isometry_cross_block_coherence(&self) -> bool {
        match self {
            Self::Euclidean | Self::Circle { .. } => true,
            Self::Sphere { .. } | Self::Interval { .. } => false,
            Self::Product(parts)
            | Self::ProductWithMetric {
                manifolds: parts, ..
            } => parts
                .iter()
                .all(|part| part.preserves_isometry_cross_block_coherence()),
        }
    }

    pub fn ambient_dim(&self, fallback_dim: usize) -> usize {
        match self {
            Self::Euclidean => fallback_dim,
            Self::Circle { .. } | Self::Interval { .. } => 1,
            Self::Sphere { dim } => *dim,
            Self::Product(parts)
            | Self::ProductWithMetric {
                manifolds: parts, ..
            } => parts.iter().map(|part| part.ambient_dim(1)).sum(),
        }
    }

    /// Per-axis weights for the Riemannian trust-region metric.
    ///
    /// Defaults use `1/scale²`: Circle scale is `2π`, Sphere scale is `π`,
    /// Interval scale is `hi - lo`, and Euclidean scale is `1`. Product
    /// manifolds recurse and concatenate; [`Self::ProductWithMetric`] uses
    /// the caller-supplied weights directly.
    pub fn metric_weights(&self) -> Vec<f64> {
        match self {
            Self::Euclidean => vec![1.0],
            Self::Circle { period } => {
                assert!(
                    period.is_finite() && *period > 0.0,
                    "LatentManifold::Circle requires a finite positive period; got {period}"
                );
                vec![1.0 / (period * period)]
            }
            Self::Sphere { dim } => {
                let w = 1.0 / (std::f64::consts::PI * std::f64::consts::PI);
                vec![w; *dim]
            }
            Self::Interval { lo, hi } => {
                let scale = hi - lo;
                assert!(
                    scale.is_finite() && scale > 0.0,
                    "LatentManifold::Interval requires finite lo < hi; got lo={lo}, hi={hi}"
                );
                vec![1.0 / (scale * scale)]
            }
            Self::Product(parts) => {
                let mut out = Vec::with_capacity(self.ambient_dim(1));
                for part in parts {
                    out.extend(part.metric_weights());
                }
                out
            }
            Self::ProductWithMetric { manifolds, weights } => {
                let expected: usize = manifolds.iter().map(|part| part.ambient_dim(1)).sum();
                assert_eq!(
                    weights.len(),
                    expected,
                    "LatentManifold::ProductWithMetric weights length must match ambient dimension"
                );
                weights.clone()
            }
        }
    }

    /// Per-ambient-axis periodicity: `Some(period)` for an axis that wraps
    /// modulo a finite period (a `Circle` factor, including the longitude of
    /// the lat/lon sphere chart), `None` for a non-periodic axis (Euclidean,
    /// Interval, or an embedded `Sphere` axis whose retraction is smooth and
    /// has no cut).
    ///
    /// Used by the SAE-manifold ARD prior to switch from the cut-discontinuous
    /// Euclidean `½α t²` to a smooth von-Mises energy on periodic axes. The
    /// embedded `Sphere` is deliberately reported as non-periodic: its
    /// retraction `(t+ξ)/‖t+ξ‖` is globally smooth, so the ambient `½α‖t‖²`
    /// prior has no discontinuity there.
    pub fn axis_periods(&self) -> Vec<Option<f64>> {
        match self {
            Self::Euclidean => vec![None],
            Self::Circle { period } => {
                assert!(
                    period.is_finite() && *period > 0.0,
                    "LatentManifold::Circle requires a finite positive period; got {period}"
                );
                vec![Some(*period)]
            }
            Self::Sphere { dim } => vec![None; *dim],
            Self::Interval { .. } => vec![None],
            Self::Product(parts) => {
                let mut out = Vec::with_capacity(self.ambient_dim(1));
                for part in parts {
                    out.extend(part.axis_periods());
                }
                out
            }
            Self::ProductWithMetric { manifolds, .. } => {
                let mut out = Vec::with_capacity(self.ambient_dim(1));
                for part in manifolds {
                    out.extend(part.axis_periods());
                }
                out
            }
        }
    }

    /// Project an arbitrary ambient point back to the manifold.
    pub fn project_point(&self, t: ArrayView1<'_, f64>) -> Array1<f64> {
        match self {
            Self::Euclidean => t.to_owned(),
            Self::Circle { period } => {
                let mut out = Array1::<f64>::zeros(1);
                out[0] = wrap_to_period(t[0], *period);
                out
            }
            Self::Sphere { dim } => {
                assert_eq!(t.len(), *dim);
                normalize_or_axis(t, *dim)
            }
            Self::Interval { lo, hi } => {
                // Order the bounds defensively: `f64::clamp` panics if min > max,
                // so a reversed `Interval { lo, hi }` would otherwise crash deep
                // in projection rather than clamp into the intended range.
                let (lo, hi) = if lo <= hi { (*lo, *hi) } else { (*hi, *lo) };
                let mut out = Array1::<f64>::zeros(1);
                out[0] = t[0].clamp(lo, hi);
                out
            }
            Self::Product(parts)
            | Self::ProductWithMetric {
                manifolds: parts, ..
            } => {
                let mut out = Array1::<f64>::zeros(t.len());
                let mut offset = 0_usize;
                for part in parts {
                    let dim = part.ambient_dim(1);
                    let projected = part.project_point(t.slice(ndarray::s![offset..offset + dim]));
                    for a in 0..dim {
                        out[offset + a] = projected[a];
                    }
                    offset += dim;
                }
                assert_eq!(offset, t.len());
                out
            }
        }
    }

    /// Retraction `R_t(ξ)`, using closed-form analytic maps for every variant.
    pub fn retract(&self, t: ArrayView1<'_, f64>, xi: ArrayView1<'_, f64>) -> Array1<f64> {
        assert_eq!(t.len(), xi.len());
        match self {
            Self::Euclidean => {
                let mut out = t.to_owned();
                for a in 0..out.len() {
                    out[a] += xi[a];
                }
                out
            }
            Self::Circle { period } => {
                let mut out = Array1::<f64>::zeros(1);
                out[0] = wrap_to_period(t[0] + xi[0], *period);
                out
            }
            Self::Sphere { dim } => {
                assert_eq!(t.len(), *dim);
                let mut y = Array1::<f64>::zeros(*dim);
                for a in 0..*dim {
                    y[a] = t[a] + xi[a];
                }
                normalize_or_axis(y.view(), *dim)
            }
            Self::Interval { lo, hi } => {
                // Order the bounds defensively: `f64::clamp` panics if min > max,
                // so a reversed `Interval { lo, hi }` would otherwise crash the
                // retraction instead of clamping into the intended range.
                let (lo, hi) = if lo <= hi { (*lo, *hi) } else { (*hi, *lo) };
                let mut out = Array1::<f64>::zeros(1);
                out[0] = (t[0] + xi[0]).clamp(lo, hi);
                out
            }
            Self::Product(parts)
            | Self::ProductWithMetric {
                manifolds: parts, ..
            } => {
                let mut out = Array1::<f64>::zeros(t.len());
                let mut offset = 0_usize;
                for part in parts {
                    let dim = part.ambient_dim(1);
                    let next = part.retract(
                        t.slice(ndarray::s![offset..offset + dim]),
                        xi.slice(ndarray::s![offset..offset + dim]),
                    );
                    for a in 0..dim {
                        out[offset + a] = next[a];
                    }
                    offset += dim;
                }
                assert_eq!(offset, t.len());
                out
            }
        }
    }

    /// Orthogonal projection of an ambient vector onto `T_t M`.
    pub fn project_to_tangent(
        &self,
        t: ArrayView1<'_, f64>,
        v: ArrayView1<'_, f64>,
    ) -> Array1<f64> {
        assert_eq!(t.len(), v.len());
        match self {
            Self::Euclidean | Self::Circle { .. } => v.to_owned(),
            Self::Sphere { dim } => {
                assert_eq!(t.len(), *dim);
                let tv = t.dot(&v);
                let mut out = v.to_owned();
                for a in 0..*dim {
                    out[a] -= tv * t[a];
                }
                out
            }
            Self::Interval { lo, hi } => {
                let mut out = Array1::<f64>::zeros(1);
                let at_lo = t[0] <= *lo && v[0] < 0.0;
                let at_hi = t[0] >= *hi && v[0] > 0.0;
                out[0] = if at_lo || at_hi { 0.0 } else { v[0] };
                out
            }
            Self::Product(parts)
            | Self::ProductWithMetric {
                manifolds: parts, ..
            } => {
                let mut out = Array1::<f64>::zeros(v.len());
                let mut offset = 0_usize;
                for part in parts {
                    let dim = part.ambient_dim(1);
                    let projected = part.project_to_tangent(
                        t.slice(ndarray::s![offset..offset + dim]),
                        v.slice(ndarray::s![offset..offset + dim]),
                    );
                    for a in 0..dim {
                        out[offset + a] = projected[a];
                    }
                    offset += dim;
                }
                assert_eq!(offset, v.len());
                out
            }
        }
    }

    /// Project an objective gradient onto the linearized feasible update space.
    ///
    /// For smooth manifolds this is the usual tangent projection. For interval
    /// endpoints the sign test is applied to the descent direction `-g`: at the
    /// upper endpoint, a negative gradient would step outward, so the coordinate
    /// is held fixed; at the lower endpoint, a positive gradient would step
    /// outward. This is distinct from [`Self::project_to_tangent`], whose
    /// interval branch projects update velocities.
    pub fn project_gradient_to_tangent(
        &self,
        t: ArrayView1<'_, f64>,
        g: ArrayView1<'_, f64>,
    ) -> Array1<f64> {
        assert_eq!(t.len(), g.len());
        match self {
            Self::Euclidean | Self::Circle { .. } | Self::Sphere { .. } => {
                self.project_to_tangent(t, g)
            }
            Self::Interval { lo, hi } => {
                let mut out = Array1::<f64>::zeros(1);
                let descent_exits_lo = t[0] <= *lo && g[0] > 0.0;
                let descent_exits_hi = t[0] >= *hi && g[0] < 0.0;
                out[0] = if descent_exits_lo || descent_exits_hi {
                    0.0
                } else {
                    g[0]
                };
                out
            }
            Self::Product(parts)
            | Self::ProductWithMetric {
                manifolds: parts, ..
            } => {
                let mut out = Array1::<f64>::zeros(g.len());
                let mut offset = 0_usize;
                for part in parts {
                    let dim = part.ambient_dim(1);
                    let projected = part.project_gradient_to_tangent(
                        t.slice(ndarray::s![offset..offset + dim]),
                        g.slice(ndarray::s![offset..offset + dim]),
                    );
                    for a in 0..dim {
                        out[offset + a] = projected[a];
                    }
                    offset += dim;
                }
                // The per-part ambient widths (`part.ambient_dim(1)`) must tile
                // `g` exactly. This holds because every `Product` is built in
                // expanded scalar-factor form — a multi-dimensional Euclidean
                // atom is stored as `d` single-axis `Euclidean` children, never
                // one `d`-wide `Euclidean` (which `ambient_dim(1)` would
                // under-count as one axis, mis-tiling a mixed-dimension composite
                // and firing here — the #2295 zoo-fit panic). See
                // `SaeManifoldTerm::append_coordinate_manifold_parts`.
                assert_eq!(
                    offset,
                    g.len(),
                    "Product factor ambient widths ({offset}) must tile the gradient ({}); a \
                     Product must be in expanded scalar-factor form (see #2295)",
                    g.len()
                );
                out
            }
        }
    }

    /// Project a coordinate-space Jacobian/cross-block column with the same
    /// active interval coordinates selected by
    /// [`Self::project_gradient_to_tangent`].
    pub fn project_vector_to_gradient_tangent(
        &self,
        t: ArrayView1<'_, f64>,
        g: ArrayView1<'_, f64>,
        v: ArrayView1<'_, f64>,
    ) -> Array1<f64> {
        assert_eq!(t.len(), g.len());
        assert_eq!(t.len(), v.len());
        match self {
            Self::Euclidean | Self::Circle { .. } | Self::Sphere { .. } => {
                self.project_to_tangent(t, v)
            }
            Self::Interval { lo, hi } => {
                let mut out = Array1::<f64>::zeros(1);
                let descent_exits_lo = t[0] <= *lo && g[0] > 0.0;
                let descent_exits_hi = t[0] >= *hi && g[0] < 0.0;
                out[0] = if descent_exits_lo || descent_exits_hi {
                    0.0
                } else {
                    v[0]
                };
                out
            }
            Self::Product(parts)
            | Self::ProductWithMetric {
                manifolds: parts, ..
            } => {
                let mut out = Array1::<f64>::zeros(v.len());
                let mut offset = 0_usize;
                for part in parts {
                    let dim = part.ambient_dim(1);
                    let projected = part.project_vector_to_gradient_tangent(
                        t.slice(ndarray::s![offset..offset + dim]),
                        g.slice(ndarray::s![offset..offset + dim]),
                        v.slice(ndarray::s![offset..offset + dim]),
                    );
                    for a in 0..dim {
                        out[offset + a] = projected[a];
                    }
                    offset += dim;
                }
                assert_eq!(offset, v.len());
                out
            }
        }
    }

    /// Project every column of `matrix` with
    /// [`Self::project_vector_to_gradient_tangent`].
    pub fn project_matrix_columns_to_gradient_tangent(
        &self,
        t: ArrayView1<'_, f64>,
        g: ArrayView1<'_, f64>,
        matrix: ArrayView2<'_, f64>,
    ) -> Array2<f64> {
        let mut out = Array2::<f64>::zeros(matrix.dim());
        assert_eq!(matrix.nrows(), t.len());
        for col_idx in 0..matrix.ncols() {
            let col = self.project_vector_to_gradient_tangent(t, g, matrix.column(col_idx));
            for row_idx in 0..matrix.nrows() {
                out[[row_idx, col_idx]] = col[row_idx];
            }
        }
        out
    }

    /// Convert Euclidean Hessian action `eh · xi` to Riemannian Hessian action.
    ///
    /// For the sphere this is the Absil/Mahony/Sepulchre embedded-sphere
    /// conversion: differentiate the projected gradient and project back to
    /// the tangent space. The ambient derivative includes the normal
    /// curvature term `-<grad_R, ξ> t`; the tangent action is equivalent to
    /// `P_t(eh ξ) - <eg, t> ξ`.
    pub fn euclidean_to_riemannian_hessian(
        &self,
        t: ArrayView1<'_, f64>,
        eg: ArrayView1<'_, f64>,
        eh: ArrayView2<'_, f64>,
        xi: ArrayView1<'_, f64>,
    ) -> Array1<f64> {
        assert_eq!(t.len(), eg.len());
        assert_eq!(t.len(), xi.len());
        assert_eq!(eh.nrows(), t.len());
        assert_eq!(eh.ncols(), t.len());
        let eh_xi = eh.dot(&xi);
        self.euclidean_hessian_action_to_riemannian(t, eg, xi, eh_xi.view())
    }

    fn euclidean_hessian_action_to_riemannian(
        &self,
        t: ArrayView1<'_, f64>,
        eg: ArrayView1<'_, f64>,
        xi: ArrayView1<'_, f64>,
        eh_xi: ArrayView1<'_, f64>,
    ) -> Array1<f64> {
        assert_eq!(t.len(), eg.len());
        assert_eq!(t.len(), xi.len());
        assert_eq!(t.len(), eh_xi.len());
        match self {
            Self::Euclidean | Self::Circle { .. } => self.project_to_tangent(t, eh_xi),
            Self::Interval { .. } => self.project_vector_to_gradient_tangent(t, eg, eh_xi),
            Self::Sphere { dim } => {
                assert_eq!(t.len(), *dim);
                let grad_r = self.project_to_tangent(t, eg);
                let mut ambient = self.project_to_tangent(t, eh_xi);
                let eg_normal = eg.dot(&t);
                let normal_curve = grad_r.dot(&xi);
                for a in 0..*dim {
                    ambient[a] -= eg_normal * xi[a];
                    ambient[a] -= normal_curve * t[a];
                }
                self.project_to_tangent(t, ambient.view())
            }
            Self::Product(parts)
            | Self::ProductWithMetric {
                manifolds: parts, ..
            } => {
                let mut out = Array1::<f64>::zeros(t.len());
                let mut offset = 0_usize;
                for part in parts {
                    let dim = part.ambient_dim(1);
                    let converted = part.euclidean_hessian_action_to_riemannian(
                        t.slice(ndarray::s![offset..offset + dim]),
                        eg.slice(ndarray::s![offset..offset + dim]),
                        xi.slice(ndarray::s![offset..offset + dim]),
                        eh_xi.slice(ndarray::s![offset..offset + dim]),
                    );
                    for a in 0..dim {
                        out[offset + a] = converted[a];
                    }
                    offset += dim;
                }
                assert_eq!(offset, t.len());
                out
            }
        }
    }

    /// Dense ambient matrix representation of the tangent Hessian action.
    ///
    /// Normal directions are pinned with an identity block for embedded
    /// constrained factors so existing BA Cholesky code can factor the ambient
    /// matrix while RHS/cross blocks stay tangent-projected.
    pub fn riemannian_hessian_matrix(
        &self,
        t: ArrayView1<'_, f64>,
        eg: ArrayView1<'_, f64>,
        eh: ArrayView2<'_, f64>,
    ) -> Array2<f64> {
        let d = t.len();
        let mut out = Array2::<f64>::zeros((d, d));
        let mut xi = Array1::<f64>::zeros(d);
        for a in 0..d {
            xi.fill(0.0);
            xi[a] = 1.0;
            let tangent_xi = self.project_vector_to_gradient_tangent(t, eg, xi.view());
            let col = self.euclidean_to_riemannian_hessian(t, eg, eh, tangent_xi.view());
            for b in 0..d {
                out[[b, a]] = col[b];
            }
        }
        self.add_normal_pinning(t, &mut out);
        symmetrize(&mut out);
        out
    }

    /// Project every column of an ambient matrix into `T_t M`.
    pub fn project_matrix_columns_to_tangent(
        &self,
        t: ArrayView1<'_, f64>,
        matrix: ArrayView2<'_, f64>,
    ) -> Array2<f64> {
        let mut out = Array2::<f64>::zeros(matrix.dim());
        self.project_matrix_columns_to_tangent_into(t, matrix, out.view_mut());
        out
    }

    /// In-place column-wise tangent projection: writes the projection of every
    /// column of `matrix` into the matching column of `out`. Both `matrix` and
    /// `out` must have shape `(ambient_dim × ncols)`. Callers that project the
    /// same `(q × p)` scratch every row hoist `out` outside the loop to avoid
    /// reallocating an `Array2` per row; the projection itself reuses the
    /// allocation-free [`Self::project_to_tangent`] per column.
    pub fn project_matrix_columns_to_tangent_into(
        &self,
        t: ArrayView1<'_, f64>,
        matrix: ArrayView2<'_, f64>,
        mut out: ndarray::ArrayViewMut2<'_, f64>,
    ) {
        assert_eq!(
            matrix.dim(),
            out.dim(),
            "project_matrix_columns_to_tangent_into: matrix {:?} != out {:?}",
            matrix.dim(),
            out.dim(),
        );
        for col_idx in 0..matrix.ncols() {
            let col = self.project_to_tangent(t, matrix.column(col_idx));
            for row_idx in 0..matrix.nrows() {
                out[[row_idx, col_idx]] = col[row_idx];
            }
        }
    }

    fn add_normal_pinning(&self, t: ArrayView1<'_, f64>, matrix: &mut Array2<f64>) {
        match self {
            Self::Sphere { dim } => {
                assert_eq!(t.len(), *dim);
                for a in 0..*dim {
                    for b in 0..*dim {
                        matrix[[a, b]] += SPHERE_NORMAL_PIN * t[a] * t[b];
                    }
                }
            }
            Self::Product(parts)
            | Self::ProductWithMetric {
                manifolds: parts, ..
            } => {
                let mut offset = 0_usize;
                for part in parts {
                    let dim = part.ambient_dim(1);
                    let mut block =
                        matrix.slice_mut(ndarray::s![offset..offset + dim, offset..offset + dim]);
                    let mut owned = block.to_owned();
                    part.add_normal_pinning(t.slice(ndarray::s![offset..offset + dim]), &mut owned);
                    block.assign(&owned);
                    offset += dim;
                }
            }
            Self::Euclidean | Self::Circle { .. } | Self::Interval { .. } => {}
        }
    }
}

impl LatentIdMode {
    /// Fixes the audit finding that ARD/DimSelection alone is rotation
    /// symmetric and therefore not a standalone identifiability mode.
    pub fn is_identifiable(&self) -> bool {
        match self {
            Self::AuxPrior { .. } | Self::AuxPriorDimSelection { .. } => true,
            // A fixed-reference anchor pins the full gauge on its own.
            Self::IsometryToReference { .. } => true,
            // The behavioral head anchors the gauge through the label channel
            // and always composes with ARD axis-selection; it is a standalone
            // identifiable mode provided the head actually carries labels (an
            // empty head pins nothing, rejected by `validate`).
            Self::AuxOutcome { head, .. } => head.effective_labeled_count() > 0.0,
            Self::DimSelection { .. } | Self::None => false,
        }
    }

    /// Validate the mode's identifiability composition (issue #912 step 2).
    ///
    /// `AuxOutcome` must carry a non-vacuous head (at least one labeled row)
    /// and composes with ARD — a bare label channel with no axis-selection
    /// under-pins the gauge. Returns the offending reason on failure so the
    /// builder can reject before fitting. (The former `reject_dim_selection_alone`
    /// guard was unified here into the Result path for a panic-free gate.)
    pub fn validate(&self) -> Result<(), String> {
        if matches!(self, Self::DimSelection { .. }) {
            // `DimSelection` alone is rotation-symmetric — not a valid
            // gauge fix; callers must pair ARD with `AuxPrior`/`Isometry`.
            // Beautiful unification: return a proper error instead of a
            // panic guard (removes the tracked ban stub while keeping the
            // gate).
            return Err("LatentIdMode::DimSelection is not a standalone gauge fix; \
                 pair ARD with AuxPrior or Isometry"
                .to_string());
        }
        if let Self::AuxOutcome { head, .. } = self
            && head.effective_labeled_count() <= 0.0
        {
            return Err(
                "LatentIdMode::AuxOutcome: the behavioral head has no labeled rows \
                 (Σ row-weights = 0); a label-free head pins no gauge dimension. \
                 Provide labels or use AuxPrior/DimSelection composition."
                    .to_string(),
            );
        }
        Ok(())
    }
}

/// Carrier for the `∂Φ/∂t` chain-rule input, dispatched on basis kind by
/// [`LatentCoordValues::design_gradient_wrt_t_dispatch`].
///
/// * [`InputLocationDerivative::Radial`] is the *radial-kernel* path: the
///   caller supplies the radial kernel family together with the center
///   coordinates, and the chain rule
///   `∂Φ/∂t = q(r) · (t − c)` is applied internally. This covers every
///   isotropic radial basis — Duchon (any nullspace order), Matérn (every
///   supported half-integer ν), and anything else whose pointwise
///   gradient is radial. Helpers:
///   [`crate::basis::duchon_radial_first_derivative_nd`],
///   [`crate::basis::matern_radial_first_derivative_nd`].
/// * [`InputLocationDerivative::Jet`] is the *pre-computed jet* path: the
///   caller has already assembled a closed-form `(N, K, d)` tensor for a
///   basis whose chain rule is not a simple radial scalar times a unit
///   vector. Sphere kernels carry the tangent-direction times `K'(cos γ)`;
///   periodic-cyclic B-splines carry the closed-form cardinal derivative;
///   tensor-product B-splines carry the product-rule mix. Helpers:
///   [`crate::basis::sphere_first_derivative_nd`],
///   [`crate::basis::periodic_bspline_first_derivative_nd`],
///   [`crate::basis::bspline_tensor_first_derivative`].
///
/// The dispatch is an enum rather than a trait because each path's
/// arguments differ structurally (radial bases reuse scalar radial kernels shared with
/// the kernel-shape chain machinery; jet bases ship the full tensor). All chain rules
/// are analytic and closed-form; no autodiff, no finite differences.
pub enum InputLocationDerivative<'a> {
    /// Radial-kernel chain rule. The chain rule `(t − c)/r` is reconstructed
    /// internally from the finite `q = φ'(r)/r` scalar and the center coordinates.
    Radial {
        centers: ArrayView2<'a, f64>,
        radial_kind: &'a RadialScalarKind,
    },
    /// Pre-computed analytic `(n_obs, n_centers, latent_dim)` jet.
    Jet(ArrayView3<'a, f64>),
}

/// Per-row latent coordinates `t ∈ ℝ^{N × d}` stored as a flat
/// row-major `Array1<f64>` of length `n_obs * latent_dim`.
///
/// The flat-`Array1` layout mirrors [`crate::smooth::SpatialLogKappaCoords`]
/// so the same `HyperDesignDerivative::from_implicit` / `DirectionalHyperParam`
/// outer plumbing can consume it without modification.
#[derive(Debug, Clone)]
pub struct LatentCoordValues {
    /// Stable process-local identity for this latent-coordinate block.
    id: u64,
    /// Flattened (n_obs, latent_dim) latent matrix, row-major
    /// (so `values[n * d + k] = t_n[k]`).
    values: Array1<f64>,
    /// Number of rows `N`.
    n_obs: usize,
    /// Number of latent dimensions `d`.
    latent_dim: usize,
    /// Identifiability / gauge-fix mode.
    id_mode: LatentIdMode,
    /// Manifold used for per-row Riemannian updates.
    manifold: LatentManifold,
    /// Explicit update-side retraction. The empty registry is Euclidean.
    retraction_registry: LatentRetractionRegistry,
}

impl LatentCoordValues {
    /// Construct from a dense `(n_obs, latent_dim)` matrix.
    pub fn from_matrix(matrix: ArrayView2<'_, f64>, id_mode: LatentIdMode) -> Self {
        Self::from_matrix_with_manifold(matrix, id_mode, LatentManifold::Euclidean)
    }

    /// Construct from a dense matrix and explicit latent manifold.
    pub fn from_matrix_with_manifold(
        matrix: ArrayView2<'_, f64>,
        id_mode: LatentIdMode,
        manifold: LatentManifold,
    ) -> Self {
        Self::from_matrix_with_manifold_and_retraction(
            matrix,
            id_mode,
            manifold,
            LatentRetractionRegistry::all_euclidean(),
        )
    }

    pub fn from_matrix_with_manifold_and_retraction(
        matrix: ArrayView2<'_, f64>,
        id_mode: LatentIdMode,
        manifold: LatentManifold,
        retraction_registry: LatentRetractionRegistry,
    ) -> Self {
        id_mode
            .validate()
            .expect("invalid LatentIdMode for LatentCoordValues::from_matrix_with_manifold");
        let n_obs = matrix.nrows();
        let latent_dim = matrix.ncols();
        retraction_registry
            .validate_dim(latent_dim, "LatentCoordValues::from_matrix_with_manifold")
            .expect("invalid latent retraction dimension");
        let mut values = Array1::<f64>::zeros(n_obs * latent_dim);
        for n in 0..n_obs {
            for k in 0..latent_dim {
                values[n * latent_dim + k] = matrix[[n, k]];
            }
        }
        let mut out = Self {
            id: next_latent_coord_id(),
            values,
            n_obs,
            latent_dim,
            id_mode,
            manifold,
            retraction_registry,
        };
        out.project_all_rows_to_manifold();
        out
    }

    /// Construct directly from a flat (`n_obs * latent_dim`) array.
    pub fn from_flat(
        values: Array1<f64>,
        n_obs: usize,
        latent_dim: usize,
        id_mode: LatentIdMode,
    ) -> Self {
        Self::from_flat_with_manifold(
            values,
            n_obs,
            latent_dim,
            id_mode,
            LatentManifold::Euclidean,
        )
    }

    /// Construct directly from a flat array and explicit latent manifold.
    pub fn from_flat_with_manifold(
        values: Array1<f64>,
        n_obs: usize,
        latent_dim: usize,
        id_mode: LatentIdMode,
        manifold: LatentManifold,
    ) -> Self {
        Self::from_flat_with_manifold_and_retraction_and_id(
            values,
            n_obs,
            latent_dim,
            id_mode,
            manifold,
            LatentRetractionRegistry::all_euclidean(),
            next_latent_coord_id(),
        )
    }

    pub fn from_flat_with_manifold_and_retraction_and_id(
        values: Array1<f64>,
        n_obs: usize,
        latent_dim: usize,
        id_mode: LatentIdMode,
        manifold: LatentManifold,
        retraction_registry: LatentRetractionRegistry,
        id: u64,
    ) -> Self {
        id_mode
            .validate()
            .expect("invalid LatentIdMode for LatentCoordValues::from_flat");
        assert_eq!(
            values.len(),
            n_obs * latent_dim,
            "LatentCoordValues::from_flat: length {} != n_obs * latent_dim = {}",
            values.len(),
            n_obs * latent_dim
        );
        retraction_registry
            .validate_dim(latent_dim, "LatentCoordValues::from_flat_with_manifold")
            .expect("invalid latent retraction dimension");
        let mut out = Self {
            id,
            values,
            n_obs,
            latent_dim,
            id_mode,
            manifold,
            retraction_registry,
        };
        out.project_all_rows_to_manifold();
        out
    }

    pub fn latent_id(&self) -> u64 {
        self.id
    }

    pub fn n_obs(&self) -> usize {
        self.n_obs
    }

    pub fn latent_dim(&self) -> usize {
        self.latent_dim
    }

    /// Total length of the flat value array (= `n_obs * latent_dim`).
    pub fn len(&self) -> usize {
        self.values.len()
    }

    pub fn is_empty(&self) -> bool {
        self.values.is_empty()
    }

    pub fn id_mode(&self) -> &LatentIdMode {
        &self.id_mode
    }

    pub fn manifold(&self) -> &LatentManifold {
        &self.manifold
    }

    pub fn retraction_registry(&self) -> &LatentRetractionRegistry {
        &self.retraction_registry
    }

    /// Effective "is all Euclidean" check used by the inner solver:
    /// returns `true` only when *both* the declared `LatentManifold` and the
    /// optional override retraction registry are Euclidean. The registry's
    /// own `is_all_euclidean` answers a strictly narrower question (was an
    /// explicit non-Euclidean override installed?) and would silently miss
    /// non-Euclidean manifolds installed via `from_matrix_with_manifold` /
    /// `with_manifold`, which left the registry at its `all_euclidean`
    /// default. See `retract_flat_delta` for the matching update path.
    pub fn effective_is_all_euclidean(&self) -> bool {
        self.manifold.is_euclidean() && self.retraction_registry.is_all_euclidean()
    }

    /// Effective per-axis trust-region metric weights. When the manifold is
    /// non-Euclidean it is the authoritative geometric description (it
    /// covers `Interval` and `ProductWithMetric`, which the registry's
    /// `RetractionKind` cannot express), so we read weights from it. When
    /// the manifold is Euclidean but an explicit override retraction was
    /// supplied (e.g. via the JSON `retraction:` key) the registry's
    /// weights win.
    pub fn effective_metric_weights(&self) -> Vec<f64> {
        if self.manifold.is_euclidean() {
            self.retraction_registry.metric_weights(self.latent_dim)
        } else {
            self.manifold.metric_weights()
        }
    }

    /// Effective per-axis periodicity (`Some(period)` on wrapped axes). When
    /// the declared manifold is non-Euclidean it is authoritative; when it is
    /// Euclidean, an explicit override retraction (if any) decides. Returns a
    /// `Vec` of length `latent_dim`.
    pub fn effective_axis_periods(&self) -> Vec<Option<f64>> {
        let periods = if self.manifold.is_euclidean() {
            self.retraction_registry.axis_periods(self.latent_dim)
        } else {
            self.manifold.axis_periods()
        };
        assert_eq!(
            periods.len(),
            self.latent_dim,
            "effective_axis_periods length {} != latent_dim {}",
            periods.len(),
            self.latent_dim
        );
        periods
    }

    pub fn with_manifold(&self, manifold: LatentManifold) -> Self {
        Self::from_flat_with_manifold_and_retraction_and_id(
            self.values.clone(),
            self.n_obs,
            self.latent_dim,
            self.id_mode.clone(),
            manifold,
            self.retraction_registry.clone(),
            self.id,
        )
    }

    /// View the flat value array.
    pub fn as_flat(&self) -> &Array1<f64> {
        &self.values
    }

    /// View row `n` as a length-`d` slice.
    pub fn row(&self, n: usize) -> &[f64] {
        let start = n * self.latent_dim;
        let end = start + self.latent_dim;
        &self.values.as_slice().expect("contiguous")[start..end]
    }

    /// Materialize as a dense `(n_obs, latent_dim)` matrix view.
    /// Useful when handing `t` to a row-major basis evaluator
    /// (e.g. `build_duchon_basis`).
    pub fn as_matrix(&self) -> Array2<f64> {
        let mut out = Array2::<f64>::zeros((self.n_obs, self.latent_dim));
        for n in 0..self.n_obs {
            for k in 0..self.latent_dim {
                out[[n, k]] = self.values[n * self.latent_dim + k];
            }
        }
        out
    }

    /// Mutable write back of the flat value array, e.g. after a Newton step.
    pub fn set_flat(&mut self, flat: ArrayView1<'_, f64>) {
        assert_eq!(flat.len(), self.values.len());
        self.values.assign(&flat);
        self.project_all_rows_to_manifold();
    }

    /// Apply a flat tangent update row-by-row through the manifold retraction.
    pub fn retract_flat_delta(&mut self, delta: ArrayView1<'_, f64>) {
        assert_eq!(delta.len(), self.values.len());
        if self.retraction_registry.is_all_euclidean() {
            if self.manifold.is_euclidean() {
                for (t, dt) in self.values.iter_mut().zip(delta.iter()) {
                    *t += *dt;
                }
                return;
            }
            assert_eq!(
                self.manifold.ambient_dim(self.latent_dim),
                self.latent_dim,
                "LatentCoordValues::retract_flat_delta: manifold ambient dim does not match latent_dim",
            );
            for n in 0..self.n_obs {
                let start = n * self.latent_dim;
                let end = start + self.latent_dim;
                let next = self.manifold.retract(
                    self.values.slice(ndarray::s![start..end]),
                    delta.slice(ndarray::s![start..end]),
                );
                for a in 0..self.latent_dim {
                    self.values[start + a] = next[a];
                }
            }
            return;
        }
        for n in 0..self.n_obs {
            let start = n * self.latent_dim;
            let end = start + self.latent_dim;
            let mut current = self.values.slice_mut(ndarray::s![start..end]);
            let xi = delta.slice(ndarray::s![start..end]);
            self.retraction_registry.retract(&mut current, xi);
        }
    }

    fn project_all_rows_to_manifold(&mut self) {
        if self.manifold.is_euclidean() {
            return;
        }
        assert_eq!(self.manifold.ambient_dim(self.latent_dim), self.latent_dim);
        for n in 0..self.n_obs {
            let start = n * self.latent_dim;
            let end = start + self.latent_dim;
            let projected = self
                .manifold
                .project_point(self.values.slice(ndarray::s![start..end]));
            for a in 0..self.latent_dim {
                self.values[start + a] = projected[a];
            }
        }
    }

    /// Apply this latent block back to a `TermCollectionSpec`-style covariate
    /// table: returns the `(N, d)` materialized matrix that downstream basis
    /// evaluators (Duchon, Matérn, ...) take as their feature input.
    ///
    /// This mirrors [`crate::smooth::SpatialLogKappaCoords::apply_tospec`],
    /// but the carrier on the spec side is the data-row covariate block rather
    /// than the per-term `length_scale`. The spec-mutation is handled at the
    /// call site (the consuming term needs to know which columns of its
    /// feature view to overwrite).
    pub fn apply_tospec(&self) -> Array2<f64> {
        self.as_matrix()
    }

    /// Compute `∂Φ/∂t` for a radial-kernel design Φ — the original
    /// Duchon/Matérn path. See [`Self::design_gradient_wrt_t_dispatch`] for
    /// the basis-agnostic dispatch entry point.
    ///
    /// `centers` is `(n_centers, d)`.
    /// Returns a `(n_obs, n_centers, d)` jet whose `(n, k, a)` entry is
    /// `∂Φ_{n,k} / ∂t_{n,a} = q(r_{n,k}) · (t_{n,a} − c_{k,a})`.
    ///
    /// At `r = 0` the unit vector `(t − c)/r` is undefined; the radial scalar
    /// path therefore asks the kernel for the finite `q` limit and surfaces
    /// `BasisError::DegenerateAtCollision` when that limit does not exist.
    pub(crate) fn design_gradient_wrt_t(
        &self,
        centers: ArrayView2<'_, f64>,
        radial_kind: &RadialScalarKind,
    ) -> Result<Array3<f64>, BasisError> {
        let n_obs = self.n_obs;
        let d = self.latent_dim;
        let n_centers = centers.nrows();
        if centers.ncols() != d {
            crate::bail_dim_basis!(
                "LatentCoordValues::design_gradient_wrt_t center dimension mismatch: centers have {} cols but latent_dim is {}",
                centers.ncols(),
                d
            );
        }
        let mut jet = Array3::<f64>::zeros((n_obs, n_centers, d));
        for n in 0..n_obs {
            let t_n = self.row(n);
            for k in 0..n_centers {
                let mut r2 = 0.0_f64;
                for a in 0..d {
                    let delta = t_n[a] - centers[[k, a]];
                    r2 += delta * delta;
                }
                let r = r2.sqrt();
                let (_, q, _) = radial_kind.eval_design_triplet(r)?;
                if q == 0.0 {
                    continue;
                }
                for a in 0..d {
                    jet[[n, k, a]] = q * (t_n[a] - centers[[k, a]]);
                }
            }
        }
        Ok(jet)
    }

    /// Compute `∂Φ/∂t` for an arbitrary supported basis kind, by dispatching
    /// to the right closed-form chain rule.
    ///
    /// All radial-kernel bases (Duchon, Matérn) reduce to the same
    /// `q(r) · (t − c)` chain that `design_gradient_wrt_t` already implements.
    /// Non-radial bases (sphere, periodic-cyclic B-spline, tensor
    /// B-spline) carry their own analytic `(N, K, d)` jet — the caller
    /// pre-builds that jet using the matching `*_first_derivative_nd` helper
    /// in [`crate::basis`] and passes it in via
    /// [`InputLocationDerivative::Jet`].
    ///
    /// This is the single entry point the outer optimizer should call; it
    /// stays in lock-step with the kernel-parameter chain rule that
    /// `SpatialLogKappaCoords` uses (re-pointed at the first kernel argument
    /// rather than at kernel anisotropy).
    pub(crate) fn design_gradient_wrt_t_dispatch(
        &self,
        input: InputLocationDerivative<'_>,
    ) -> Result<Array3<f64>, BasisError> {
        match input {
            InputLocationDerivative::Radial {
                centers,
                radial_kind,
            } => self.design_gradient_wrt_t(centers, radial_kind),
            InputLocationDerivative::Jet(jet) => {
                if jet.shape() != [self.n_obs, jet.shape()[1], self.latent_dim] {
                    crate::bail_dim_basis!(
                        "LatentCoordValues::design_gradient_wrt_t_dispatch jet shape {:?} does not match latent shape ({}, {}, {})",
                        jet.shape(),
                        self.n_obs,
                        jet.shape()[1],
                        self.latent_dim
                    );
                }
                // The non-radial helpers already produce a (N, K, d) tensor
                // in the layout downstream contraction consumes. Return a copy
                // so the caller owns the data and is decoupled from the source
                // array's lifetime.
                Ok(jet.to_owned())
            }
        }
    }
}

/// Minimum total active assignment mass a per-row atom code must retain after a
/// Whether the active assignment mass of a per-row code is degenerate
/// (non-finite), i.e. a NaN/Inf assignment.
///
/// #1074: the magic `LATENT_ACTIVE_MASS_FLOOR = 1e-6` collapse-detect-and-reseed
/// floor was DELETED. It masked the real defect — rows being driven dark by the
/// assignment optimizer — with a detect-then-reseed bandaid rather than
/// preventing the collapse. Only the genuine NaN/Inf guard remains. The proper
/// fix (prevent the active-set collapse in the assignment step) is tracked
/// separately.
///
/// `active_weights` is the slice of per-atom assignment weights on the row's
/// active support. Returns `true` only when the summed magnitude is non-finite.
pub fn active_mass_breached(active_weights: &[f64]) -> bool {
    let mut mass = 0.0_f64;
    for &w in active_weights {
        mass += w.abs();
    }
    !mass.is_finite()
}

fn wrap_to_period(x: f64, period: f64) -> f64 {
    assert!(
        period.is_finite() && period > 0.0,
        "wrap_to_period requires a finite positive period; got {period}"
    );
    let y = x.rem_euclid(period);
    if y == period { 0.0 } else { y }
}

/// Normalize `v[0..dim]` to a unit vector (for `LatentManifold::Sphere`
/// projection and retraction).
///
/// "Or axis": if the input is zero or non-finite (degenerate or numerical
/// mishap in caller), gracefully fall back to the canonical first axis
/// unit vector `[1, 0, …, 0]`. This removes a hard panic while preserving
/// the sphere contract that every returned point has unit Euclidean norm.
/// Callers (project_point / retract on Sphere) already ensure dim matches
/// the view length for the manifold component.
fn normalize_or_axis(v: ArrayView1<'_, f64>, dim: usize) -> Array1<f64> {
    let mut norm_sq = 0.0_f64;
    for a in 0..dim {
        norm_sq += v[a] * v[a];
    }
    const EPS: f64 = 1e-300; // protect against underflow/denorm that would give Inf
    if norm_sq > EPS && norm_sq.is_finite() {
        let inv = 1.0 / norm_sq.sqrt();
        let mut out = Array1::<f64>::zeros(dim);
        for a in 0..dim {
            out[a] = v[a] * inv;
        }
        out
    } else {
        // "or axis" fallback — beautiful, non-panicking resolution for
        // degenerate ambient vector on the sphere.
        let mut out = Array1::<f64>::zeros(dim);
        if dim > 0 {
            out[0] = 1.0;
        }
        out
    }
}

#[inline]
fn symmetrize(a: &mut Array2<f64>) {
    // Callers in this module always pass square (d, d) matrices; delegate to
    // the canonical helper in `linalg::utils`.
    gam_linalg::matrix::symmetrize_in_place(a)
}

/// Auxiliary-prior penalty contribution: returns the per-row reference
/// coordinates `ĥ(u_n)` shape `(n_obs, d)` and the effective strength `μ`.
///
/// `t_target` is broadcast across the inner ridge of `½ μ · ‖t − t_target‖²`,
/// which the call site folds into the Y-stack via a virtual-row augmentation
/// (`y' = [y; √μ · t_target]`, `X' = [X; √μ · I_d ⊗ row-block]`). This
/// keeps the inner solver Gaussian-closed-form.
///
/// For `AuxPriorFamily::Ridge` the conditional mean is the closed-form ridge
/// regression `(UᵀU + ε I)⁻¹ UᵀT` evaluated at each row's `u_n`. For
/// `Linear` the ridge is zero (which raises if `UᵀU` is singular).
/// Closed-form auxiliary-prior REML statistics at a fixed outer coordinate `t`.
pub struct AuxPriorRemlStats {
    pub residual_sq: f64,
    pub log_mu: f64,
    pub mu: f64,
    pub auto: bool,
    pub score: f64,
}

/// Auxiliary-prior REML statistics for a fixed outer coordinate `t`, given the
/// precomputed `targets` (see [`aux_prior_targets`]). Returns the residual sum of
/// squares, the precision `mu` (the supplied `aux_strength` when `Some`, else the
/// closed-form REML optimum `mu = K / Σr²`), whether it was auto-selected, and
/// the prior score `0.5·mu·Σr² − 0.5·K·ln(mu)`. The `log_mu` coordinate has this
/// closed-form optimum at fixed `t` because only the normalized auxiliary prior
/// depends on it.
///
/// `K = n_obs · latent_dim` is the number of scalar latent coordinates the single
/// shared precision `mu` governs. The normalizer term `−0.5·K·ln(mu)` is the prior
/// log-determinant `−0.5·log det₊(mu · I_K)`, so it counts every governed
/// coordinate. Counting only `n_obs` undercounts a `latent_dim`-dimensional latent
/// by exactly `latent_dim`, which biases the REML precision toward under-shrinkage
/// (the per-axis ARD path emits `−0.5·n_obs·ln(α)` for each of `latent_dim` axes;
/// a single shared `mu` must match that sum).
pub fn aux_prior_reml_stats(
    t_mat: ArrayView2<'_, f64>,
    targets: ArrayView2<'_, f64>,
    aux_strength: Option<f64>,
) -> Result<AuxPriorRemlStats, String> {
    let n_obs = t_mat.nrows();
    let latent_dim = t_mat.ncols();
    let mut residual_sq = 0.0_f64;
    for n in 0..n_obs {
        for a in 0..latent_dim {
            let diff = t_mat[[n, a]] - targets[[n, a]];
            residual_sq += diff * diff;
        }
    }
    if !residual_sq.is_finite() {
        return Err("auxiliary prior residual norm must be finite".to_string());
    }
    let (log_mu, mu, auto) = match aux_strength {
        Some(mu) => {
            if !(mu.is_finite() && mu > 0.0) {
                return Err(format!(
                    "aux_strength must be finite and positive; got {mu}"
                ));
            }
            (mu.ln(), mu, false)
        }
        None => {
            if residual_sq <= 0.0 {
                return Err(
                    "aux_strength='auto' has no finite REML optimum when the auxiliary residual is zero"
                        .to_string(),
                );
            }
            let mu = ((n_obs * latent_dim) as f64) / residual_sq;
            if !(mu.is_finite() && mu > 0.0) {
                return Err(format!(
                    "auto aux_strength selected a non-finite precision: {mu}"
                ));
            }
            (mu.ln(), mu, true)
        }
    };
    let score = 0.5 * mu * residual_sq - 0.5 * ((n_obs * latent_dim) as f64) * log_mu;
    Ok(AuxPriorRemlStats {
        residual_sq,
        log_mu,
        mu,
        auto,
        score,
    })
}

pub fn aux_prior_targets(
    t: ArrayView2<'_, f64>,
    u: ArrayView2<'_, f64>,
    family: AuxPriorFamily,
) -> Result<Array2<f64>, String> {
    let n_obs = t.nrows();
    let d = t.ncols();
    if u.nrows() != n_obs {
        return Err(format!(
            "aux_prior_targets: u has {} rows but t has {}",
            u.nrows(),
            n_obs
        ));
    }
    let p = u.ncols();
    if p == 0 {
        return Err("aux_prior_targets: auxiliary u must have at least one column".into());
    }
    // gram = UᵀU  (p × p)
    let mut gram = Array2::<f64>::zeros((p, p));
    for n in 0..n_obs {
        for i in 0..p {
            for j in 0..p {
                gram[[i, j]] += u[[n, i]] * u[[n, j]];
            }
        }
    }
    let ridge_eps = match family {
        AuxPriorFamily::Ridge => {
            let trace: f64 = (0..p).map(|i| gram[[i, i]]).sum();
            (1e-6 * trace / p as f64).max(1e-12)
        }
        AuxPriorFamily::Linear => 0.0,
    };
    for i in 0..p {
        gram[[i, i]] += ridge_eps;
    }
    // rhs = UᵀT  (p × d)
    let mut rhs = Array2::<f64>::zeros((p, d));
    for n in 0..n_obs {
        for i in 0..p {
            for k in 0..d {
                rhs[[i, k]] += u[[n, i]] * t[[n, k]];
            }
        }
    }
    let coeffs = solve_spd(gram.view(), rhs.view())?;
    // targets = U · coeffs  (n_obs × d)
    let mut targets = Array2::<f64>::zeros((n_obs, d));
    for n in 0..n_obs {
        for k in 0..d {
            let mut acc = 0.0_f64;
            for i in 0..p {
                acc += u[[n, i]] * coeffs[[i, k]];
            }
            targets[[n, k]] = acc;
        }
    }
    Ok(targets)
}

/// Lightweight Cholesky-based SPD solve. Keeps this module dependency-free
/// from the broader faer-wrapping surface; matrices here are tiny
/// (`p × p` with p = aux-feature count, typically O(10)).
fn solve_spd(a: ArrayView2<'_, f64>, b: ArrayView2<'_, f64>) -> Result<Array2<f64>, String> {
    let n = a.nrows();
    if a.ncols() != n {
        return Err("solve_spd: A must be square".into());
    }
    if b.nrows() != n {
        return Err("solve_spd: RHS row count mismatch".into());
    }
    // In-place Cholesky factorization. We pay the O(n³) copy + O(n³) factor
    // up front; n is tiny in the auxiliary-prior path.
    let mut l = Array2::<f64>::zeros((n, n));
    for i in 0..n {
        for j in 0..=i {
            let mut sum = a[[i, j]];
            for k in 0..j {
                sum -= l[[i, k]] * l[[j, k]];
            }
            if i == j {
                if sum <= 0.0 {
                    return Err(format!(
                        "solve_spd: non-positive pivot {sum} at index {i} \
                         (matrix is not positive definite)"
                    ));
                }
                l[[i, j]] = sum.sqrt();
            } else {
                l[[i, j]] = sum / l[[j, j]];
            }
        }
    }
    // Solve L y = b, then Lᵀ x = y, column by column.
    let d = b.ncols();
    let mut out = Array2::<f64>::zeros((n, d));
    for col in 0..d {
        let mut y = Array1::<f64>::zeros(n);
        for i in 0..n {
            let mut sum = b[[i, col]];
            for k in 0..i {
                sum -= l[[i, k]] * y[k];
            }
            y[i] = sum / l[[i, i]];
        }
        for i in (0..n).rev() {
            let mut sum = y[i];
            for k in (i + 1)..n {
                sum -= l[[k, i]] * out[[k, col]];
            }
            out[[i, col]] = sum / l[[i, i]];
        }
    }
    Ok(out)
}

#[cfg(test)]
mod tests {
    use super::*;
    use ndarray::array;

    #[test]
    fn from_matrix_roundtrip() {
        let m = array![[1.0_f64, 2.0], [3.0, 4.0], [5.0, 6.0]];
        let lc = LatentCoordValues::from_matrix(m.view(), LatentIdMode::None);
        assert_eq!(lc.n_obs(), 3);
        assert_eq!(lc.latent_dim(), 2);
        let back = lc.as_matrix();
        assert_eq!(back, m);
    }

    #[test]
    fn row_access() {
        let m = array![[1.0_f64, 2.0], [3.0, 4.0]];
        let lc = LatentCoordValues::from_matrix(m.view(), LatentIdMode::None);
        assert_eq!(lc.row(0), &[1.0, 2.0]);
        assert_eq!(lc.row(1), &[3.0, 4.0]);
    }

    /// `preserves_isometry_cross_block_coherence` must report exactly the
    /// charts whose Euclidean→Riemannian geometry transform is the identity on
    /// the per-row gradient / `H_tt` blocks. Keying the SAE isometry
    /// cross-block coupling decision on `is_euclidean()` instead of this
    /// predicate dropped the cross-block on the flat `Circle` chart, leaving a
    /// block-diagonal Hessian whose joint Newton step never reached KKT
    /// stationarity — the arrow-Schur proximal ridge then saturated at 1e15
    /// (issue #795, regression of #681). We pin the predicate AND its grounding
    /// invariant: on `Circle` the geometry transform really is the identity, so
    /// coherence is preserved; on `Sphere` / `Interval` it is not.
    #[test]
    fn isometry_cross_block_coherence_tracks_identity_geometry_transform() {
        assert!(LatentManifold::Euclidean.preserves_isometry_cross_block_coherence());
        assert!(
            LatentManifold::Circle {
                period: std::f64::consts::TAU
            }
            .preserves_isometry_cross_block_coherence()
        );
        assert!(!LatentManifold::Sphere { dim: 3 }.preserves_isometry_cross_block_coherence());
        assert!(
            !LatentManifold::Interval { lo: -1.0, hi: 1.0 }
                .preserves_isometry_cross_block_coherence()
        );
        // A Product is coherent iff every factor is.
        assert!(
            LatentManifold::Product(vec![
                LatentManifold::Euclidean,
                LatentManifold::Circle {
                    period: std::f64::consts::TAU
                },
            ])
            .preserves_isometry_cross_block_coherence()
        );
        assert!(
            !LatentManifold::Product(vec![
                LatentManifold::Circle {
                    period: std::f64::consts::TAU
                },
                LatentManifold::Sphere { dim: 3 },
            ])
            .preserves_isometry_cross_block_coherence()
        );

        // Grounding invariant: on the Circle chart the geometry transform that
        // `apply_riemannian_latent_geometry` applies — gradient projection and
        // the Euclidean→Riemannian Hessian conversion — is the EXACT identity,
        // so the coupled `μ AᵀA` block survives intact and the cross-block must
        // be kept.
        let circle = LatentManifold::Circle {
            period: std::f64::consts::TAU,
        };
        let t = array![0.73_f64];
        let eg = array![2.4_f64];
        let eh = array![[1.7_f64]];
        let projected_g = circle.project_gradient_to_tangent(t.view(), eg.view());
        assert_eq!(
            projected_g, eg,
            "Circle gradient projection must be identity"
        );
        let rhess = circle.riemannian_hessian_matrix(t.view(), eg.view(), eh.view());
        assert_eq!(
            rhess, eh,
            "Circle Riemannian Hessian must equal the Euclidean Hessian"
        );
    }

    /// `project_matrix_columns_to_tangent_into` (the hoisted, allocation-reuse
    /// projection used by the SAE arrow-Schur assembler) must match the
    /// per-column ground truth `project_to_tangent`, and must agree exactly
    /// with the allocating `project_matrix_columns_to_tangent` it backs, on a
    /// non-Euclidean (Sphere) manifold where the tangent projection is
    /// non-trivial. This pins the in-place projection introduced for the SAE
    /// hot-path scratch hoist.
    #[test]
    fn project_matrix_columns_to_tangent_into_matches_columnwise() {
        let manifold = LatentManifold::Sphere { dim: 3 };
        // Unit base point on S².
        let norm = (1.0_f64 + 4.0 + 4.0).sqrt();
        let t = array![1.0 / norm, 2.0 / norm, 2.0 / norm];
        let matrix = array![
            [0.3_f64, -1.1, 0.7, 2.0],
            [1.5, 0.2, -0.4, 0.9],
            [-0.6, 0.8, 1.3, -1.7],
        ];
        let mut into = Array2::<f64>::zeros(matrix.dim());
        manifold.project_matrix_columns_to_tangent_into(t.view(), matrix.view(), into.view_mut());
        let allocating = manifold.project_matrix_columns_to_tangent(t.view(), matrix.view());
        for col_idx in 0..matrix.ncols() {
            let expected = manifold.project_to_tangent(t.view(), matrix.column(col_idx));
            for row_idx in 0..matrix.nrows() {
                assert!(
                    (into[[row_idx, col_idx]] - expected[row_idx]).abs() < 1e-12,
                    "in-place projection deviates from columnwise truth at ({row_idx},{col_idx})"
                );
                assert_eq!(
                    into[[row_idx, col_idx]],
                    allocating[[row_idx, col_idx]],
                    "in-place and allocating projection differ at ({row_idx},{col_idx})"
                );
            }
        }
    }

    /// Regression for #2295: a composite `Product` mixing a d=1 factor with a
    /// d=2 factor must split the flat gradient at the correct per-part offsets
    /// (each factor is projected at ITS OWN ambient width), round-trip with
    /// `offset == g.len()`, and reproduce every factor's standalone projection.
    /// The joint mixed-dimension superposition path (zoo dims=[1,1,2,2,2,2,2,1])
    /// drives exactly this split; a per-part width that ignored a factor's true
    /// dimension miscounted the offsets and tripped `assert_eq!(offset, g.len())`
    /// after the first sub-dimensional factor.
    #[test]
    fn product_gradient_projection_splits_mixed_dimensional_factors() {
        let circle = LatentManifold::Circle {
            period: std::f64::consts::TAU,
        };
        // `Sphere { dim: 2 }` is S¹ embedded in R², i.e. a genuinely 2-wide
        // ambient block whose tangent projection removes the radial component —
        // a non-trivial d=2 factor next to the flat d=1 circle.
        let sphere = LatentManifold::Sphere { dim: 2 };
        let product = LatentManifold::Product(vec![circle.clone(), sphere.clone()]);

        // Ambient width = 1 (circle) + 2 (sphere) = 3, split at offsets 0 and 1.
        assert_eq!(product.ambient_dim(3), 3);

        // Base point: the circle coordinate, then a unit 2-vector for the sphere.
        let t = array![0.5_f64, 0.6, 0.8];
        let g = array![1.3_f64, 2.0, -0.7];

        let projected = product.project_gradient_to_tangent(t.view(), g.view());
        assert_eq!(
            projected.len(),
            3,
            "composite output tiles the full ambient"
        );

        // Each factor projected standalone at its own offset/width must match the
        // composite's corresponding block.
        let circle_block = circle
            .project_gradient_to_tangent(t.slice(ndarray::s![0..1]), g.slice(ndarray::s![0..1]));
        let sphere_block = sphere
            .project_gradient_to_tangent(t.slice(ndarray::s![1..3]), g.slice(ndarray::s![1..3]));
        assert_eq!(projected[0], circle_block[0], "d=1 circle factor block");
        for a in 0..2 {
            assert_eq!(projected[1 + a], sphere_block[a], "d=2 sphere factor block");
        }

        // Non-triviality guard: the sphere block genuinely removed the radial
        // component, so this is not a vacuous identity round-trip.
        let radial = g[1] * t[1] + g[2] * t[2];
        assert!(
            radial.abs() > 1e-6,
            "fixture must exercise a non-tangent gradient on the sphere factor"
        );
        assert!(
            (sphere_block[0] - (g[1] - radial * t[1])).abs() < 1e-12
                && (sphere_block[1] - (g[2] - radial * t[2])).abs() < 1e-12,
            "sphere tangent projection must remove the radial component"
        );
    }

    /// Regression for issue #191 (and the K=2 periodic case of #174):
    /// `from_matrix_with_manifold(Circle)` must produce a value whose
    /// update path wraps into `[0, 2π)` even though the override
    /// `LatentRetractionRegistry` is left at its `all_euclidean` default.
    /// Before the fix, the retraction silently decayed to Euclidean and
    /// values drifted outside the circle on every Newton step.
    #[test]
    fn circle_manifold_update_wraps_into_canonical_interval() {
        let two_pi = std::f64::consts::TAU;
        let near_top = 6.2_f64;
        let m = array![[near_top]];
        let mut lc = LatentCoordValues::from_matrix_with_manifold(
            m.view(),
            LatentIdMode::None,
            LatentManifold::Circle { period: two_pi },
        );
        let delta = Array1::from(vec![1.5_f64]);
        lc.retract_flat_delta(delta.view());
        let updated = lc.row(0)[0];
        let expected = (near_top + 1.5).rem_euclid(two_pi);
        assert!(
            (0.0..two_pi).contains(&updated),
            "Circle retraction did not wrap into [0, 2π): got {updated}",
        );
        assert!(
            (updated - expected).abs() < 1e-12,
            "Circle retraction value mismatch: got {updated}, expected {expected}",
        );

        let large_delta = Array1::from(vec![10.0 * two_pi + 0.25_f64]);
        lc.retract_flat_delta(large_delta.view());
        let after_big = lc.row(0)[0];
        assert!(
            (0.0..two_pi).contains(&after_big),
            "Circle retraction did not wrap a large delta: got {after_big}",
        );
    }

    /// Mirror of the Circle regression for `LatentManifold::Sphere`: the
    /// per-row update must preserve unit norm. Before the fix the registry
    /// stayed Euclidean and the additive update broke the constraint.
    #[test]
    fn sphere_manifold_update_preserves_unit_norm() {
        let m = array![[1.0_f64, 0.0, 0.0]];
        let mut lc = LatentCoordValues::from_matrix_with_manifold(
            m.view(),
            LatentIdMode::None,
            LatentManifold::Sphere { dim: 3 },
        );
        let delta = Array1::from(vec![0.3_f64, 0.7, -0.2]);
        lc.retract_flat_delta(delta.view());
        let row = lc.row(0);
        let norm_sq: f64 = row.iter().map(|x| x * x).sum();
        assert!(
            (norm_sq.sqrt() - 1.0).abs() < 1e-12,
            "Sphere retraction did not preserve unit norm: ||t|| = {}",
            norm_sq.sqrt(),
        );

        let big_delta = Array1::from(vec![50.0_f64, -25.0, 13.0]);
        lc.retract_flat_delta(big_delta.view());
        let row2 = lc.row(0);
        let norm_sq2: f64 = row2.iter().map(|x| x * x).sum();
        assert!(
            (norm_sq2.sqrt() - 1.0).abs() < 1e-12,
            "Sphere retraction failed to renormalize after large delta: ||t|| = {}",
            norm_sq2.sqrt(),
        );
    }
}