Skip to main content

gam_geometry/
response_geometry.rs

1//! User-selectable response geometries beyond Sphere and Simplex.
2//!
3//! The fit DSL exposes `response_geometry="..."`: one scalar Gaussian GAM is
4//! fitted per tangent coordinate at a fixed base point (the intrinsic Fréchet
5//! mean when none is supplied), and predictions are mapped back to the manifold
6//! by the exponential map. Sphere and Simplex have bespoke batched wrappers in
7//! their own modules; this module supplies the same `(values 2-D, base 1-D) →
8//! tangent 2-D` / `(tangent 2-D, base 1-D) → values 2-D` contract for the
9//! curved matrix manifolds whose per-point math is already wired in
10//! [`crate::geometry`] but which were never reachable as a *fittable* response
11//! geometry: the SPD cone `Sym⁺(n)`, the Grassmannian `Gr(k, n)`, the Stiefel
12//! manifold `St(k, n)`, and the Poincaré ball `B^d_κ`.
13//!
14//! Every primitive here delegates to the canonical landed math
15//! ([`RiemannianManifold::exp_map`]/[`log_map`](RiemannianManifold::log_map) and
16//! the Poincaré [`exp_map`](crate::manifolds::poincare::exp_map)/[`log_map`](crate::manifolds::poincare::log_map));
17//! the only new code is the batched row loop, the base-point dimension wiring,
18//! and a generic Riemannian Karcher (Fréchet) mean shared by all four. There is
19//! no separate per-manifold mean: the SPD safeguarded Karcher iteration is
20//! generalised once, over the metric supplied by
21//! [`RiemannianManifold::metric_tensor`], so adding a curved response geometry
22//! is a single resolver arm.
23
24use ndarray::{Array1, Array2, ArrayView1, ArrayView2};
25
26use crate::manifold::{
27    GEOMETRY_EPS, RiemannianManifold, flatten, from_flat, jacobi_symmetric, spectral_map_symmetric,
28    sym,
29};
30use crate::manifolds::constant_curvature::ConstantCurvature;
31use crate::{GeometryError, GeometryResult, GrassmannManifold, SpdManifold, StiefelManifold};
32
33/// Split a parenthesised `key=value, key=value` parameter list into trimmed,
34/// lower-cased `(key, value)` pairs. An empty list is valid (`spd()`).
35fn parse_kv(inner: &str) -> Result<Vec<(String, String)>, String> {
36    let trimmed = inner.trim();
37    if trimmed.is_empty() {
38        return Ok(Vec::new());
39    }
40    let mut out = Vec::new();
41    for piece in trimmed.split(',') {
42        let piece = piece.trim();
43        if piece.is_empty() {
44            continue;
45        }
46        let (k, v) = piece
47            .split_once('=')
48            .ok_or_else(|| format!("response_geometry parameter {piece:?} must be key=value"))?;
49        out.push((k.trim().to_ascii_lowercase(), v.trim().to_string()));
50    }
51    Ok(out)
52}
53
54/// A fittable curved response geometry. Each variant carries the shape the user
55/// requested; the embedding/ambient flat dimension is fixed by that shape and
56/// is the column count of the `values` matrix the caller supplies.
57#[derive(Debug, Clone, Copy, PartialEq)]
58pub enum ResponseManifold {
59    /// Symmetric positive-definite `n×n` matrices, flattened row-major to `n²`
60    /// ambient coordinates (the layout [`SpdManifold`] uses).
61    Spd { n: usize },
62    /// `k`-dimensional subspaces of `ℝⁿ`, represented by an orthonormal `n×k`
63    /// frame flattened to `n·k` ambient coordinates.
64    Grassmann { k: usize, n: usize },
65    /// Orthonormal `k`-frames in `ℝⁿ`, flattened to `n·k` ambient coordinates.
66    Stiefel { k: usize, n: usize },
67    /// The Poincaré ball of dimension `d` with curvature `κ < 0`.
68    Poincare { dim: usize, curvature: f64 },
69    /// Constant-curvature manifold `M_κ` of dimension `d` with curvature `κ`
70    /// (any finite real value). `κ > 0` → spherical, `κ = 0` → flat (Euclidean
71    /// up to scale), `κ < 0` → hyperbolic (Poincaré ball). Unlike `Poincare`,
72    /// which fixes `κ < 0`, this variant accepts any curvature including zero
73    /// and positive values, and is the target for curvature-as-estimand fits
74    /// where `κ̂` is optimized over all of ℝ (#1104).
75    ConstantCurvature { dim: usize, kappa: f64 },
76}
77
78impl ResponseManifold {
79    /// Resolve a lower-cased geometry label and its shape parameters into a
80    /// response manifold. Shape parameters are passed positionally exactly as
81    /// the FFI marshals them; absent/zero values are rejected here so the error
82    /// surfaces at selection time rather than mid-fit.
83    ///
84    /// - `"spd"` needs `n` (matrix side).
85    /// - `"grassmann"` / `"stiefel"` need `k` and `n` with `1 ≤ k ≤ n`.
86    /// - `"poincare"` needs `dim` and a strictly negative `curvature`.
87    pub fn resolve(
88        kind: &str,
89        n: Option<usize>,
90        k: Option<usize>,
91        dim: Option<usize>,
92        curvature: Option<f64>,
93    ) -> Result<Self, String> {
94        match kind {
95            "spd" => {
96                let n = n.ok_or_else(|| "response_geometry='spd' requires n".to_string())?;
97                if n == 0 {
98                    return Err("response_geometry='spd' requires n >= 1".to_string());
99                }
100                Ok(Self::Spd { n })
101            }
102            "grassmann" => {
103                let k = k.ok_or_else(|| "response_geometry='grassmann' requires k".to_string())?;
104                let n = n.ok_or_else(|| "response_geometry='grassmann' requires n".to_string())?;
105                if k == 0 || n == 0 || k > n {
106                    return Err("response_geometry='grassmann' requires 1 <= k <= n".to_string());
107                }
108                Ok(Self::Grassmann { k, n })
109            }
110            "stiefel" => {
111                let k = k.ok_or_else(|| "response_geometry='stiefel' requires k".to_string())?;
112                let n = n.ok_or_else(|| "response_geometry='stiefel' requires n".to_string())?;
113                if k == 0 || n == 0 || k > n {
114                    return Err("response_geometry='stiefel' requires 1 <= k <= n".to_string());
115                }
116                Ok(Self::Stiefel { k, n })
117            }
118            "poincare" => {
119                let dim =
120                    dim.ok_or_else(|| "response_geometry='poincare' requires dim".to_string())?;
121                if dim == 0 {
122                    return Err("response_geometry='poincare' requires dim >= 1".to_string());
123                }
124                let curvature = curvature
125                    .ok_or_else(|| "response_geometry='poincare' requires curvature".to_string())?;
126                if !(curvature.is_finite() && curvature < 0.0) {
127                    return Err(
128                        "response_geometry='poincare' requires finite curvature < 0".to_string()
129                    );
130                }
131                Ok(Self::Poincare { dim, curvature })
132            }
133            "constant_curvature" => {
134                let dim = dim.ok_or_else(|| {
135                    "response_geometry='constant_curvature' requires dim".to_string()
136                })?;
137                if dim == 0 {
138                    return Err(
139                        "response_geometry='constant_curvature' requires dim >= 1".to_string()
140                    );
141                }
142                // curvature defaults to 0 (flat) when not supplied — the user can
143                // supply any finite value; the κ-estimand outer loop will optimize it.
144                let kappa = curvature.unwrap_or(0.0);
145                if !kappa.is_finite() {
146                    return Err(
147                        "response_geometry='constant_curvature' requires finite curvature"
148                            .to_string(),
149                    );
150                }
151                Ok(Self::ConstantCurvature { dim, kappa })
152            }
153            other => Err(format!(
154                "response_geometry must be one of 'spd', 'grassmann', 'stiefel', 'poincare', \
155                 'constant_curvature', 'spherical', or 'simplex'; got {other:?}"
156            )),
157        }
158    }
159
160    /// Parse a user-facing `response_geometry` label, magic-by-default: the head
161    /// is the geometry name, an optional parenthesised `key=value` list carries
162    /// shape parameters, and anything not given is inferred from the ambient
163    /// column count `cols` of the response matrix.
164    ///
165    /// Recognised forms (case-insensitive, whitespace tolerant):
166    /// - `"spd"` — `n = √cols` (must be a perfect square).
167    /// - `"grassmann(k=2)"` or `"grassmann(k=2,n=5)"` — `n` defaults to
168    ///   `cols / k`; `k` is required (it cannot be inferred from `n·k`).
169    /// - `"stiefel(k=2)"` / `"stiefel(k=2,n=5)"` — same inference as Grassmann.
170    /// - `"poincare"` or `"poincare(curvature=-0.5)"` — `dim = cols`; curvature
171    ///   defaults to `-1.0`.
172    ///
173    /// This is the single mapping from the formula-DSL string to a constructed
174    /// response manifold; the FFI passes the raw label straight through.
175    pub fn parse(label: &str, cols: usize) -> Result<Self, String> {
176        let lowered = label.trim().to_ascii_lowercase();
177        let (head, params) = match lowered.split_once('(') {
178            Some((h, rest)) => {
179                let rest = rest.trim_end();
180                let inner = rest
181                    .strip_suffix(')')
182                    .ok_or_else(|| format!("response_geometry {label:?}: missing closing ')'"))?;
183                (h.trim().to_string(), parse_kv(inner)?)
184            }
185            None => (lowered.clone(), Vec::new()),
186        };
187        let get_usize = |key: &str| -> Result<Option<usize>, String> {
188            for (k, v) in &params {
189                if k == key {
190                    let parsed: usize = v.parse().map_err(|_| {
191                        format!("response_geometry {label:?}: {key} must be a non-negative integer")
192                    })?;
193                    return Ok(Some(parsed));
194                }
195            }
196            Ok(None)
197        };
198        let get_f64 = |key: &str| -> Result<Option<f64>, String> {
199            for (k, v) in &params {
200                if k == key {
201                    let parsed: f64 = v.parse().map_err(|_| {
202                        format!("response_geometry {label:?}: {key} must be a real number")
203                    })?;
204                    return Ok(Some(parsed));
205                }
206            }
207            Ok(None)
208        };
209
210        match head.as_str() {
211            "spd" => {
212                let n = match get_usize("n")? {
213                    Some(n) => n,
214                    None => {
215                        let r = (cols as f64).sqrt().round() as usize;
216                        if r * r != cols {
217                            return Err(format!(
218                                "response_geometry='spd': {cols} response columns is not a perfect \
219                                 square; pass spd(n=...) explicitly"
220                            ));
221                        }
222                        r
223                    }
224                };
225                Self::resolve("spd", Some(n), None, None, None)
226            }
227            "grassmann" | "stiefel" => {
228                let k = get_usize("k")?.ok_or_else(|| {
229                    format!("response_geometry='{head}' requires k, e.g. {head}(k=2)")
230                })?;
231                let n = match get_usize("n")? {
232                    Some(n) => n,
233                    None => {
234                        if k == 0 || cols % k != 0 {
235                            return Err(format!(
236                                "response_geometry='{head}': {cols} response columns is not \
237                                 divisible by k={k}; pass {head}(k=..,n=..) explicitly"
238                            ));
239                        }
240                        cols / k
241                    }
242                };
243                Self::resolve(&head, Some(n), Some(k), None, None)
244            }
245            "poincare" => {
246                let dim = get_usize("dim")?.unwrap_or(cols);
247                let curvature = get_f64("curvature")?.unwrap_or(-1.0);
248                Self::resolve("poincare", None, None, Some(dim), Some(curvature))
249            }
250            "constant_curvature" => {
251                let dim = get_usize("dim")?.unwrap_or(cols);
252                // κ defaults to 0 (flat initial point for the REML optimizer).
253                let kappa = get_f64("kappa")?
254                    .or_else(|| get_f64("curvature").ok().flatten())
255                    .unwrap_or(0.0);
256                Self::resolve("constant_curvature", None, None, Some(dim), Some(kappa))
257            }
258            other => Err(format!(
259                "response_geometry must be one of 'spd', 'grassmann(k=..)', 'stiefel(k=..)', \
260                 'poincare', 'constant_curvature', 'spherical', or 'simplex'; got {other:?}"
261            )),
262        }
263    }
264
265    /// Canonical, fully-specified label echoed back to the caller (mirrors the
266    /// way the sphere/simplex dispatch reports its resolved coordinate label).
267    pub fn canonical_label(&self) -> String {
268        match self {
269            Self::Spd { n } => format!("spd(n={n})"),
270            Self::Grassmann { k, n } => format!("grassmann(k={k},n={n})"),
271            Self::Stiefel { k, n } => format!("stiefel(k={k},n={n})"),
272            Self::Poincare { dim, curvature } => {
273                format!("poincare(dim={dim},curvature={curvature})")
274            }
275            Self::ConstantCurvature { dim, kappa } => {
276                format!("constant_curvature(dim={dim},kappa={kappa})")
277            }
278        }
279    }
280
281    /// Ambient (flattened) coordinate count: the column width of the `values`
282    /// matrix and the `base` vector.
283    pub fn ambient_dim(&self) -> usize {
284        match self {
285            Self::Spd { n } => n * n,
286            Self::Grassmann { k, n } | Self::Stiefel { k, n } => n * k,
287            Self::Poincare { dim, .. } | Self::ConstantCurvature { dim, .. } => *dim,
288        }
289    }
290
291    /// Build the underlying [`RiemannianManifold`] for the matrix geometries.
292    /// `None` for Poincaré, whose primitives are free functions parameterised
293    /// by curvature rather than a trait object.
294    fn riemannian(&self) -> Option<Box<dyn RiemannianManifold>> {
295        match self {
296            Self::Spd { n } => Some(Box::new(SpdManifold::new(*n))),
297            Self::Grassmann { k, n } => GrassmannManifold::new(*k, *n)
298                .ok()
299                .map(|m| Box::new(m) as _),
300            Self::Stiefel { k, n } => StiefelManifold::new(*k, *n).ok().map(|m| Box::new(m) as _),
301            Self::ConstantCurvature { dim, kappa } => {
302                Some(Box::new(ConstantCurvature::new(*dim, *kappa)))
303            }
304            Self::Poincare { .. } => None,
305        }
306    }
307
308    /// Per-point logarithm `log_base(value)` in flat ambient coordinates.
309    fn log_point(
310        &self,
311        base: ArrayView1<'_, f64>,
312        value: ArrayView1<'_, f64>,
313    ) -> GeometryResult<Array1<f64>> {
314        match self {
315            Self::Poincare { curvature, .. } => {
316                crate::manifolds::poincare::log_map(base, value, *curvature)
317            }
318            // ConstantCurvature implements RiemannianManifold::log_map directly.
319            Self::ConstantCurvature { .. }
320            | Self::Spd { .. }
321            | Self::Grassmann { .. }
322            | Self::Stiefel { .. } => self
323                .riemannian()
324                .expect("riemannian response manifold")
325                .log_map(base, value),
326        }
327    }
328
329    /// Per-point exponential `exp_base(tangent)` in flat ambient coordinates.
330    fn exp_point(
331        &self,
332        base: ArrayView1<'_, f64>,
333        tangent: ArrayView1<'_, f64>,
334    ) -> GeometryResult<Array1<f64>> {
335        match self {
336            Self::Poincare { curvature, .. } => {
337                crate::manifolds::poincare::exp_map(base, tangent, *curvature)
338            }
339            Self::ConstantCurvature { .. }
340            | Self::Spd { .. }
341            | Self::Grassmann { .. }
342            | Self::Stiefel { .. } => self
343                .riemannian()
344                .expect("riemannian response manifold")
345                .exp_map(base, tangent),
346        }
347    }
348
349    /// Euclidean / Frobenius distance from an arbitrary ambient row to the
350    /// candidate response geometry, in flat ambient coordinates — the extrinsic
351    /// constraint-violation distance behind [`response_projection_residual`].
352    ///
353    /// Unlike [`log_point`](Self::log_point), which is gatekept to *genuine*
354    /// manifold points on both arguments, this accepts off-manifold `value`. The
355    /// distance is computed in closed form per geometry and is **well-defined for
356    /// every input** — there is no rank-deficiency error path, because the
357    /// distance to a set is defined even where the nearest point is not unique:
358    ///
359    /// * `Gr(k, n)` / `St(k, n)` — distance to the orthonormal-frame set,
360    ///   `√Σ_i (σ_i − 1)²` with `σ_i = √max(λ_i(YᵀY), 0)` the singular values of
361    ///   the `n × k` frame `Y`. Exact for every rank (`σ_i = 0` columns
362    ///   contribute `1` each). Grassmann and Stiefel coincide because this module
363    ///   represents Grassmann points by frames — it is a *representation*
364    ///   distance, not a subspace/principal-angle distance.
365    /// * SPD cone — distance to the *closed* PSD cone,
366    ///   `√(‖skew(A)‖_F² + Σ_{λ_i<0} λ_i²)` with `λ_i` the eigenvalues of the
367    ///   symmetric part `sym(A)`. This is the infimum distance to the open SPD
368    ///   cone; a zero distance means PSD, **not** strictly PD.
369    /// * Poincaré ball — distance to the *manifold* open ball of radius
370    ///   `R = 1/√(−c)`: `max(0, ‖x‖ − R)`. (This uses the true radius `R`, not
371    ///   the slightly smaller numerical safety radius used when projecting points
372    ///   for a fit, so interior points score exactly zero.)
373    /// * `ConstantCurvature` — distance to the chart *domain*: `0` for `κ ≥ 0`
374    ///   (chart is all of `ℝ^d`), else `max(0, ‖x‖ − 1/√(−κ))`. The curvature
375    ///   lives in the metric, not the domain, so this is a domain-admissibility
376    ///   check only and carries little curvature information.
377    fn manifold_residual(&self, value: ArrayView1<'_, f64>) -> GeometryResult<f64> {
378        match self {
379            Self::Poincare { curvature, .. } => ball_domain_residual(value, *curvature),
380            Self::ConstantCurvature { kappa, .. } => {
381                if *kappa >= 0.0 {
382                    Ok(0.0)
383                } else {
384                    ball_domain_residual(value, *kappa)
385                }
386            }
387            Self::Spd { n } => {
388                let mat = from_flat(value, *n, *n)?;
389                let symm = sym(&mat);
390                let psd = spectral_map_symmetric(&symm, |lam| Ok(lam.max(0.0)))?;
391                // Distance to the closed PSD cone, measured against the original
392                // (skew included) input so the skew-symmetric part is counted.
393                Ok(frobenius_distance(value, flatten(&psd).view()))
394            }
395            Self::Grassmann { k, n } | Self::Stiefel { k, n } => {
396                use gam_linalg::faer_ndarray::fast_atb;
397                let frame = from_flat(value, *n, *k)?;
398                let gram = fast_atb(&frame, &frame);
399                let (evals, _) = jacobi_symmetric(&gram)?;
400                let mut sq = 0.0_f64;
401                for &lam in evals.iter() {
402                    let sigma = lam.max(0.0).sqrt();
403                    let d = sigma - 1.0;
404                    sq += d * d;
405                }
406                Ok(sq.sqrt())
407            }
408        }
409    }
410
411    /// Squared metric norm `‖v‖²_base` of a tangent at `base`. Used by the
412    /// Karcher iteration's stationarity test. Poincaré uses the conformal
413    /// factor squared; the matrix manifolds and ConstantCurvature use the trait
414    /// metric tensor.
415    fn sq_metric_norm(
416        &self,
417        base: ArrayView1<'_, f64>,
418        v: ArrayView1<'_, f64>,
419    ) -> GeometryResult<f64> {
420        match self {
421            Self::Poincare { curvature, .. } => {
422                let lam = crate::manifolds::poincare::conformal_factor(base, *curvature)?;
423                Ok(lam * lam * v.iter().map(|x| x * x).sum::<f64>())
424            }
425            Self::ConstantCurvature { .. }
426            | Self::Spd { .. }
427            | Self::Grassmann { .. }
428            | Self::Stiefel { .. } => {
429                let g = self
430                    .riemannian()
431                    .expect("riemannian response manifold")
432                    .metric_tensor(base)?;
433                let gv = g.dot(&v);
434                Ok(v.dot(&gv).max(0.0))
435            }
436        }
437    }
438}
439
440/// Batched response-geometry logarithm: map every manifold-valued response row
441/// to its tangent coordinate at `base`. `values` is `(n_rows, ambient)`, `base`
442/// is `(ambient,)`, and the returned tangent is `(n_rows, ambient)` (the same
443/// flat ambient layout — the tangent of a matrix manifold is itself a flattened
444/// matrix). The scalar Gaussian GAMs the caller fits operate column-wise on
445/// this matrix exactly as they do for the sphere.
446pub fn response_log_map(
447    manifold: ResponseManifold,
448    values: ArrayView2<'_, f64>,
449    base: ArrayView1<'_, f64>,
450) -> Result<Array2<f64>, String> {
451    let ambient = manifold.ambient_dim();
452    let (n_rows, cols) = values.dim();
453    if base.len() != ambient {
454        return Err(format!(
455            "response geometry base point has length {}; expected {ambient}",
456            base.len()
457        ));
458    }
459    if cols != ambient {
460        return Err(format!(
461            "response geometry values have {cols} columns; expected {ambient}"
462        ));
463    }
464    let mut out = Array2::<f64>::zeros((n_rows, ambient));
465    for row in 0..n_rows {
466        let tangent = manifold
467            .log_point(base, values.row(row))
468            .map_err(|e| format!("response geometry log map (row {row}): {e}"))?;
469        out.row_mut(row).assign(&tangent);
470    }
471    Ok(out)
472}
473
474/// Batched response-geometry exponential: map predicted tangent coordinates
475/// back to manifold-valued responses at `base`. Inverse of [`response_log_map`]
476/// with the same shapes.
477pub fn response_exp_map(
478    manifold: ResponseManifold,
479    tangent: ArrayView2<'_, f64>,
480    base: ArrayView1<'_, f64>,
481) -> Result<Array2<f64>, String> {
482    let ambient = manifold.ambient_dim();
483    let (n_rows, cols) = tangent.dim();
484    if base.len() != ambient {
485        return Err(format!(
486            "response geometry base point has length {}; expected {ambient}",
487            base.len()
488        ));
489    }
490    if cols != ambient {
491        return Err(format!(
492            "response geometry tangent has {cols} columns; expected {ambient}"
493        ));
494    }
495    if !tangent.iter().all(|v| v.is_finite()) {
496        return Err("response geometry tangent must contain only finite values".to_string());
497    }
498    let mut out = Array2::<f64>::zeros((n_rows, ambient));
499    for row in 0..n_rows {
500        let value = manifold
501            .exp_point(base, tangent.row(row))
502            .map_err(|e| format!("response geometry exp map (row {row}): {e}"))?;
503        out.row_mut(row).assign(&value);
504    }
505    Ok(out)
506}
507
508/// Numerically-stable Euclidean norm `‖v‖₂`, scaled by the largest-magnitude
509/// entry so the squared sum cannot overflow for large but finite inputs.
510fn scaled_l2_norm(v: ArrayView1<'_, f64>) -> f64 {
511    let mut scale = 0.0_f64;
512    for &x in v.iter() {
513        let a = x.abs();
514        if a > scale {
515            scale = a;
516        }
517    }
518    if scale == 0.0 {
519        return 0.0;
520    }
521    let mut ssq = 0.0_f64;
522    for &x in v.iter() {
523        let t = x / scale;
524        ssq += t * t;
525    }
526    scale * ssq.sqrt()
527}
528
529/// Numerically-stable Frobenius distance `‖a − b‖₂` over equal-length flat
530/// vectors, scaled by the largest entrywise difference to avoid overflow.
531fn frobenius_distance(a: ArrayView1<'_, f64>, b: ArrayView1<'_, f64>) -> f64 {
532    let mut scale = 0.0_f64;
533    for (x, y) in a.iter().zip(b.iter()) {
534        let d = (x - y).abs();
535        if d > scale {
536            scale = d;
537        }
538    }
539    if scale == 0.0 {
540        return 0.0;
541    }
542    let mut ssq = 0.0_f64;
543    for (x, y) in a.iter().zip(b.iter()) {
544        let t = (x - y) / scale;
545        ssq += t * t;
546    }
547    scale * ssq.sqrt()
548}
549
550/// Distance from `value` to the open ball of radius `R = 1/√(−c)` (`c < 0`):
551/// `max(0, ‖value‖ − R)`, the true Euclidean infimum distance to the ball.
552/// Errors if the curvature is not a finite negative number.
553fn ball_domain_residual(value: ArrayView1<'_, f64>, curvature: f64) -> GeometryResult<f64> {
554    if !curvature.is_finite() || curvature >= 0.0 {
555        return Err(GeometryError::InvalidPoint(
556            "ball distance requires a finite negative curvature",
557        ));
558    }
559    let radius = (-curvature).sqrt().recip();
560    Ok((scaled_l2_norm(value) - radius).max(0.0))
561}
562
563/// Per-row extrinsic distance from ambient observations to a *candidate*
564/// response geometry — a coordinate-dependent constraint / closure-distance
565/// diagnostic.
566///
567/// What this is (and is not)
568/// -------------------------
569/// This is a cheap, pre-fit **constraint-violation** measure: given a candidate
570/// response geometry, how far does each raw row sit from that geometry's
571/// extrinsic representation (the unit-norm frame, the PSD cone, the Poincaré
572/// ball)? It is **not** the post-fit on/off-manifold membership signal (which
573/// comes from a fitted geometric smooth's residual and posterior predictive
574/// density), and it is **not** a universal cross-geometry model-selection score:
575/// it measures extrinsic constraint violation *in a chosen coordinate chart*,
576/// not intrinsic topology or curvature. Different candidate geometries have
577/// different chart codimensions (a full-dimensional Poincaré/`κ ≥ 0` chart can
578/// score zero trivially), so residuals are not directly comparable across
579/// candidates without a noise model and per-candidate calibration. Use it as a
580/// fast per-candidate gate, with candidate-specific thresholds.
581///
582/// What it computes
583/// ----------------
584/// For each ambient row `x`, [`manifold_residual`](Self::manifold_residual)
585/// returns the closed-form distance to the candidate geometry (well-defined for
586/// every input and every rank — see that method for the per-geometry formulas),
587/// and this returns:
588///
589/// * `residual[i]` — the absolute distance-to-geometry (zero for genuinely
590///   admissible rows; for the matrix manifolds, exact to machine precision).
591/// * `relative[i] = residual[i] / (‖x‖ + eps)` — the distance normalised by the
592///   row's ambient magnitude. **Note:** this is dimensionless but *not*
593///   scale-invariant for the fixed-radius geometries (Stiefel/Grassmann/ball)
594///   and is *not* bounded by `1` (it diverges as `‖x‖ → 0`); it is scale-free
595///   only for the homogeneous SPD cone. Treat it as `input_norm_relative`, not
596///   an off-manifold fraction.
597///
598/// Unlike [`response_log_map`], **no base point is needed**. `values` is
599/// `(n_rows, ambient)`; both returned arrays are `(n_rows,)`. Every fittable
600/// response geometry — including `ConstantCurvature` — has a closed-form
601/// distance, so no variant errors on a valid, finite input.
602pub fn response_projection_residual(
603    manifold: ResponseManifold,
604    values: ArrayView2<'_, f64>,
605) -> Result<(Array1<f64>, Array1<f64>), String> {
606    let ambient = manifold.ambient_dim();
607    let (n_rows, cols) = values.dim();
608    if cols != ambient {
609        return Err(format!(
610            "response geometry values have {cols} columns; expected {ambient}"
611        ));
612    }
613    if !values.iter().all(|v| v.is_finite()) {
614        return Err("response geometry values must contain only finite values".to_string());
615    }
616
617    let mut residual = Array1::<f64>::zeros(n_rows);
618    let mut relative = Array1::<f64>::zeros(n_rows);
619    for row in 0..n_rows {
620        let value = values.row(row);
621        let dist = manifold
622            .manifold_residual(value)
623            .map_err(|e| format!("response geometry residual (row {row}): {e}"))?;
624        let rel = dist / (scaled_l2_norm(value) + GEOMETRY_EPS);
625        if !dist.is_finite() || !rel.is_finite() {
626            return Err(format!(
627                "response geometry residual (row {row}) is non-finite"
628            ));
629        }
630        residual[row] = dist;
631        relative[row] = rel;
632    }
633    Ok((residual, relative))
634}
635
636/// String-driven response-geometry log map: parse the user `label` (with shape
637/// inference from the response column count), pick the base point (intrinsic
638/// Fréchet mean when `base` is `None`), map every row to its tangent, and report
639/// the canonical resolved label. This is the curved-manifold analogue of the
640/// sphere/simplex dispatch and the single entry the FFI calls for these
641/// geometries.
642pub fn dispatch_log_map(
643    values: ArrayView2<'_, f64>,
644    label: &str,
645    base: Option<ArrayView1<'_, f64>>,
646) -> Result<(Array2<f64>, Array1<f64>, String), String> {
647    let manifold = ResponseManifold::parse(label, values.ncols())?;
648    let base_point = match base {
649        Some(b) => b.to_owned(),
650        None => response_frechet_mean(manifold, values, None, 1.0e-12, 256)?,
651    };
652    let tangent = response_log_map(manifold, values, base_point.view())?;
653    Ok((tangent, base_point, manifold.canonical_label()))
654}
655
656/// String-driven response-geometry exponential map: inverse of
657/// [`dispatch_log_map`] given an explicit base point.
658pub fn dispatch_exp_map(
659    tangent: ArrayView2<'_, f64>,
660    label: &str,
661    base: ArrayView1<'_, f64>,
662) -> Result<Array2<f64>, String> {
663    let manifold = ResponseManifold::parse(label, tangent.ncols())?;
664    response_exp_map(manifold, tangent, base)
665}
666
667/// Intrinsic (Karcher) Fréchet mean of manifold-valued responses, the default
668/// base point when the user supplies none. `values` is `(n_rows, ambient)`.
669///
670/// This is the SPD safeguarded Karcher iteration generalised over an arbitrary
671/// [`ResponseManifold`]: a Riemannian gradient-descent on the weighted
672/// dispersion `V(P) = Σ_i w_i ‖log_P(X_i)‖²_P` with the descent direction
673/// `ξ = Σ_i w_i log_P(X_i)` (`= −½ grad V`), a unit Karcher step `exp_P(t·ξ)`
674/// with Armijo backtracking plus a round-off cushion, a best-iterate stall
675/// guard, and the metric-norm stationarity test `‖ξ‖_P ≤ tol`. The SPD-specific
676/// version in [`crate::manifolds::spd::spd_frechet_mean`] remains for the affine
677/// inverse it caches per step; this generic form pays a metric-tensor solve but
678/// covers all four geometries uniformly.
679pub fn response_frechet_mean(
680    manifold: ResponseManifold,
681    values: ArrayView2<'_, f64>,
682    weights: Option<ArrayView1<'_, f64>>,
683    tol: f64,
684    max_iter: usize,
685) -> Result<Array1<f64>, String> {
686    let ambient = manifold.ambient_dim();
687    let (m, cols) = values.dim();
688    if m == 0 || cols != ambient {
689        return Err(format!(
690            "response geometry Fréchet mean: values must be M×{ambient} with M >= 1"
691        ));
692    }
693    if !(tol.is_finite() && tol > 0.0) {
694        return Err("response geometry Fréchet mean tolerance must be finite and positive".into());
695    }
696    let w = crate::normalize_weights(m, weights)
697        .map_err(|_| "response geometry Fréchet mean: invalid weights".to_string())?;
698    let samples: Vec<Array1<f64>> = (0..m).map(|i| values.row(i).to_owned()).collect();
699
700    let dispersion = |p: ArrayView1<'_, f64>| -> Result<f64, String> {
701        let mut acc = 0.0_f64;
702        for (i, x) in samples.iter().enumerate() {
703            let lg = manifold
704                .log_point(p, x.view())
705                .map_err(|e| format!("response geometry Fréchet mean log map: {e}"))?;
706            let sq = manifold
707                .sq_metric_norm(p, lg.view())
708                .map_err(|e| format!("response geometry Fréchet mean metric: {e}"))?;
709            acc += w[i] * sq;
710        }
711        Ok(acc)
712    };
713
714    // Seed the Karcher iteration at a sample whose tangent star is fully
715    // defined, then take one Riemannian averaging step for an interior start.
716    //
717    // A fixed seed at `samples[0]` is fragile: if any *other* sample lies at
718    // that seed's cut locus the seeding log is undefined and the whole mean
719    // aborts, even though the Fréchet mean itself is well defined. On
720    // `Gr(1,n) = ℝP^{n-1}` two orthogonal lines (principal angle π/2) are
721    // exactly such a cut-locus pair, so a design whose first response happens
722    // to be orthogonal to another could never be averaged. Instead, try each
723    // sample as the seed and keep the first whose log-tangents to *every*
724    // sample land: the safeguarded descent below converges to the same mean
725    // from any admissible seed, so this only changes which interior point the
726    // iteration starts from — and the very first sample is chosen whenever it
727    // is admissible (so SPD/Stiefel/Poincaré data with no cut-locus pair seed
728    // exactly as before). A design where every sample sits at another's cut
729    // locus has a genuinely ambiguous mean and is reported as such.
730    let mut seeded: Option<Array1<f64>> = None;
731    let mut last_seed_err = String::new();
732    for seed in &samples {
733        let base = match manifold.exp_point(seed.view(), Array1::<f64>::zeros(ambient).view()) {
734            Ok(base) => base,
735            Err(e) => {
736                last_seed_err = e.to_string();
737                continue;
738            }
739        };
740        let mut xi = Array1::<f64>::zeros(ambient);
741        let mut admissible = true;
742        for (i, x) in samples.iter().enumerate() {
743            match manifold.log_point(base.view(), x.view()) {
744                Ok(lg) => xi.scaled_add(w[i], &lg),
745                Err(e) => {
746                    last_seed_err = e.to_string();
747                    admissible = false;
748                    break;
749                }
750            }
751        }
752        if !admissible {
753            continue;
754        }
755        match manifold.exp_point(base.view(), xi.view()) {
756            Ok(stepped) => {
757                seeded = Some(stepped);
758                break;
759            }
760            Err(e) => {
761                last_seed_err = e.to_string();
762            }
763        }
764    }
765    let mut p = seeded.ok_or_else(|| {
766        format!(
767            "response geometry Fréchet mean init: no admissible seed among samples \
768             (every sample lies at another's cut locus; last error: {last_seed_err})"
769        )
770    })?;
771
772    let mut f_cur = dispersion(p.view())?;
773    let mut best_p = p.clone();
774    let mut best_grad = f64::INFINITY;
775    const STALL_REL: f64 = 5.0e-3;
776    const STALL_PATIENCE: usize = 10;
777    let mut stall = 0_usize;
778    const ARMIJO_C1: f64 = 1.0e-4;
779    const MAX_BACKTRACK_HALVINGS: usize = 60;
780    const ARMIJO_ROUNDOFF_EPS_MULTIPLE: f64 = 8.0;
781
782    for _ in 0..max_iter {
783        let mut xi = Array1::<f64>::zeros(ambient);
784        for (i, x) in samples.iter().enumerate() {
785            let lg = manifold
786                .log_point(p.view(), x.view())
787                .map_err(|e| format!("response geometry Fréchet mean log map: {e}"))?;
788            xi.scaled_add(w[i], &lg);
789        }
790        let grad_norm = manifold
791            .sq_metric_norm(p.view(), xi.view())
792            .map_err(|e| format!("response geometry Fréchet mean metric: {e}"))?
793            .sqrt();
794        if grad_norm <= tol {
795            return Ok(p);
796        }
797
798        let improved = grad_norm < best_grad * (1.0 - STALL_REL);
799        if grad_norm < best_grad {
800            best_grad = grad_norm;
801            best_p.assign(&p);
802        }
803        if improved {
804            stall = 0;
805        } else {
806            stall += 1;
807            if stall >= STALL_PATIENCE {
808                return Ok(best_p);
809            }
810        }
811
812        let pred = grad_norm * grad_norm;
813        let f_tol = ARMIJO_ROUNDOFF_EPS_MULTIPLE * f64::EPSILON * (1.0 + f_cur.abs());
814        let mut t = 1.0_f64;
815        let mut accepted = false;
816        for _ in 0..MAX_BACKTRACK_HALVINGS {
817            let step = &xi * t;
818            let cand = match manifold.exp_point(p.view(), step.view()) {
819                Ok(c) => c,
820                Err(_) => {
821                    // The step left the manifold's domain (e.g. a Poincaré
822                    // overshoot past the ball boundary); shrink and retry.
823                    t *= 0.5;
824                    continue;
825                }
826            };
827            let f_cand = match dispersion(cand.view()) {
828                Ok(f) => f,
829                Err(_) => {
830                    t *= 0.5;
831                    continue;
832                }
833            };
834            if f_cand <= f_cur - 2.0 * ARMIJO_C1 * t * pred + f_tol {
835                p = cand;
836                f_cur = f_cand;
837                accepted = true;
838                break;
839            }
840            t *= 0.5;
841        }
842        if !accepted {
843            return Ok(best_p);
844        }
845    }
846    Err("response geometry Fréchet mean did not reach stationarity within max_iter".into())
847}
848
849// ── Curvature as an estimand on the response geometry (#944 stage 4 / #1104) ──
850//
851// `response_geometry="constant_curvature(dim=d)"` does NOT take a fixed κ from
852// the user: κ is ESTIMATED from the manifold-valued responses. At each κ the
853// family `ConstantCurvature{dim, κ}` is laid down and κ is scored by the HONEST
854// change-of-variables likelihood of the observed chart coordinates `yᵢ` w.r.t.
855// ambient Lebesgue measure `dy` — the density that is automatically normalised on
856// the SAME measure in which the data are observed, regardless of how the manifold
857// is parameterised. This is the crux of the #1104 fix.
858//
859// ## Why dispersion alone (and the self-normalising wrapped Gaussian) is degenerate
860//
861// The generative model is the wrapped normal `yᵢ = exp_μ(vᵢ)`, `vᵢ` isotropic at
862// geodesic scale σ. Its density w.r.t. the Riemannian volume `dvol_κ` is
863// `N(sᵢ;0,σ²)/Jᵧ_κ(sᵢ)` with `sᵢ = d_κ(μ,yᵢ)` the geodesic radius and
864// `J_κ(s) = (sn_κ(s)/s)^{d−1}` the exp-map volume Jacobian
865// (`ConstantCurvature::jacobian_radial`). The naive criterion
866// `½nd·ln(Σsᵢ²/nd)` (dispersion only), and even the full `dvol_κ`-density NLL
867// `Σ[sᵢ²/2σ² + (d/2)ln2πσ² + ln J_κ(sᵢ)]`, are SCALE-DEGENERATE: rescaling the
868// manifold radius `R = 1/√|κ|` rescales every `sᵢ` and every volume element, and
869// the σ-profile absorbs the change with no κ information left. That is exactly
870// why a `dvol_κ`-normalised (self-normalising) wrapped Gaussian rails, and why an
871// intrinsic-volume partition function double-counts: the density is already
872// normalised on `dvol_κ`, so re-integrating its volume adds nothing identifying.
873//
874// ## The restoring force is the ambient (chart) volume element at the DATA points
875//
876// Curvature is identified only when the abstract manifold is tied to the CONCRETE
877// observed chart coordinates `yᵢ`. The data are observed as points of `ℝ^d` under
878// Lebesgue `dy`, so the likelihood must be the density w.r.t. `dy`, obtained from
879// the `dvol_κ`-density by the chart volume factor `dvol_κ/dy = λ_{yᵢ}^d`,
880// `λ_y = 2/(1+κ‖y‖²)`:
881//
882// ```text
883//   −ℓ(κ,μ,σ²) = Σᵢ[ sᵢ²/(2σ²) + (d/2)ln(2πσ²) + ln J_κ(sᵢ) − d·ln λ_{yᵢ} ].
884// ```
885//
886// The new term `−d·Σ ln λ_{yᵢ} = d·Σ ln((1+κ‖yᵢ‖²)/2)` is evaluated at every DATA
887// point (not at the mean), so `‖yᵢ‖² > 0` even for mean-centred clouds and it
888// supplies a genuine κ-restoring force: it grows like `+d·κ·Σ‖yᵢ‖²` for small κ
889// and `→ +∞` as κ→+∞ (each `−ln λ_{yᵢ}→+∞`), exactly opposing the dispersion /
890// `ln J_κ` terms which fall as the sphere shrinks. The minimum is therefore
891// INTERIOR at the data-generating curvature. None of `ln J_κ` or `λ` depend on σ,
892// so σ profiles in closed form `σ̂² = D/(nd)`, `D = Σ sᵢ²`.
893//
894// ## Reparameterisation invariance / unit-covariance of κ̂
895//
896// κ carries units of `1/length²`. Under a global rescaling `yᵢ ↦ α·yᵢ` the chart
897// of `M_κ` at scale `α` equals the chart of `M_{κ/α²}` at scale 1 (because
898// `λ` and every geodesic primitive depend on `y` only through `κ‖y‖²`). The whole
899// criterion `V(κ, αy)` therefore equals `V(α²κ, y)`, so its minimiser transforms
900// as `κ̂(αy) = κ̂(y)/α²` — the CORRECT covariance of a curvature with units
901// `1/length²`. The base point μ is held at the κ-independent flat centroid (NOT
902// re-solved per κ): re-solving the Fréchet mean per κ is precisely what
903// re-entangles κ with the chart scale and biases the estimate, so it is removed.
904//
905// `V_p` is a negative log-evidence (lower is better) so κ̂ = argmin V_p; it is the
906// full NLL summed over all `n·d` scalar observations, so `2[V_p(0) − V_p(κ̂)]` is
907// the Wilks LR statistic with a calibrated χ²₁ flatness reference — exactly the
908// contract `profile_ci_walk` / `flatness_lr_test` in `curvature_estimand.rs`
909// consume, with no new outer machinery.
910
911/// Outcome of fitting curvature as an estimand on a constant-curvature response
912/// geometry: the optimised κ̂, its tangent base point, the profile-likelihood CI,
913/// and the interior-point flatness (Wilks) test of κ = 0.
914#[derive(Clone, Debug)]
915pub struct ResponseCurvatureFit {
916    /// The dimension `d` of the constant-curvature response manifold.
917    pub dim: usize,
918    /// The REML/evidence-optimal curvature κ̂ (argmin of the profiled criterion).
919    ///
920    /// **Units `1/length²`** — κ̂ is therefore *scale-dependent*: rescaling the
921    /// cloud `y ↦ α·y` rescales `κ̂ ↦ κ̂/α²`. For a scale-free statement of how
922    /// curved the cloud is, read [`kappa_r2`](Self::kappa_r2) instead. When the
923    /// cloud is curved BEYOND what its spread can resolve (it fills a large
924    /// fraction of the sphere `S^d(1/√κ̂)`), the optimiser rails to the
925    /// chart-resolution cap and [`railed_at_resolution_limit`](Self::railed_at_resolution_limit)
926    /// is `true`: κ̂ is then a *lower bound on |κ|*, not a point estimate.
927    pub kappa_hat: f64,
928    /// The DIMENSIONLESS geometric invariant the cloud actually determines:
929    /// `κ̂ · r²` with `r` = [`characteristic_radius`](Self::characteristic_radius).
930    /// This is scale-FREE (`κ̂·r²` is invariant under `y ↦ α·y`, since `κ̂ ↦ κ̂/α²`
931    /// and `r ↦ α·r`) — the honest answer to "how curved is this cloud relative
932    /// to its own spread". `|κ̂·r²| ≪ 1` ⇒ nearly flat at this scale; `κ̂·r² ↗ (π/2)²`
933    /// ⇒ the cloud fills the sphere and curvature is at the chart-resolution limit.
934    pub kappa_r2: f64,
935    /// Characteristic geodesic radius `r` of the cloud at κ = 0 (the doubled-gauge
936    /// chart distance `r = 2·max_i‖y_i − μ‖`): the length scale against which κ̂ is
937    /// dimensionless. Reported so the caller can convert between scale-dependent κ̂
938    /// and the scale-free `κ̂·r²` without re-deriving the chart gauge.
939    pub characteristic_radius: f64,
940    /// The intrinsic Fréchet-mean base point at κ̂ (the tangent expansion point
941    /// the scalar GAMs are fitted around).
942    pub base: Array1<f64>,
943    /// Profiled criterion value `V_p(κ̂)` (concentrated negative log-evidence).
944    pub v_p_hat: f64,
945    /// `true` when the κ̂ search converged ONTO the chart-resolution cap rather
946    /// than an interior optimum: the data want curvature at or beyond the
947    /// conjugate radius of their geodesic spread (the cloud fills the sphere).
948    /// In that case κ̂ / the CI upper end are NOT a resolved point estimate but a
949    /// HONEST "curvature exceeds chart-resolvable range at this scale" flag — the
950    /// caller must report it as such and never as a silent `κ̂ = ci_hi`. The
951    /// hyperbolic side cannot rail this way (κ < 0 has no conjugate radius), so a
952    /// rail here always means strongly spherical relative to the spread.
953    pub railed_at_resolution_limit: bool,
954    /// `true` only when the SIGN of κ̂ is statistically resolved — i.e. the
955    /// profile-likelihood CI excludes 0 (`profile_ci.verdict ≠ Flat`).
956    ///
957    /// ## Why a point estimate alone is not enough (the #944/#1059 flat-floor)
958    ///
959    /// Curvature is resolvable only through the dimensionless product `κ·r²`
960    /// (see [`kappa_r2`](Self::kappa_r2)); the per-point Fisher information for κ
961    /// scales like `σ⁴`. When the cloud is nearly flat at its own scale
962    /// (`|κ·r²| ≪ 1`), the profiled criterion is so shallow that its single-cloud
963    /// argmin κ̂ can land on the WRONG SIDE OF ZERO purely by Monte-Carlo
964    /// fluctuation — empirically a coin-flip below `|κ·r²| ≈ 0.03`, reliable above
965    /// `≈ 0.09` (the #944 power curve). The estimand itself is UNBIASED (the
966    /// criterion averaged over clouds minimises exactly at κ⋆), so this is a
967    /// resolution limit, not a bias.
968    ///
969    /// The CI, in contrast, is honest in this regime: at an under-resolved
970    /// operating point it reports `Flat` (straddles 0) rather than a confident
971    /// wrong sign — it essentially never claims the wrong-signed geometry. So the
972    /// SIGN-bearing summary the caller may quote is the CI verdict, not the bare
973    /// κ̂. This flag exposes that contract on the point-estimate surface: when it
974    /// is `false`, κ̂'s sign is noise — the caller must report "curvature not
975    /// resolved at this scale (|κ·r²| too small)" and quote the CI / `kappa_r2`,
976    /// never a sign-confident κ̂. It is the flat-floor twin of
977    /// [`railed_at_resolution_limit`](Self::railed_at_resolution_limit) (the
978    /// spherical-cap rail); together they bracket the two ends of the resolvable
979    /// `κ·r²` band where κ̂ is a genuine interior point estimate.
980    pub sign_resolved: bool,
981    /// Profile-likelihood CI for κ and the geometry verdict from its sign.
982    pub profile_ci: crate::curvature_estimand::KappaProfileCi,
983    /// Interior-point χ²₁ likelihood-ratio test of flatness (κ = 0).
984    pub flatness: crate::curvature_estimand::FlatnessTest,
985}
986
987/// Chart-validity bounds on κ for a constant-curvature response geometry built
988/// from the supplied responses, plus the characteristic geodesic radius
989/// `ρ_max = 2·max_i‖y_i − μ‖` against which κ is made dimensionless.
990///
991/// Returns `(kappa_min, kappa_max, rho_max)`.
992///
993/// * **Lower (hyperbolic) bound.** The κ-stereographic chart requires
994///   `1 + κ‖x‖² > 0` at every point measured from the chart origin, i.e.
995///   `κ > −1/R²` with `R² = max_i ‖y_i‖²`. With a safety margin: `−0.999/R²`.
996/// * **Upper (spherical) bound.** Unlike the hyperbolic side this is NOT
997///   unbounded: on a sphere of curvature κ the geodesic radius cannot exceed the
998///   conjugate radius `π/√κ`, beyond which the exp-map volume Jacobian
999///   `J_κ = (sn_κ/·)^{d−1}` changes sign (clamped to 0 here) and `ln J_κ` would
1000///   collapse `V_p` toward `−∞`, railing the optimiser onto a spurious shell.
1001///   The κ = 0 geodesic radius of the farthest point from the centroid is
1002///   `ρ_max = 2·max_i‖y_i − μ‖` (doubled-gauge chart). We cap κ so that radius
1003///   stays strictly inside the first conjugate shell with a 10% margin:
1004///   `√κ·ρ_max ≤ 0.9π ⇒ κ_max = (0.9π / ρ_max)²`. This keeps every geodesic
1005///   radius before the antipodal singularity along the whole search/CI walk.
1006///
1007/// `κ_max` is the chart-RESOLUTION limit of the cloud: at it the geodesic spread
1008/// fills `(0.9π)² ≈ (π/2·1.8)²` of the conjugate shell, i.e. the cloud nearly
1009/// fills the sphere `S^d(1/√κ_max)`. The DIMENSIONLESS product `κ_max·ρ_max²
1010/// = (0.9π)²` is fixed and data-scale-free — it is the natural "the cloud is
1011/// maximally curved relative to its spread" sentinel the rail check compares κ̂ to.
1012fn response_kappa_bounds(values: ArrayView2<'_, f64>) -> (f64, f64, f64) {
1013    let (n_rows, dim) = values.dim();
1014    // ‖y_i‖² from the chart origin (governs the λ / hyperbolic-chart constraint).
1015    let mut r2_max = 0.0_f64;
1016    for row in values.outer_iter() {
1017        let r2 = row.dot(&row);
1018        if r2 > r2_max {
1019            r2_max = r2;
1020        }
1021    }
1022    // ‖y_i − μ‖² from the centroid (governs the spherical conjugate-radius cap).
1023    let mut centroid = Array1::<f64>::zeros(dim.max(1));
1024    if n_rows > 0 && dim > 0 {
1025        for row in values.outer_iter() {
1026            centroid += &row;
1027        }
1028        centroid.mapv_inplace(|v| v / n_rows as f64);
1029    }
1030    let mut s2_max = 0.0_f64;
1031    if dim > 0 {
1032        for row in values.outer_iter() {
1033            let diff = &row - &centroid;
1034            let r2 = diff.dot(&diff);
1035            if r2 > s2_max {
1036                s2_max = r2;
1037            }
1038        }
1039    }
1040    if r2_max <= 0.0 && s2_max <= 0.0 {
1041        // Degenerate (all points at the origin): κ is unidentified; use a wide
1042        // symmetric default so the optimiser/CI report a flat, unbounded result.
1043        return (-1.0e6, 1.0e6, 0.0);
1044    }
1045    // Keep a safety margin off the singular hyperbolic boundary.
1046    let kappa_min = if r2_max > 0.0 {
1047        -0.999 / r2_max
1048    } else {
1049        -1.0e6
1050    };
1051    // Conjugate-radius cap: ρ_max = 2·max‖y_i − μ‖ is the κ=0 geodesic radius.
1052    let rho_max = 2.0 * s2_max.sqrt();
1053    let kappa_max = if s2_max > 0.0 {
1054        let edge = 0.9 * std::f64::consts::PI / rho_max;
1055        edge * edge
1056    } else {
1057        1.0e6
1058    };
1059    (kappa_min, kappa_max, rho_max)
1060}
1061
1062/// Profiled curvature criterion `V_p(κ)` for the constant-curvature response
1063/// geometry: the σ-profiled HONEST change-of-variables negative log-likelihood of
1064/// the observed chart coordinates `y_i` at curvature `κ`, expressed w.r.t. ambient
1065/// Lebesgue measure `dy`. Lower is better (κ̂ = argmin). Returns `(V_p, base)`;
1066/// the base point is the κ-INDEPENDENT flat centroid (the tangent expansion point
1067/// that the scalar GAMs are fitted around), held fixed across κ so the estimate is
1068/// not re-entangled with the chart scale.
1069///
1070/// The model is the wrapped normal `y_i = exp_{μ,κ}(v_i)` with isotropic geodesic
1071/// scale σ; `s_i = d_κ(μ, y_i)` is the geodesic radius and `J_κ(s)` the exp-map
1072/// volume Jacobian. The density on the Riemannian volume `dvol_κ` is
1073/// `N(s_i;0,σ²)/J_κ(s_i)`; converting to ambient `dy` multiplies by the chart
1074/// volume factor `λ_{y_i}^d`, `λ_y = 2/(1+κ‖y‖²)`. The negative log-likelihood is
1075///
1076/// ```text
1077///   −ℓ(κ,σ²) = Σ_i[ s_i²/(2σ²) + (d/2)ln(2πσ²) + ln J_κ(s_i) − d·ln λ_{y_i} ].
1078/// ```
1079///
1080/// `ln J_κ` and `λ` do not depend on σ, so σ profiles in closed form
1081/// `σ̂² = D/(nd)`, `D = Σ s_i²`. The `−d·Σ ln λ_{y_i}` term — evaluated at the DATA
1082/// points, not the mean — is the κ-restoring force that breaks the scale
1083/// degeneracy of the dispersion / `dvol_κ`-density alone (see the module notes).
1084/// Additive constants independent of κ are kept implicit; they cancel in every
1085/// LR / profile-drop the CI machinery forms. μ is the closed-form flat centroid,
1086/// so the criterion is a pure function of κ with no inner tolerance/iteration
1087/// budget (the outer κ̂ search owns those).
1088pub fn response_curvature_criterion(
1089    values: ArrayView2<'_, f64>,
1090    dim: usize,
1091    kappa: f64,
1092) -> Result<(f64, Array1<f64>), String> {
1093    if !kappa.is_finite() {
1094        return Err("response curvature criterion: kappa must be finite".into());
1095    }
1096    let (n_rows, cols) = values.dim();
1097    if n_rows == 0 || cols != dim || dim == 0 {
1098        return Err(format!(
1099            "response curvature criterion: values must be N×{dim} with N >= 1"
1100        ));
1101    }
1102    // κ-independent base point: the flat (ambient) centroid. Holding μ fixed across
1103    // κ is the de-entangling move — re-solving the Fréchet mean per κ couples the
1104    // base to the chart scale and biases κ̂ (#1104 root cause).
1105    let mut base = Array1::<f64>::zeros(dim);
1106    for row in values.outer_iter() {
1107        base += &row;
1108    }
1109    base.mapv_inplace(|v| v / n_rows as f64);
1110
1111    let chart = ConstantCurvature::new(dim, kappa);
1112    // Reject κ at/over the chart boundary (1 + κ‖x‖² ≤ 0) at the centroid or any
1113    // data point: the geodesic primitives are undefined there. The bracket in
1114    // `response_kappa_bounds` keeps the optimiser strictly inside, but a CI/LR
1115    // probe can still land on the edge, so guard rather than panic.
1116    chart
1117        .conformal_factor(base.view())
1118        .map_err(|e| format!("response curvature criterion: base off chart: {e}"))?;
1119
1120    let d = dim as f64;
1121    let mut dispersion = 0.0_f64; // D = Σ s_i²
1122    let mut ln_jac = 0.0_f64; // Σ ln J_κ(s_i)
1123    let mut ln_lambda = 0.0_f64; // Σ ln λ_{y_i}
1124    // Geodesic radii s_i = d_κ(μ, y_i) for every row, computed in a single
1125    // batched pass (four rows per SIMD lane-group). `distance_batch` is
1126    // bit-for-bit identical to the per-row `distance`, so D, Σ ln J, and Σ ln λ
1127    // below are unchanged; it also validates each y_i is in-chart.
1128    let mut radii = vec![0.0_f64; n_rows];
1129    chart
1130        .distance_batch(base.view(), values, &mut radii)
1131        .map_err(|e| format!("response curvature criterion distance: {e}"))?;
1132    for (row, &s) in values.outer_iter().zip(radii.iter()) {
1133        dispersion += s * s;
1134        // ln J_κ(s_i): exp-map volume Jacobian (≥ 0); floor before the log so the
1135        // conjugate-shell clamp (J → 0 on the κ>0 antipodal shell) is a large
1136        // finite penalty rather than −∞.
1137        ln_jac += chart.jacobian_radial(s).max(1.0e-300).ln();
1138        // ln λ_{y_i} = ln(2) − ln(1 + κ‖y_i‖²); `conformal_factor` validates chart.
1139        let lam = chart
1140            .conformal_factor(row)
1141            .map_err(|e| format!("response curvature criterion conformal factor: {e}"))?;
1142        ln_lambda += lam.ln();
1143    }
1144    let nobs = (n_rows * dim) as f64;
1145    // Floor the dispersion so a (near-)perfect flat fit does not blow ln up; the
1146    // floor is far below any genuine residual scale and cancels in profile drops.
1147    let disp = dispersion.max(1.0e-300 * nobs.max(1.0));
1148
1149    // σ profiles in closed form: σ̂² = D/(nd). Substituting and dropping the
1150    // κ-independent constant (nd/2)(1 + ln 2π):
1151    //   V_p(κ) = (nd/2)·ln(D/(nd)) + Σ ln J_κ(s_i) − d·Σ ln λ_{y_i}.
1152    let v_p = 0.5 * nobs * (disp / nobs).ln() + ln_jac - d * ln_lambda;
1153    Ok((v_p, base))
1154}
1155
1156/// Fit curvature as an estimand on a constant-curvature response geometry.
1157///
1158/// κ̂ is the minimiser of the profiled criterion [`response_curvature_criterion`]
1159/// (the σ-profiled honest change-of-variables negative log-evidence of the wrapped
1160/// normal w.r.t. ambient measure), found by a golden-section search inside the
1161/// chart-validity bracket. The base point μ is the κ-independent flat centroid, so
1162/// every `V_p` evaluation scores the SAME geometry without re-entangling κ with the
1163/// chart scale (the #1104 fix). The exact outer
1164/// curvature `V_p''(κ̂)` is taken by a central second difference of the same
1165/// criterion and handed to [`profile_ci_walk`](crate::profile_ci_walk)
1166/// to size the initial Wald step; the CI itself is the exact χ²₁ profile crossing.
1167/// Flatness is the interior-point χ²₁ LR test
1168/// [`flatness_lr_test`](crate::flatness_lr_test). κ = 0 is an interior
1169/// point of the analytic `S^d ← ℝ^d → H^d` family, so no boundary correction is
1170/// applied. Returns the κ̂, its tangent base point, the profile CI, and the Wilks
1171/// flatness test for the fit summary.
1172///
1173/// ## Scale-awareness and honest railing (#1104)
1174///
1175/// κ has units `1/length²`, so a cloud of characteristic geodesic radius `r`
1176/// resolves only the DIMENSIONLESS product `κ·r²` (every chart primitive depends
1177/// on `y` through `κ‖y‖²`, hence `V(κ, αy) = V(α²κ, y)` and `κ̂ ↦ κ̂/α²` under
1178/// `y ↦ αy`). The fit therefore also returns:
1179/// * `kappa_r2 = κ̂·r²` — the scale-FREE invariant the cloud actually determines
1180///   (how curved relative to its own spread), and `characteristic_radius = r`;
1181/// * `railed_at_resolution_limit` — `true` when the data want curvature at or
1182///   beyond the conjugate radius of their spread (the cloud fills the sphere),
1183///   so the search converges onto the spherical cap. There κ̂ is a LOWER BOUND on
1184///   `|κ|`, not a resolved point estimate, and the caller must report "curvature
1185///   exceeds chart-resolvable range at this scale" rather than silently quoting
1186///   `κ̂ = ci_hi`. This is the #1104 fix: a tightly-concentrated near-spherical
1187///   cloud (e.g. unit-normalised OLMo activations) no longer SILENTLY rails to a
1188///   huge scale-dependent `ci_hi` while claiming a point estimate + CI.
1189pub fn fit_response_curvature(
1190    values: ArrayView2<'_, f64>,
1191    dim: usize,
1192    level: f64,
1193    tol: f64,
1194    max_iter: usize,
1195) -> Result<ResponseCurvatureFit, String> {
1196    if dim == 0 {
1197        return Err("constant-curvature response geometry requires dim >= 1".into());
1198    }
1199    let (n_rows, cols) = values.dim();
1200    if n_rows == 0 || cols != dim {
1201        return Err(format!(
1202            "constant-curvature response geometry: values must be N×{dim} with N >= 1"
1203        ));
1204    }
1205    if !(level > 0.0 && level < 1.0) {
1206        return Err("response curvature CI level must lie in (0, 1)".into());
1207    }
1208    let (kappa_min, kappa_max, rho_max) = response_kappa_bounds(values);
1209
1210    // `V_p` as a closure over the criterion; threaded through both the κ̂ search
1211    // and the CI walk. Every evaluation uses the same κ-independent flat-centroid
1212    // base, so the criterion is a clean 1-D function of κ.
1213    let mut v_p = |kappa: f64| -> Result<f64, String> {
1214        response_curvature_criterion(values, dim, kappa).map(|(v, _)| v)
1215    };
1216
1217    // ── κ̂: golden-section minimisation inside the chart bracket. ────────────
1218    // The dispersion criterion is smooth and unimodal in practice; golden
1219    // section is derivative-free and respects the bracket bounds exactly.
1220    const GOLDEN_INV: f64 = 0.618_033_988_749_894_8; // 1/φ
1221    let mut a = kappa_min;
1222    let mut b = kappa_max;
1223    let mut c = b - GOLDEN_INV * (b - a);
1224    let mut d_pt = a + GOLDEN_INV * (b - a);
1225    let mut fc = v_p(c)?;
1226    let mut fd = v_p(d_pt)?;
1227    let ktol = (tol * (kappa_max - kappa_min)).max(tol).max(1.0e-12);
1228    for _ in 0..max_iter {
1229        if (b - a).abs() <= ktol {
1230            break;
1231        }
1232        if fc < fd {
1233            b = d_pt;
1234            d_pt = c;
1235            fd = fc;
1236            c = b - GOLDEN_INV * (b - a);
1237            fc = v_p(c)?;
1238        } else {
1239            a = c;
1240            c = d_pt;
1241            fc = fd;
1242            d_pt = a + GOLDEN_INV * (b - a);
1243            fd = v_p(d_pt)?;
1244        }
1245    }
1246    let kappa_hat = 0.5 * (a + b);
1247    let (v_p_hat, base) = response_curvature_criterion(values, dim, kappa_hat)?;
1248
1249    // ── Honest chart-resolution-rail detection. ─────────────────────────────
1250    // The spherical cap κ_max is the curvature at which the cloud's geodesic
1251    // spread ρ_max fills `(0.9π)²` of the conjugate shell — i.e. the cloud nearly
1252    // fills the sphere S^d(1/√κ_max). When the criterion's optimum sits AT that
1253    // cap (the data want κ ≥ κ_max, but the chart cannot resolve a sphere smaller
1254    // than the cloud), the search converges onto the upper bracket and κ̂ ≈ κ_max
1255    // is NOT a resolved point estimate — it is a lower bound on |κ|. We flag this
1256    // so the caller reports "curvature exceeds chart-resolvable range at this
1257    // scale" instead of silently quoting κ̂ / ci_hi as if interior. The detection
1258    // is scale-free: it triggers when κ̂ lands within the final golden-section
1259    // resolution of κ_max (the dimensionless product κ̂·ρ_max² ↗ (0.9π)²), never
1260    // by an absolute κ threshold. The hyperbolic side has no conjugate radius, so
1261    // only the spherical (upper) cap can rail this way.
1262    let span = kappa_max - kappa_min;
1263    let rail_margin = (0.02 * span).max(ktol);
1264    let railed_at_resolution_limit = kappa_hat >= kappa_max - rail_margin;
1265
1266    // Dimensionless scale-free invariant κ̂·r²: the geometric content the cloud
1267    // actually determines (invariant under y ↦ αy). r = ρ_max is the κ=0 doubled-
1268    // gauge characteristic radius; for a degenerate (point) cloud r = 0 and the
1269    // product is 0 (κ unidentified). This is what the caller should report as the
1270    // honest "how curved relative to its spread" number alongside the dimensional κ̂.
1271    let kappa_r2 = kappa_hat * rho_max * rho_max;
1272
1273    // Exact outer curvature V_p''(κ̂) by a central second difference, on a step
1274    // scaled to the bracket; only used to size the Wald bracket of the CI walk.
1275    let h = (1.0e-3 * (kappa_max - kappa_min)).max(1.0e-6);
1276    let v_pp = if (kappa_hat - h) > kappa_min && (kappa_hat + h) < kappa_max {
1277        let vp = v_p(kappa_hat + h)?;
1278        let vm = v_p(kappa_hat - h)?;
1279        (vp - 2.0 * v_p_hat + vm) / (h * h)
1280    } else {
1281        // Near a bound: leave it to the walk's default step.
1282        f64::NAN
1283    };
1284
1285    let profile_ci = crate::curvature_estimand::profile_ci_walk(
1286        &mut v_p, kappa_hat, v_pp, kappa_min, kappa_max, level, ktol,
1287    )?;
1288    let flatness = crate::curvature_estimand::flatness_lr_test(&mut v_p, kappa_hat)?;
1289
1290    // The sign of κ̂ is statistically resolved iff the profile CI excludes 0 — the
1291    // CI is the honest sign-bearing summary (it reports Flat under-resolution rather
1292    // than a confident wrong sign), so we mirror its verdict onto the point-estimate
1293    // surface. Below the resolvable `κ·r²` floor (`|κ·r²| ≪ 1`) the bare κ̂ argmin can
1294    // flip sign on Monte-Carlo noise, so `false` here means "do not quote κ̂'s sign".
1295    let sign_resolved = !matches!(
1296        profile_ci.verdict,
1297        crate::curvature_estimand::CurvatureVerdict::Flat
1298    );
1299
1300    Ok(ResponseCurvatureFit {
1301        dim,
1302        kappa_hat,
1303        kappa_r2,
1304        characteristic_radius: rho_max,
1305        railed_at_resolution_limit,
1306        sign_resolved,
1307        base,
1308        v_p_hat,
1309        profile_ci,
1310        flatness,
1311    })
1312}
1313
1314#[cfg(test)]
1315mod tests {
1316    use super::*;
1317    use ndarray::{Array2, array};
1318
1319    fn round_trip(manifold: ResponseManifold, values: Array2<f64>) {
1320        let base =
1321            response_frechet_mean(manifold, values.view(), None, 1e-12, 500).expect("frechet mean");
1322        let tangent = response_log_map(manifold, values.view(), base.view()).expect("log map");
1323        let back = response_exp_map(manifold, tangent.view(), base.view()).expect("exp map");
1324        for row in 0..values.nrows() {
1325            for col in 0..values.ncols() {
1326                assert!(
1327                    (back[[row, col]] - values[[row, col]]).abs() < 1e-6,
1328                    "{manifold:?} exp∘log mismatch at ({row},{col}): {} vs {}",
1329                    back[[row, col]],
1330                    values[[row, col]]
1331                );
1332            }
1333        }
1334    }
1335
1336    #[test]
1337    fn spd_round_trip_and_mean() {
1338        // Three 2×2 SPD matrices, row-major flat.
1339        let values = array![
1340            [2.0, 0.0, 0.0, 1.0],
1341            [1.0, 0.3, 0.3, 2.0],
1342            [3.0, -0.5, -0.5, 1.5],
1343        ];
1344        round_trip(ResponseManifold::Spd { n: 2 }, values);
1345    }
1346
1347    #[test]
1348    fn grassmann_round_trip_and_mean() {
1349        // Gr(1, 3): unit columns (lines through the origin), n·k = 3 flat.
1350        let values = array![[1.0, 0.0, 0.0], [0.0, 1.0, 0.0], [0.6, 0.8, 0.0],];
1351        round_trip(ResponseManifold::Grassmann { k: 1, n: 3 }, values);
1352    }
1353
1354    #[test]
1355    fn stiefel_round_trip_and_mean() {
1356        // St(1, 3): unit 1-frames in ℝ³ (== sphere S²).
1357        let values = array![[1.0, 0.0, 0.0], [0.0, 1.0, 0.0], [0.0, 0.6, 0.8],];
1358        round_trip(ResponseManifold::Stiefel { k: 1, n: 3 }, values);
1359    }
1360
1361    #[test]
1362    fn stiefel_k2_round_trip_and_mean_n_lt_2k() {
1363        // St(3, 2): three orthonormal 2-frames in ℝ³ clustered near [e0, e1],
1364        // exercising the genuine canonical-metric logarithm (k ≥ 2) through the
1365        // full Karcher-mean → log → exp round trip. This is the n < 2k regime
1366        // (n = 3 < 2k = 4) where the economical 2k-block form is rank-deficient.
1367        // Before the k ≥ 2 Stiefel logarithm existed this aborted in
1368        // Fréchet-mean init with a misleading cut-locus error (#1637).
1369        let (c2, s2) = (0.2_f64.cos(), 0.2_f64.sin());
1370        let (c1, s1) = (0.15_f64.cos(), 0.15_f64.sin());
1371        let values = array![
1372            [1.0, 0.0, 0.0, 1.0, 0.0, 0.0],
1373            [c2, 0.0, 0.0, 1.0, s2, 0.0],
1374            [1.0, 0.0, 0.0, c1, 0.0, s1],
1375        ];
1376        round_trip(ResponseManifold::Stiefel { k: 2, n: 3 }, values);
1377    }
1378
1379    #[test]
1380    fn stiefel_k2_round_trip_and_mean_n_ge_2k() {
1381        // St(4, 2): the n ≥ 2k regime (n = 4 = 2k), clustered 2-frames in ℝ⁴.
1382        let (c0, s0) = (0.1_f64.cos(), 0.1_f64.sin());
1383        let (c1, s1) = (0.12_f64.cos(), 0.12_f64.sin());
1384        let values = array![
1385            [1.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0],
1386            [c0, 0.0, 0.0, 1.0, s0, 0.0, 0.0, 0.0],
1387            [1.0, 0.0, 0.0, c1, 0.0, 0.0, 0.0, s1],
1388        ];
1389        round_trip(ResponseManifold::Stiefel { k: 2, n: 4 }, values);
1390    }
1391
1392    #[test]
1393    fn poincare_round_trip_and_mean() {
1394        let values = array![[0.1, 0.2], [-0.3, 0.1], [0.2, -0.25],];
1395        round_trip(
1396            ResponseManifold::Poincare {
1397                dim: 2,
1398                curvature: -1.0,
1399            },
1400            values,
1401        );
1402    }
1403
1404    #[test]
1405    fn resolver_rejects_bad_shapes() {
1406        assert!(ResponseManifold::resolve("grassmann", Some(2), Some(3), None, None).is_err());
1407        assert!(ResponseManifold::resolve("spd", None, None, None, None).is_err());
1408        assert!(ResponseManifold::resolve("poincare", None, None, Some(2), Some(1.0)).is_err());
1409        assert!(ResponseManifold::resolve("nonsense", None, None, None, None).is_err());
1410        assert_eq!(
1411            ResponseManifold::resolve("spd", Some(3), None, None, None).unwrap(),
1412            ResponseManifold::Spd { n: 3 }
1413        );
1414    }
1415
1416    #[test]
1417    fn parse_infers_shapes_from_columns() {
1418        // SPD: n from the perfect-square column count.
1419        assert_eq!(
1420            ResponseManifold::parse("spd", 9).unwrap(),
1421            ResponseManifold::Spd { n: 3 }
1422        );
1423        assert!(ResponseManifold::parse("spd", 8).is_err());
1424        // Grassmann/Stiefel: n inferred as cols / k.
1425        assert_eq!(
1426            ResponseManifold::parse("grassmann(k=2)", 10).unwrap(),
1427            ResponseManifold::Grassmann { k: 2, n: 5 }
1428        );
1429        assert_eq!(
1430            ResponseManifold::parse("Stiefel( k = 2 , n = 4 )", 8).unwrap(),
1431            ResponseManifold::Stiefel { k: 2, n: 4 }
1432        );
1433        assert!(ResponseManifold::parse("grassmann", 10).is_err());
1434        assert!(ResponseManifold::parse("grassmann(k=3)", 10).is_err());
1435        // Poincaré: dim = cols, default curvature -1.
1436        assert_eq!(
1437            ResponseManifold::parse("poincare", 3).unwrap(),
1438            ResponseManifold::Poincare {
1439                dim: 3,
1440                curvature: -1.0
1441            }
1442        );
1443        assert_eq!(
1444            ResponseManifold::parse("poincare(curvature=-0.5)", 3).unwrap(),
1445            ResponseManifold::Poincare {
1446                dim: 3,
1447                curvature: -0.5
1448            }
1449        );
1450        assert!(ResponseManifold::parse("hyperbolic", 3).is_err());
1451    }
1452
1453    #[test]
1454    fn dispatch_round_trips_through_user_label() {
1455        // Drive the full string-selected user path for each geometry: parse the
1456        // label, build the intrinsic base, log to the tangent, exp back.
1457        let cases: Vec<(&str, Array2<f64>)> = vec![
1458            (
1459                "spd",
1460                array![
1461                    [2.0, 0.0, 0.0, 1.0],
1462                    [1.0, 0.3, 0.3, 2.0],
1463                    [3.0, -0.5, -0.5, 1.5],
1464                ],
1465            ),
1466            (
1467                "grassmann(k=1)",
1468                array![[1.0, 0.0, 0.0], [0.0, 1.0, 0.0], [0.6, 0.8, 0.0]],
1469            ),
1470            (
1471                "stiefel(k=1)",
1472                array![[1.0, 0.0, 0.0], [0.0, 1.0, 0.0], [0.0, 0.6, 0.8]],
1473            ),
1474            ("poincare", array![[0.1, 0.2], [-0.3, 0.1], [0.2, -0.25]]),
1475        ];
1476        for (label, values) in cases {
1477            let (tangent, base, canonical) =
1478                dispatch_log_map(values.view(), label, None).expect("dispatch log");
1479            assert!(canonical.starts_with(label.split('(').next().unwrap()));
1480            let back = dispatch_exp_map(tangent.view(), label, base.view()).expect("dispatch exp");
1481            for row in 0..values.nrows() {
1482                for col in 0..values.ncols() {
1483                    assert!(
1484                        (back[[row, col]] - values[[row, col]]).abs() < 1e-6,
1485                        "{label} exp∘log mismatch at ({row},{col}): {} vs {}",
1486                        back[[row, col]],
1487                        values[[row, col]]
1488                    );
1489                }
1490            }
1491        }
1492    }
1493
1494    #[test]
1495    fn ambient_dim_matches_layout() {
1496        assert_eq!(ResponseManifold::Spd { n: 3 }.ambient_dim(), 9);
1497        assert_eq!(ResponseManifold::Grassmann { k: 2, n: 5 }.ambient_dim(), 10);
1498        assert_eq!(ResponseManifold::Stiefel { k: 2, n: 4 }.ambient_dim(), 8);
1499        assert_eq!(
1500            ResponseManifold::Poincare {
1501                dim: 4,
1502                curvature: -1.0
1503            }
1504            .ambient_dim(),
1505            4
1506        );
1507    }
1508
1509    /// Deterministic xorshift64* + Box–Muller standard normals — a dependency-free
1510    /// reproducible source for the synthetic known-κ clouds. Seeded per call so
1511    /// the test is bit-stable across runs and platforms.
1512    struct DetNormal {
1513        state: u64,
1514        spare: Option<f64>,
1515    }
1516    impl DetNormal {
1517        fn new(seed: u64) -> Self {
1518            Self {
1519                state: seed | 1,
1520                spare: None,
1521            }
1522        }
1523        fn u01(&mut self) -> f64 {
1524            // xorshift64*; take the top 53 bits as a (0,1) double.
1525            let mut x = self.state;
1526            x ^= x >> 12;
1527            x ^= x << 25;
1528            x ^= x >> 27;
1529            self.state = x;
1530            let v = x.wrapping_mul(0x2545_F491_4F6C_DD1D);
1531            ((v >> 11) as f64 + 0.5) / (1u64 << 53) as f64
1532        }
1533        fn normal(&mut self) -> f64 {
1534            if let Some(z) = self.spare.take() {
1535                return z;
1536            }
1537            // Box–Muller; clamp u1 away from 0 so ln is finite.
1538            let u1 = self.u01().max(1e-12);
1539            let u2 = self.u01();
1540            let r = (-2.0 * u1.ln()).sqrt();
1541            let theta = 2.0 * std::f64::consts::PI * u2;
1542            self.spare = Some(r * theta.sin());
1543            r * theta.cos()
1544        }
1545    }
1546
1547    /// Build a synthetic cloud at known curvature `k_star`: `n` points whose
1548    /// geodesic normal coordinates about `center` are i.i.d. isotropic Gaussian
1549    /// of scale `sigma`, exp-mapped onto `M_{k_star}`, then mean-centred in the
1550    /// ambient chart to mimic the real (mean-subtracted) response clouds.
1551    fn synth_cloud(dim: usize, k_star: f64, n: usize, sigma: f64, seed: u64) -> Array2<f64> {
1552        let manifold = ResponseManifold::ConstantCurvature { dim, kappa: k_star };
1553        let center = Array1::<f64>::zeros(dim);
1554        let mut rng = DetNormal::new(seed);
1555        let mut values = Array2::<f64>::zeros((n, dim));
1556        for i in 0..n {
1557            let t: Array1<f64> = (0..dim).map(|_| sigma * rng.normal()).collect();
1558            let y = manifold
1559                .exp_point(center.view(), t.view())
1560                .expect("exp tangent to response");
1561            values.row_mut(i).assign(&y);
1562        }
1563        // Mean-centre in the ambient chart (the real-data preprocessing).
1564        let mut mean = Array1::<f64>::zeros(dim);
1565        for row in values.outer_iter() {
1566            mean += &row;
1567        }
1568        mean.mapv_inplace(|v| v / n as f64);
1569        for mut row in values.outer_iter_mut() {
1570            row -= &mean;
1571        }
1572        values
1573    }
1574
1575    /// The #1104 reparameterisation-invariant curvature estimator: on synthetic
1576    /// clouds generated at known κ⋆ the fitted κ̂ must be (a) INTERIOR to the
1577    /// chart bracket (never railed), (b) close to κ⋆ and MONOTONE in κ⋆, (c)
1578    /// produce a smooth (non-degenerate) χ²₁ flatness p-value that does not reject
1579    /// the flat truth, and (d) be correctly COVARIANT under a global rescaling of
1580    /// the cloud (κ has units 1/length², so `y ↦ α y ⇒ κ̂ ↦ κ̂/α²`).
1581    #[test]
1582    fn fit_response_curvature_is_reparameterization_invariant() {
1583        let dim = 3usize;
1584        // Unit-ish scale: σ=0.15 keeps every geodesic radius (≈ a few·σ) well
1585        // inside the κ-stereographic chart for the most hyperbolic κ⋆ = −1.5
1586        // (chart needs ‖y‖² < 1/1.5 ≈ 0.667).
1587        let sigma = 0.15;
1588        let n = 300usize;
1589        let k_stars = [-1.5_f64, -0.5, 0.0, 0.6, 1.2];
1590        let mut k_hats = Vec::new();
1591        for (idx, &k_star) in k_stars.iter().enumerate() {
1592            let values = synth_cloud(dim, k_star, n, sigma, 0xC0FFEE ^ (idx as u64 + 1));
1593            let (kmin, kmax, _rho) = response_kappa_bounds(values.view());
1594            let fit = fit_response_curvature(values.view(), dim, 0.95, 1e-12, 256)
1595                .expect("response curvature fit");
1596            k_hats.push(fit.kappa_hat);
1597
1598            // (a) INTERIOR: κ̂ strictly inside the bracket, not railed to either end.
1599            let span = kmax - kmin;
1600            assert!(
1601                fit.kappa_hat > kmin + 0.02 * span && fit.kappa_hat < kmax - 0.02 * span,
1602                "κ⋆={k_star}: κ̂={} railed to bracket [{kmin}, {kmax}]",
1603                fit.kappa_hat
1604            );
1605
1606            // (b-direct) recovery within a sane tolerance (finite-sample bias is
1607            // O(1/n); the estimator only needs the right region and sign).
1608            assert!(
1609                (fit.kappa_hat - k_star).abs() <= 0.6 + 0.3 * k_star.abs(),
1610                "κ⋆={k_star}: κ̂={} too far",
1611                fit.kappa_hat
1612            );
1613
1614            // (c) the profile CI is a valid interval bracketing κ̂.
1615            assert!(
1616                fit.profile_ci.ci_lo <= fit.kappa_hat && fit.kappa_hat <= fit.profile_ci.ci_hi,
1617                "κ⋆={k_star}: CI [{}, {}] excludes κ̂={}",
1618                fit.profile_ci.ci_lo,
1619                fit.profile_ci.ci_hi,
1620                fit.kappa_hat
1621            );
1622            // The flatness LR statistic and p-value are valid; the p-value is a
1623            // genuine probability strictly between 0 and 1 (smooth, not 0/1).
1624            assert!(fit.flatness.lr_stat >= 0.0);
1625            assert!(
1626                fit.flatness.p_value > 0.0 && fit.flatness.p_value < 1.0,
1627                "κ⋆={k_star}: degenerate flatness p={}",
1628                fit.flatness.p_value
1629            );
1630            // The flat truth κ⋆ = 0 must NOT be rejected at 5% (lr < χ²_{1,.95}).
1631            if k_star == 0.0 {
1632                assert!(
1633                    fit.flatness.lr_stat < 3.84,
1634                    "flat truth wrongly rejected: lr={}",
1635                    fit.flatness.lr_stat
1636                );
1637            }
1638
1639            // (d) RESCALING COVARIANCE: scale the SAME cloud by α and refit; κ̂
1640            // must transform as κ̂/α² (curvature has units 1/length²). We reuse the
1641            // identical points so the only change is the global scale.
1642            let alpha = 1.5_f64;
1643            let scaled = values.mapv(|v| alpha * v);
1644            let fit_scaled = fit_response_curvature(scaled.view(), dim, 0.95, 1e-12, 256)
1645                .expect("scaled response curvature fit");
1646            let expected = fit.kappa_hat / (alpha * alpha);
1647            // Tolerance scales with magnitude; the transform is exact in the
1648            // criterion (V(κ, αy) = V(α²κ, y)) up to the finite golden-section /
1649            // bracket discretisation.
1650            assert!(
1651                (fit_scaled.kappa_hat - expected).abs() <= 0.05 + 0.05 * expected.abs(),
1652                "κ⋆={k_star}: rescale covariance broken: κ̂(αy)={} vs κ̂(y)/α²={}",
1653                fit_scaled.kappa_hat,
1654                expected
1655            );
1656        }
1657
1658        // (b-monotone) κ̂ is monotone increasing in κ⋆ across the whole sweep.
1659        for w in k_hats.windows(2) {
1660            assert!(w[1] > w[0] - 0.05, "κ̂ not monotone in κ⋆: {:?}", k_hats);
1661        }
1662    }
1663
1664    /// d = 1 carries REDUCED curvature information: the transverse volume
1665    /// Jacobian is identically 1 (radial isometry), so κ is identified by the
1666    /// conformal-factor restoring force `−d·Σ ln λ_{y_i}` alone (#944 power
1667    /// analysis). The estimator must still run end-to-end, return an INTERIOR
1668    /// κ̂, and produce a valid CI — never divide/exponentiate the absent
1669    /// transverse direction.
1670    #[test]
1671    fn fit_response_curvature_d1_uses_conformal_term_only() {
1672        let sigma = 0.12;
1673        let n = 400usize;
1674        for &k_star in &[-1.0_f64, 0.0, 0.8] {
1675            let values = synth_cloud(1, k_star, n, sigma, 0xD1 ^ (k_star.to_bits()));
1676            let (kmin, kmax, _rho) = response_kappa_bounds(values.view());
1677            let fit = fit_response_curvature(values.view(), 1, 0.95, 1e-12, 256)
1678                .expect("d=1 curvature fit");
1679            let span = kmax - kmin;
1680            assert!(
1681                fit.kappa_hat > kmin + 0.01 * span && fit.kappa_hat < kmax - 0.01 * span,
1682                "d=1 κ⋆={k_star}: κ̂={} railed to [{kmin},{kmax}]",
1683                fit.kappa_hat
1684            );
1685            assert!(
1686                fit.profile_ci.ci_lo <= fit.kappa_hat && fit.kappa_hat <= fit.profile_ci.ci_hi,
1687                "d=1 κ⋆={k_star}: CI excludes κ̂"
1688            );
1689            assert!(fit.kappa_hat.is_finite() && fit.v_p_hat.is_finite());
1690        }
1691    }
1692
1693    /// The criterion guard must reject κ probes AT or PAST the chart boundary
1694    /// gracefully (an `Err`, never a panic / NaN): on the hyperbolic edge
1695    /// `1 + κ‖y‖² ≤ 0` and on the spherical antipode. The `response_kappa_bounds`
1696    /// bracket stays strictly interior, but a stray CI/LR probe can land on the
1697    /// edge, so the criterion itself must be defensive.
1698    #[test]
1699    fn response_curvature_criterion_rejects_boundary_probes() {
1700        // A cloud with a known max radius R²; the hyperbolic edge is κ = −1/R².
1701        let values = array![[0.5_f64, 0.0], [-0.4, 0.3], [0.1, -0.5]];
1702        let r2_max = values
1703            .outer_iter()
1704            .map(|r| r.dot(&r))
1705            .fold(0.0_f64, f64::max);
1706        // Exactly on / past the hyperbolic edge: 1 + κ‖y‖² = 0 (or < 0).
1707        let kappa_edge = -1.0 / r2_max;
1708        assert!(
1709            response_curvature_criterion(values.view(), 2, kappa_edge).is_err(),
1710            "criterion must reject the hyperbolic chart edge κ=−1/R²"
1711        );
1712        assert!(
1713            response_curvature_criterion(values.view(), 2, 1.5 * kappa_edge).is_err(),
1714            "criterion must reject past the hyperbolic chart edge"
1715        );
1716        // Interior κ just inside the edge succeeds and is finite.
1717        let (v, _) = response_curvature_criterion(values.view(), 2, 0.9 * kappa_edge)
1718            .expect("interior κ valid");
1719        assert!(v.is_finite());
1720        // Non-finite κ is rejected up front.
1721        assert!(response_curvature_criterion(values.view(), 2, f64::NAN).is_err());
1722        assert!(response_curvature_criterion(values.view(), 2, f64::INFINITY).is_err());
1723    }
1724
1725    // ── Projection residual (distance to candidate manifold) ───────────────
1726
1727    #[test]
1728    fn projection_residual_is_zero_for_on_manifold_points() {
1729        // On-manifold rows are their own nearest point, so the residual is ~0
1730        // row-wise. No base point / Fréchet mean is involved — projection is
1731        // base-independent — so this no longer depends on the inputs forming an
1732        // admissible Karcher seed.
1733        let cases: Vec<(ResponseManifold, Array2<f64>)> = vec![
1734            (
1735                ResponseManifold::Spd { n: 2 }, // PD: eigenvalues {2,1} and {2,1}
1736                array![[2.0, 0.0, 0.0, 1.0], [1.5, 0.5, 0.5, 1.5]],
1737            ),
1738            (
1739                ResponseManifold::Grassmann { k: 1, n: 3 }, // unit columns
1740                array![[1.0, 0.0, 0.0], [0.6, 0.8, 0.0]],
1741            ),
1742            (
1743                ResponseManifold::Poincare {
1744                    dim: 2,
1745                    curvature: -1.0,
1746                }, // strictly inside the ball
1747                array![[0.1, 0.2], [-0.3, 0.1]],
1748            ),
1749        ];
1750        for (manifold, values) in cases {
1751            let (resid, rel) =
1752                response_projection_residual(manifold, values.view()).expect("projection residual");
1753            for row in 0..values.nrows() {
1754                assert!(
1755                    resid[row] < 1e-9,
1756                    "{manifold:?} on-manifold row {row} should have ~0 residual, got {}",
1757                    resid[row]
1758                );
1759                assert!(rel[row] < 1e-9 && rel[row] >= 0.0);
1760            }
1761        }
1762    }
1763
1764    #[test]
1765    fn projection_residual_recovers_known_off_manifold_displacement() {
1766        // Closed-form checks against the exact nearest-point distance.
1767
1768        // Gr(1,3) / sphere: nearest unit vector to x is x/‖x‖, so the distance
1769        // is |‖x‖ − 1|. [2,0,0] ⇒ 1; [0,3,0] ⇒ 2. Relative = dist/‖x‖.
1770        let g = ResponseManifold::Grassmann { k: 1, n: 3 };
1771        let gv = array![[2.0, 0.0, 0.0], [0.0, 3.0, 0.0]];
1772        let (gres, grel) = response_projection_residual(g, gv.view()).expect("grassmann");
1773        assert!((gres[0] - 1.0).abs() < 1e-12, "got {}", gres[0]);
1774        assert!((gres[1] - 2.0).abs() < 1e-12, "got {}", gres[1]);
1775        assert!((grel[0] - 0.5).abs() < 1e-12);
1776        assert!((grel[1] - 2.0 / 3.0).abs() < 1e-12);
1777
1778        // SPD(2): nearest PSD matrix clamps negative eigenvalues to 0, so the
1779        // distance is the norm of the discarded negative part. [[1,0],[0,-1]]
1780        // has eigenvalue −1 discarded ⇒ distance 1; ‖x‖_F = √2.
1781        let s = ResponseManifold::Spd { n: 2 };
1782        let sv = array![[1.0, 0.0, 0.0, -1.0]];
1783        let (sres, srel) = response_projection_residual(s, sv.view()).expect("spd");
1784        assert!((sres[0] - 1.0).abs() < 1e-9, "got {}", sres[0]);
1785        assert!((srel[0] - 1.0 / 2.0_f64.sqrt()).abs() < 1e-9);
1786
1787        // Poincaré ball (c = −1, true radius R = 1): the distance to the open
1788        // ball is max(0, ‖x‖ − R). [3,0] ⇒ exactly 2 (not 3 − (1 − BOUNDARY_EPS)
1789        // — the diagnostic uses the manifold radius, not the safety radius).
1790        let p = ResponseManifold::Poincare {
1791            dim: 2,
1792            curvature: -1.0,
1793        };
1794        let pv = array![[3.0, 0.0]];
1795        let (pres, _prel) = response_projection_residual(p, pv.view()).expect("poincare");
1796        assert!((pres[0] - 2.0).abs() < 1e-12, "got {}", pres[0]);
1797
1798        // A different curvature (c = −4, R = 1/2): [2,0] ⇒ 2 − 0.5 = 1.5.
1799        let p4 = ResponseManifold::Poincare {
1800            dim: 2,
1801            curvature: -4.0,
1802        };
1803        let (p4res, _) =
1804            response_projection_residual(p4, array![[2.0, 0.0]].view()).expect("poincare c=-4");
1805        assert!((p4res[0] - 1.5).abs() < 1e-12, "got {}", p4res[0]);
1806    }
1807
1808    #[test]
1809    fn projection_residual_validates_shapes_and_finiteness() {
1810        let manifold = ResponseManifold::Spd { n: 2 }; // ambient = 4
1811        // Wrong column count.
1812        let bad_cols = array![[1.0, 2.0, 3.0]];
1813        assert!(response_projection_residual(manifold, bad_cols.view()).is_err());
1814        // Non-finite value.
1815        let nan_vals = array![[f64::NAN, 0.0, 0.0, 1.0]];
1816        assert!(response_projection_residual(manifold, nan_vals.view()).is_err());
1817        let inf_vals = array![[f64::INFINITY, 0.0, 0.0, 1.0]];
1818        assert!(response_projection_residual(manifold, inf_vals.view()).is_err());
1819    }
1820
1821    #[test]
1822    fn projection_residual_separates_on_and_off_manifold() {
1823        // The motivating case, now honestly answered: an on-manifold row sits
1824        // at zero distance from the candidate shape; a row pushed off it has a
1825        // clearly positive distance. This is the shape-plausibility signal that
1826        // gates which topology is worth fitting — not the post-fit membership
1827        // decision, which comes from the fitted surface's residual instead.
1828        let manifold = ResponseManifold::Grassmann { k: 1, n: 3 };
1829        let on = array![[0.6, 0.8, 0.0]]; // a genuine unit direction
1830        let off = array![[0.6, 0.8, 1.4]]; // same direction, pushed off-sphere
1831
1832        let (resid_on, _) = response_projection_residual(manifold, on.view()).expect("on");
1833        let (resid_off, _) = response_projection_residual(manifold, off.view()).expect("off");
1834
1835        assert!(
1836            resid_on[0] < 1e-9,
1837            "on-manifold should be ~0, got {}",
1838            resid_on[0]
1839        );
1840        assert!(
1841            resid_off[0] > 1e-2 && resid_off[0] > resid_on[0],
1842            "off-manifold distance ({}) must clearly exceed on-manifold ({})",
1843            resid_off[0],
1844            resid_on[0]
1845        );
1846    }
1847
1848    #[test]
1849    fn projection_residual_supports_k_greater_than_one_frames() {
1850        // k > 1 frames use the closed form √Σ(σ_i − 1)². St(2,3), ambient = 6,
1851        // row-major n×k.
1852        let manifold = ResponseManifold::Stiefel { k: 2, n: 3 };
1853
1854        // An orthonormal frame [e1 | e2] is its own nearest point ⇒ residual 0.
1855        let on = array![[1.0, 0.0, 0.0, 1.0, 0.0, 0.0]];
1856        let (resid_on, _) = response_projection_residual(manifold, on.view()).expect("on");
1857        assert!(
1858            resid_on[0] < 1e-9,
1859            "orthonormal frame should be ~0, got {}",
1860            resid_on[0]
1861        );
1862
1863        // Scale the first column by 2: Y = [2·e1 | e2]. YᵀY = diag(4,1) ⇒
1864        // σ = (2,1), distance √((2−1)²+(1−1)²) = 1, relative = 1/‖Y‖_F = 1/√5.
1865        let off = array![[2.0, 0.0, 0.0, 1.0, 0.0, 0.0]];
1866        let (resid_off, rel_off) = response_projection_residual(manifold, off.view()).expect("off");
1867        assert!((resid_off[0] - 1.0).abs() < 1e-9, "got {}", resid_off[0]);
1868        assert!(
1869            (rel_off[0] - 1.0 / 5.0_f64.sqrt()).abs() < 1e-9,
1870            "got {}",
1871            rel_off[0]
1872        );
1873
1874        // Grassmann(2,4) gives the identical score for the same frame data.
1875        let g = ResponseManifold::Grassmann { k: 2, n: 4 };
1876        let g_on = array![[1.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0]];
1877        let (g_resid, _) = response_projection_residual(g, g_on.view()).expect("grassmann");
1878        assert!(g_resid[0] < 1e-9, "got {}", g_resid[0]);
1879    }
1880
1881    #[test]
1882    fn projection_residual_handles_nontrivial_eigenvectors() {
1883        // A frame whose Gram is NOT diagonal, so the singular values come from a
1884        // genuine eigendecomposition. Y = [[1,1],[0,1],[0,0]] (St(2,3)):
1885        // YᵀY = [[1,1],[1,2]], eigenvalues (3±√5)/2, σ = ((1+√5)/2, (√5−1)/2).
1886        // distance² = (σ₁−1)² + (σ₂−1)².
1887        let manifold = ResponseManifold::Stiefel { k: 2, n: 3 };
1888        let y = array![[1.0, 1.0, 0.0, 1.0, 0.0, 0.0]]; // row-major rows [1,1],[0,1],[0,0]
1889        let (resid, _) = response_projection_residual(manifold, y.view()).expect("frame");
1890        let s5 = 5.0_f64.sqrt();
1891        let sig1 = (1.0 + s5) / 2.0;
1892        let sig2 = (s5 - 1.0) / 2.0;
1893        let expect = ((sig1 - 1.0).powi(2) + (sig2 - 1.0).powi(2)).sqrt();
1894        assert!(
1895            (resid[0] - expect).abs() < 1e-9,
1896            "got {} want {}",
1897            resid[0],
1898            expect
1899        );
1900    }
1901
1902    #[test]
1903    fn projection_residual_is_defined_for_rank_deficient_frames() {
1904        // A rank-deficient frame has a well-defined distance even though the
1905        // nearest orthonormal frame is not unique — distance to a compact set is
1906        // always defined, so this must NOT error. Two identical columns e1 give
1907        // YᵀY = [[1,1],[1,1]], σ = (√2, 0), distance √((√2−1)²+(0−1)²) = √(4−2√2).
1908        let manifold = ResponseManifold::Stiefel { k: 2, n: 3 };
1909        let degenerate = array![[1.0, 1.0, 0.0, 0.0, 0.0, 0.0]]; // both columns = e1
1910        let (resid, _) =
1911            response_projection_residual(manifold, degenerate.view()).expect("rank-deficient ok");
1912        let expect = (4.0 - 2.0 * 2.0_f64.sqrt()).sqrt(); // ≈ 1.0823922
1913        assert!(
1914            (resid[0] - expect).abs() < 1e-9,
1915            "got {} want {}",
1916            resid[0],
1917            expect
1918        );
1919
1920        // Minimal case: zero vector on the sphere (Gr(1,3)). Every unit vector is
1921        // a nearest point and the distance is exactly 1 — also must not error.
1922        let sphere = ResponseManifold::Grassmann { k: 1, n: 3 };
1923        let (zres, _) =
1924            response_projection_residual(sphere, array![[0.0, 0.0, 0.0]].view()).expect("zero");
1925        assert!((zres[0] - 1.0).abs() < 1e-12, "got {}", zres[0]);
1926    }
1927
1928    #[test]
1929    fn projection_residual_handles_tiny_full_rank_frame() {
1930        // A tiny but full-rank frame must NOT be rejected as rank-deficient: the
1931        // distance is scale-correct. Y = 1e-7·[e1 | e2] (St(2,3)) ⇒ σ = (1e-7,
1932        // 1e-7), distance √2·(1 − 1e-7) ≈ 1.41421342.
1933        let manifold = ResponseManifold::Stiefel { k: 2, n: 3 };
1934        let tiny = array![[1e-7, 0.0, 0.0, 1e-7, 0.0, 0.0]];
1935        let (resid, _) = response_projection_residual(manifold, tiny.view()).expect("tiny ok");
1936        let expect = 2.0_f64.sqrt() * (1.0 - 1e-7);
1937        assert!(
1938            (resid[0] - expect).abs() < 1e-9,
1939            "got {} want {}",
1940            resid[0],
1941            expect
1942        );
1943    }
1944
1945    #[test]
1946    fn projection_residual_spd_nonsymmetric_and_singular() {
1947        // Non-symmetric input: A = [[1,1],[-1,1]] has sym(A) = I (no negative
1948        // part), but the distance to the PSD cone still counts the skew part:
1949        // ‖A − I‖_F = √2.
1950        let spd = ResponseManifold::Spd { n: 2 };
1951        let asym = array![[1.0, 1.0, -1.0, 1.0]]; // row-major [[1,1],[-1,1]]
1952        let (ares, _) = response_projection_residual(spd, asym.view()).expect("nonsym");
1953        assert!((ares[0] - 2.0_f64.sqrt()).abs() < 1e-9, "got {}", ares[0]);
1954
1955        // A singular PSD matrix diag(1,0) is in the closed cone ⇒ distance 0
1956        // (even though it is not strictly positive definite).
1957        let singular = array![[1.0, 0.0, 0.0, 0.0]];
1958        let (sres, _) = response_projection_residual(spd, singular.view()).expect("singular psd");
1959        assert!(
1960            sres[0] < 1e-12,
1961            "singular PSD should be ~0, got {}",
1962            sres[0]
1963        );
1964    }
1965
1966    #[test]
1967    fn projection_residual_poincare_interior_shell_is_zero() {
1968        // A point in the numerical safety shell R_safe < ‖x‖ < R is a genuine
1969        // interior point of the manifold ball, so it must score exactly 0 — the
1970        // diagnostic uses the true radius, not the projection safety radius.
1971        let p = ResponseManifold::Poincare {
1972            dim: 2,
1973            curvature: -1.0,
1974        };
1975        let shell = array![[0.999999, 0.0]]; // inside R = 1, outside R_safe ≈ 0.99999
1976        let (resid, _) = response_projection_residual(p, shell.view()).expect("shell");
1977        assert!(
1978            resid[0] < 1e-12,
1979            "interior point must be 0, got {}",
1980            resid[0]
1981        );
1982    }
1983
1984    #[test]
1985    fn projection_residual_handles_constant_curvature_domain() {
1986        // ConstantCurvature is a fittable response geometry produced by the
1987        // resolver/parser, so it must return a closed-form distance, not error.
1988        // κ ≥ 0: chart is all of ℝ^d ⇒ every finite row scores 0.
1989        let pos = ResponseManifold::parse("constant_curvature(dim=3,kappa=1.0)", 3)
1990            .expect("parse constant_curvature");
1991        assert!(matches!(pos, ResponseManifold::ConstantCurvature { .. }));
1992        let (pres, _) =
1993            response_projection_residual(pos, array![[0.1, 9.0, -100.0]].view()).expect("kappa>=0");
1994        assert!(pres[0] < 1e-12, "κ≥0 finite row must be 0, got {}", pres[0]);
1995
1996        // κ < 0: chart is the ball of radius 1/√(−κ). For κ = −1, R = 1, so a
1997        // point of norm 3 is at distance 2; an interior point is at 0.
1998        let neg = ResponseManifold::ConstantCurvature {
1999            dim: 2,
2000            kappa: -1.0,
2001        };
2002        let (nres, _) = response_projection_residual(neg, array![[3.0, 0.0], [0.2, 0.1]].view())
2003            .expect("kappa<0");
2004        assert!((nres[0] - 2.0).abs() < 1e-12, "got {}", nres[0]);
2005        assert!(nres[1] < 1e-12, "interior row must be 0, got {}", nres[1]);
2006    }
2007
2008    #[test]
2009    fn projection_residual_accepts_empty_batch() {
2010        // A zero-row batch is valid and returns empty arrays for every geometry.
2011        let manifold = ResponseManifold::Spd { n: 2 }; // ambient = 4
2012        let empty = Array2::<f64>::zeros((0, 4));
2013        let (resid, rel) = response_projection_residual(manifold, empty.view()).expect("empty");
2014        assert_eq!(resid.len(), 0);
2015        assert_eq!(rel.len(), 0);
2016    }
2017}