ferrolearn-linear 0.5.0

Linear models for the ferrolearn ML framework
Documentation
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
//! Ridge regression with built-in cross-validation for alpha selection.
//!
//! This module provides [`RidgeCV`], which automatically selects the best
//! regularization parameter `alpha` from a candidate list. It mirrors
//! scikit-learn's `RidgeCV` (`sklearn/linear_model/_ridge.py`): by **default**
//! (`cv = None`) it uses efficient leave-one-out Generalized Cross-Validation
//! (`_RidgeGCV`, `_ridge.py:1688`), computing the LOO prediction errors in
//! closed form from a single matrix decomposition reused across all `alphas`;
//! when an explicit fold count `cv = Some(k)` is set it falls back to
//! brute-force k-fold cross-validation (mirroring sklearn's `GridSearchCV`
//! branch, `_ridge.py:2413-2439`). In both cases the final model is refit on
//! the full data with the chosen alpha.
//!
//! ## REQ status
//!
//! Two states only (SHIPPED / NOT-STARTED), per goal.md.
//!
//! See `.design/linear/ridge_cv.md` for the full requirements table.
//!
//! | REQ | Status | Evidence |
//! |---|---|---|
//! | REQ-9 (non-finite input rejected) | SHIPPED | `Fit for RidgeCV::fit` rejects any NaN/+/-inf in X or y BEFORE the GCV decomposition / k-fold split with `FerroError::InvalidParameter`, mirroring sklearn `_RidgeGCV.fit` `_validate_data(force_all_finite=True)` (`_ridge.py:2087`) and the `cv=Some(k)` `GridSearchCV`→`Ridge.fit` (`_ridge.py:1242`) path, both raising `ValueError("Input X contains NaN.")` / `"... contains infinity ..."`. ferrolearn's `Fit::fit` takes only `(x, y)` (no `sample_weight` in the trait surface), so X and y are validated. `.iter().any(|v| !v.is_finite())` catches both NaN and Inf; the finite path is byte-identical (alpha_/coef_ unchanged). Verified vs the live sklearn 1.5.2 oracle (R-CHAR-3): NaN/+inf/-inf in X and NaN/inf in y all raise `ValueError`; all-finite `alpha_=0.1` unchanged (`tests/divergence_linear_nonfinite_batch5.rs::ridge_cv_*`). Non-test consumer: the existing `pub use ridge_cv::RidgeCV` boundary API. (#2265) |
//!
//! # Examples
//!
//! ```
//! use ferrolearn_linear::RidgeCV;
//! use ferrolearn_core::{Fit, Predict};
//! use ndarray::{array, Array1, Array2};
//!
//! let model = RidgeCV::<f64>::new();
//! let x = Array2::from_shape_vec((6, 1), vec![1.0, 2.0, 3.0, 4.0, 5.0, 6.0]).unwrap();
//! let y = array![2.0, 4.0, 6.0, 8.0, 10.0, 12.0];
//!
//! let fitted = model.fit(&x, &y).unwrap();
//! let preds = fitted.predict(&x).unwrap();
//! assert_eq!(preds.len(), 6);
//! ```

use ferray::linalg::LinalgFloat;
use ferray::{Array as FerrayArray, Ix2};
use ferrolearn_core::error::FerroError;
use ferrolearn_core::introspection::HasCoefficients;
use ferrolearn_core::traits::{Fit, Predict};
use ndarray::{Array1, Array2, Axis, ScalarOperand};
use num_traits::{Float, FromPrimitive};

use crate::Ridge;

/// Ridge regression with built-in cross-validation for alpha selection.
///
/// Selects the regularization strength `alpha` from a candidate grid, then
/// refits on the full training data with the chosen alpha. The selection
/// method mirrors scikit-learn's `RidgeCV`:
///
/// - **`cv = None` (the default)** — efficient leave-one-out Generalized
///   Cross-Validation (`_RidgeGCV`, `_ridge.py:1688`). A single matrix
///   decomposition is reused across all `alphas`; for each candidate the LOO
///   prediction errors are computed in closed form and the alpha minimising
///   the mean squared LOO error is selected.
/// - **`cv = Some(k)`** — brute-force k-fold cross-validation (mirroring
///   sklearn's `GridSearchCV` branch, `_ridge.py:2429-2434`): each candidate
///   alpha is scored by averaging the per-fold R² (sklearn `GridSearchCV`
///   default scoring, `scoring=None` → `RegressorMixin.score` = R²), and the
///   alpha with the MAXIMUM averaged R² is selected.
///
/// # Type Parameters
///
/// - `F`: The floating-point type (`f32` or `f64`).
#[derive(Debug, Clone)]
pub struct RidgeCV<F> {
    /// Candidate regularization strengths to evaluate.
    alphas: Vec<F>,
    /// Cross-validation strategy: `None` selects the default leave-one-out
    /// Generalized Cross-Validation path; `Some(k)` selects brute-force
    /// k-fold cross-validation with `k` folds.
    cv: Option<usize>,
    /// Whether to fit an intercept (bias) term.
    fit_intercept: bool,
}

impl<F: Float + FromPrimitive> RidgeCV<F> {
    /// Create a new `RidgeCV` with default settings.
    ///
    /// Defaults: `alphas = [0.1, 1.0, 10.0]`, `cv = None` (leave-one-out
    /// Generalized Cross-Validation, matching sklearn's default), and
    /// `fit_intercept = true`.
    #[must_use]
    pub fn new() -> Self {
        // `F::from(_)` returns `Option`; fall back to `one` (never hit for
        // f32/f64 literals) rather than unwrap in library code (R-CODE-2).
        let one = <F as num_traits::One>::one();
        let p1 = F::from(0.1).unwrap_or(one);
        let ten = F::from(10.0).unwrap_or(one);
        Self {
            alphas: vec![p1, one, ten],
            cv: None,
            fit_intercept: true,
        }
    }

    /// Set the candidate regularization strengths.
    ///
    /// Each value must be non-negative.
    #[must_use]
    pub fn with_alphas(mut self, alphas: Vec<F>) -> Self {
        self.alphas = alphas;
        self
    }

    /// Set the number of cross-validation folds, switching to the brute-force
    /// k-fold path (mirroring sklearn `RidgeCV(cv=k)`).
    ///
    /// Must be at least 2. To use the default leave-one-out Generalized
    /// Cross-Validation path, do not call this (leave `cv = None`).
    #[must_use]
    pub fn with_cv(mut self, cv: usize) -> Self {
        self.cv = Some(cv);
        self
    }

    /// Set whether to fit an intercept term.
    #[must_use]
    pub fn with_fit_intercept(mut self, fit_intercept: bool) -> Self {
        self.fit_intercept = fit_intercept;
        self
    }
}

impl<F: Float + FromPrimitive> Default for RidgeCV<F> {
    fn default() -> Self {
        Self::new()
    }
}

/// Fitted Ridge regression model with cross-validated alpha.
///
/// Stores the selected alpha, learned coefficients, and intercept. Implements
/// [`Predict`] and [`HasCoefficients`] for introspection.
#[derive(Debug, Clone)]
pub struct FittedRidgeCV<F> {
    /// The alpha that achieved the lowest CV error.
    best_alpha: F,
    /// Learned coefficient vector (one per feature).
    coefficients: Array1<F>,
    /// Learned intercept (bias) term.
    intercept: F,
}

impl<F: Float> FittedRidgeCV<F> {
    /// Returns the alpha value that was selected by cross-validation.
    #[must_use]
    pub fn best_alpha(&self) -> F {
        self.best_alpha
    }
}

/// Split sample indices into `k` CONTIGUOUS folds, mirroring sklearn
/// `KFold(shuffle=False)` (`sklearn/model_selection/_split.py`, `_BaseKFold`):
/// `fold_sizes = full(n_splits, n_samples // n_splits)`,
/// `fold_sizes[: n_samples % n_splits] += 1`, then yield consecutive index
/// blocks. The first `n_samples % k` folds get one extra element; folds are
/// contiguous ranges, NOT round-robin. The `cv = Some(k)` k-fold path resolves
/// to this splitter because sklearn's `GridSearchCV(estimator, cv=k)`
/// (`_ridge.py:2429`) uses `KFold(n_splits=k, shuffle=False)` for a regressor.
/// (Prior identical fix for LassoCV: #421, commit abf5a14.)
fn kfold_indices(n_samples: usize, k: usize) -> Vec<Vec<usize>> {
    let base = n_samples / k;
    let rem = n_samples % k;
    let mut folds = Vec::with_capacity(k);
    let mut current = 0;
    for f in 0..k {
        let fold_size = base + if f < rem { 1 } else { 0 };
        folds.push((current..current + fold_size).collect());
        current += fold_size;
    }
    folds
}

/// Compute the R² (coefficient of determination) score between true and
/// predicted targets, mirroring sklearn `r2_score`
/// (`sklearn/metrics/_regression.py:1220-1228` numerator/denominator, then
/// `_assemble_r2_explained_variance`, `_regression.py:872-891`) for the
/// single-output `multioutput='uniform_average'` case: `1 - SS_res/SS_tot`,
/// with the `SS_tot == 0` edge returning `1.0` when `SS_res == 0` (perfect)
/// else `0.0`. This is the default `GridSearchCV` scorer (`RegressorMixin`'s
/// `.score()`) that sklearn `RidgeCV(cv=k, scoring=None)` maximizes
/// (`_ridge.py:2429-2434`).
fn r2_score<F: Float + FromPrimitive + 'static>(y_true: &Array1<F>, y_pred: &Array1<F>) -> F {
    let n = F::from(y_true.len()).unwrap_or_else(F::one);
    let mean = y_true.iter().copied().fold(F::zero(), |a, b| a + b) / n;
    let ss_res = y_true
        .iter()
        .zip(y_pred.iter())
        .map(|(&t, &p)| (t - p) * (t - p))
        .fold(F::zero(), |a, b| a + b);
    let ss_tot = y_true
        .iter()
        .map(|&t| (t - mean) * (t - mean))
        .fold(F::zero(), |a, b| a + b);
    if ss_tot == F::zero() {
        if ss_res == F::zero() {
            F::one()
        } else {
            F::zero()
        }
    } else {
        F::one() - ss_res / ss_tot
    }
}

/// Gather rows from a 2-D array by index.
fn select_rows<F: Float>(x: &Array2<F>, indices: &[usize]) -> Array2<F> {
    let ncols = x.ncols();
    let mut out = Array2::<F>::zeros((indices.len(), ncols));
    for (out_row, &idx) in indices.iter().enumerate() {
        out.row_mut(out_row).assign(&x.row(idx));
    }
    out
}

/// Gather elements from a 1-D array by index.
fn select_elements<F: Float>(y: &Array1<F>, indices: &[usize]) -> Array1<F> {
    Array1::from_iter(indices.iter().map(|&i| y[i]))
}

impl<F: Float + Send + Sync + ScalarOperand + FromPrimitive + LinalgFloat + 'static>
    Fit<Array2<F>, Array1<F>> for RidgeCV<F>
{
    type Fitted = FittedRidgeCV<F>;
    type Error = FerroError;

    /// Fit the `RidgeCV` model.
    ///
    /// Selects `alpha` by leave-one-out Generalized Cross-Validation when
    /// `cv = None` (the default; mirrors sklearn's `_RidgeGCV` path,
    /// `_ridge.py:2382`) or by brute-force k-fold cross-validation when
    /// `cv = Some(k)` (mirrors sklearn's `GridSearchCV` path,
    /// `_ridge.py:2413-2439`), then refits on the full data with the chosen
    /// alpha.
    ///
    /// # Errors
    ///
    /// - [`FerroError::ShapeMismatch`] if `x` and `y` have different numbers
    ///   of samples.
    /// - [`FerroError::InvalidParameter`] if `alphas` is empty or contains
    ///   a negative value, or if `cv = Some(k)` with `k < 2`.
    /// - [`FerroError::InsufficientSamples`] if the number of samples is less
    ///   than the number of folds (k-fold path).
    fn fit(&self, x: &Array2<F>, y: &Array1<F>) -> Result<FittedRidgeCV<F>, FerroError> {
        let (n_samples, _n_features) = x.dim();

        if n_samples != y.len() {
            return Err(FerroError::ShapeMismatch {
                expected: vec![n_samples],
                actual: vec![y.len()],
                context: "y length must match number of samples in X".into(),
            });
        }

        if self.alphas.is_empty() {
            return Err(FerroError::InvalidParameter {
                name: "alphas".into(),
                reason: "must contain at least one candidate".into(),
            });
        }

        for &a in &self.alphas {
            // `<F as num_traits::Zero>::zero()`: the `LinalgFloat` bound pulls
            // `ferray::Element` (also defining `zero`) into scope, making a
            // bare `F::zero()` ambiguous. Disambiguate to `num_traits::Zero`.
            if a < <F as num_traits::Zero>::zero() {
                return Err(FerroError::InvalidParameter {
                    name: "alphas".into(),
                    reason: "all alpha values must be non-negative".into(),
                });
            }
        }

        // RidgeCV split-count guard for the `cv=Some(k)` path, ordered BEFORE the
        // finiteness check (#2268, the MIRROR of #2267). sklearn routes the
        // `cv=Some(k)` path through `GridSearchCV(KFold(n_splits=k))`, whose
        // split-count check (`KFold._iter_test_indices`,
        // `model_selection/_split.py:408`: `if self.n_splits > n_samples: raise
        // ValueError("Cannot have number of splits n_splits={k} greater than the
        // number of samples ...")`) fires BEFORE the per-fold `Ridge.fit`
        // `_validate_data` (`_ridge.py:1242`). So for `n_samples < k` WITH a NaN
        // in X, sklearn raises the split-count (`cv`) error, NOT the non-finite
        // error. ferrolearn therefore checks the split-count for the
        // `cv=Some(k)` arm FIRST, matching that ordering. (`select_alpha_kfold`
        // re-checks `k<2`/`n<k` defensively; this up-front guard reproduces the
        // same `FerroError` variants so the finite-input behaviour is unchanged.)
        if let Some(k) = self.cv {
            if k < 2 {
                return Err(FerroError::InvalidParameter {
                    name: "cv".into(),
                    reason: "number of folds must be at least 2".into(),
                });
            }
            if n_samples < k {
                return Err(FerroError::InsufficientSamples {
                    required: k,
                    actual: n_samples,
                    context: "RidgeCV requires at least as many samples as folds".into(),
                });
            }
        }

        // Non-finite input validation (#2265 batch5). sklearn `RidgeCV.fit`
        // routes to `_RidgeGCV.fit` (`cv=None`) which calls
        // `self._validate_data(X, y, ...)` (`_ridge.py:2087`) keeping the default
        // `force_all_finite=True`, rejecting any NaN or +/-inf in X OR y with a
        // `ValueError` BEFORE the decomposition; the `cv=Some(k)` path routes
        // through `GridSearchCV` whose per-fold `Ridge.fit` (`_ridge.py:1242`)
        // also rejects non-finite input — but only AFTER the `KFold` split-count
        // check above (#2268). Checking finiteness here (after the `cv=Some(k)`
        // split-count guard, before the GCV `cv=None` decomposition) gives the
        // clean sklearn error and matches BOTH modes' ordering. ferrolearn's
        // `Fit::fit` takes only `(x, y)` (no `sample_weight` in the trait
        // surface), so X and y are the inputs validated. `.iter().any(|v|
        // !v.is_finite())` rejects both NaN and Inf (bounds-safe, no panic,
        // R-CODE-2), matching the crate idiom (`ridge.rs`, #2259). The finite
        // path is byte-identical (the guard never fires on finite input).
        if x.iter().any(|v| !v.is_finite()) {
            return Err(FerroError::InvalidParameter {
                name: "X".into(),
                reason: "Input X contains NaN or infinity.".into(),
            });
        }
        if y.iter().any(|v| !v.is_finite()) {
            return Err(FerroError::InvalidParameter {
                name: "y".into(),
                reason: "Input y contains NaN or infinity.".into(),
            });
        }

        // Route on the CV strategy, mirroring sklearn `_BaseRidgeCV.fit`
        // (`_ridge.py:2382`: `if cv is None: ... estimator = _RidgeGCV(...)`,
        // else brute-force `GridSearchCV`).
        let best_alpha = match self.cv {
            None => self.select_alpha_gcv(x, y)?,
            Some(k) => self.select_alpha_kfold(x, y, k)?,
        };

        // Refit on full data with the best alpha, matching sklearn
        // `self.coef_ = estimator.coef_` (`_ridge.py:2441`); `Ridge`'s
        // centering reproduces sklearn's intercept handling.
        let final_model = Ridge::<F>::new()
            .with_alpha(best_alpha)
            .with_fit_intercept(self.fit_intercept);
        let final_fitted = final_model.fit(x, y)?;

        Ok(FittedRidgeCV {
            best_alpha,
            coefficients: final_fitted.coefficients().clone(),
            intercept: final_fitted.intercept(),
        })
    }
}

impl<F: Float + Send + Sync + ScalarOperand + FromPrimitive + LinalgFloat + 'static> RidgeCV<F> {
    /// Brute-force k-fold alpha selection (sklearn `GridSearchCV` branch,
    /// `_ridge.py:2429-2434`). For each candidate alpha, average the per-fold
    /// R² of a `Ridge` refit (sklearn's default `scoring=None` →
    /// `RegressorMixin.score` = R²) and keep the MAXIMISER.
    fn select_alpha_kfold(&self, x: &Array2<F>, y: &Array1<F>, k: usize) -> Result<F, FerroError> {
        let n_samples = x.nrows();

        if k < 2 {
            return Err(FerroError::InvalidParameter {
                name: "cv".into(),
                reason: "number of folds must be at least 2".into(),
            });
        }

        if n_samples < k {
            return Err(FerroError::InsufficientSamples {
                required: k,
                actual: n_samples,
                context: "RidgeCV requires at least as many samples as folds".into(),
            });
        }

        let folds = kfold_indices(n_samples, k);

        let mut best_alpha = self.alphas[0];
        let mut best_r2 = F::neg_infinity();

        for &alpha in &self.alphas {
            let mut total_r2 = <F as num_traits::Zero>::zero();

            for fold_idx in 0..k {
                let test_indices = &folds[fold_idx];
                let train_indices: Vec<usize> = folds
                    .iter()
                    .enumerate()
                    .filter(|&(i, _)| i != fold_idx)
                    .flat_map(|(_, v)| v.iter().copied())
                    .collect();

                let x_train = select_rows(x, &train_indices);
                let y_train = select_elements(y, &train_indices);
                let x_test = select_rows(x, test_indices);
                let y_test = select_elements(y, test_indices);

                let model = Ridge::<F>::new()
                    .with_alpha(alpha)
                    .with_fit_intercept(self.fit_intercept);

                let fitted = model.fit(&x_train, &y_train)?;
                let preds = fitted.predict(&x_test)?;
                total_r2 += r2_score(&y_test, &preds);
            }

            let avg_r2 = total_r2 / F::from(k).unwrap_or_else(<F as num_traits::One>::one);

            if avg_r2 > best_r2 {
                best_r2 = avg_r2;
                best_alpha = alpha;
            }
        }

        Ok(best_alpha)
    }

    /// Default leave-one-out Generalized Cross-Validation alpha selection,
    /// mirroring sklearn `_RidgeGCV.fit` (`_ridge.py:2059`).
    ///
    /// Centers `X`/`y` when `fit_intercept` (uniform weights here, so
    /// `sqrt_sw = 1`; sklearn `_preprocess_data`, `_ridge.py:2106`), decomposes
    /// once via the shape-appropriate mode (sklearn `_check_gcv_mode`,
    /// `_ridge.py:1569`: SVD of the design matrix when `n_samples > n_features`,
    /// otherwise the eigendecomposition of the Gram matrix `X·Xᵀ`), then for
    /// each alpha computes the closed-form LOO errors and picks the alpha
    /// minimising the mean squared LOO error (`scoring=None` →
    /// `_score_without_scorer = -squared_errors.mean()`, `_ridge.py:2211`).
    fn select_alpha_gcv(&self, x: &Array2<F>, y: &Array1<F>) -> Result<F, FerroError> {
        let (n_samples, n_features) = x.dim();

        // Center X and y (sklearn `_preprocess_data`, `_ridge.py:2106`); with
        // uniform weights the centered design has zero column means and the
        // square-root sample weights are all 1.
        let (x_c, y_c) = if self.fit_intercept {
            let x_mean = x
                .mean_axis(Axis(0))
                .ok_or_else(|| FerroError::NumericalInstability {
                    message: "RidgeCV GCV: failed to compute column means".into(),
                })?;
            let y_mean = y.mean().ok_or_else(|| FerroError::NumericalInstability {
                message: "RidgeCV GCV: failed to compute target mean".into(),
            })?;
            (x - &x_mean, y - y_mean)
        } else {
            (x.to_owned(), y.to_owned())
        };

        // Per-alpha squared LOO errors → mean → negative score; sklearn keeps
        // the alpha with the maximal score, i.e. the minimal mean squared LOO
        // error (`_ridge.py:2146-2186`).
        let scores = if n_samples > n_features {
            self.gcv_scores_svd(&x_c, &y_c)?
        } else {
            self.gcv_scores_eigen(&x_c, &y_c)?
        };

        let mut best_alpha = self.alphas[0];
        let mut best_mse = F::infinity();
        for (&alpha, &mean_sq_err) in self.alphas.iter().zip(scores.iter()) {
            if mean_sq_err < best_mse {
                best_mse = mean_sq_err;
                best_alpha = alpha;
            }
        }

        Ok(best_alpha)
    }

    /// SVD-mode GCV scores (sklearn `_svd_decompose_design_matrix`
    /// `_ridge.py:2025` + `_solve_svd_design_matrix` `_ridge.py:2039`), used
    /// when `n_samples > n_features` and `X` is dense. Returns the mean squared
    /// LOO error for each candidate alpha (lower is better).
    fn gcv_scores_svd(&self, x_c: &Array2<F>, y_c: &Array1<F>) -> Result<Vec<F>, FerroError> {
        let n_samples = x_c.nrows();
        let one = <F as num_traits::One>::one();

        // Build the (possibly intercept-augmented) design matrix. With uniform
        // weights `sqrt_sw = 1`, so the appended intercept column is all ones
        // (sklearn `_svd_decompose_design_matrix`, `_ridge.py:2032`:
        // `intercept_column = sqrt_sw[:, None]`).
        let n_cols = if self.fit_intercept {
            x_c.ncols() + 1
        } else {
            x_c.ncols()
        };
        let mut design = Array2::<F>::zeros((n_samples, n_cols));
        design.slice_mut(ndarray::s![.., ..x_c.ncols()]).assign(x_c);
        if self.fit_intercept {
            design.column_mut(x_c.ncols()).fill(one);
        }

        // Thin SVD: U is (n_samples, k), singular values length k. sklearn uses
        // `linalg.svd(X, full_matrices=0)` (`_ridge.py:2034`).
        let (u, singvals) = svd_u_s(&design)?;
        let k = singvals.len();

        // singvals_sq and UT_y (sklearn `_ridge.py:2035-2036`).
        let singvals_sq: Vec<F> = (0..k).map(|j| singvals[j] * singvals[j]).collect();
        // UT_y[j] = sum_i U[i,j] * y[i].
        let mut ut_y = vec![<F as num_traits::Zero>::zero(); k];
        for j in 0..k {
            let mut acc = <F as num_traits::Zero>::zero();
            for i in 0..n_samples {
                acc += u[(i, j)] * y_c[i];
            }
            ut_y[j] = acc;
        }

        // Intercept dimension: the column of U most aligned with the
        // normalized sqrt_sw vector (uniform weights → ones/√n). Since the
        // query is a unit vector, the most-aligned column maximises
        // |query·U[:,j]| (sklearn `_find_smallest_angle`, `_ridge.py:1579`).
        let intercept_dim = if self.fit_intercept {
            Some(find_intercept_dim(&u, n_samples, k))
        } else {
            None
        };

        let n_f = F::from(n_samples).unwrap_or(one);
        let mut out = Vec::with_capacity(self.alphas.len());
        for &alpha in &self.alphas {
            let inv_alpha = one / alpha;
            // w_j = (singvals_sq_j + alpha)^-1 - alpha^-1 (sklearn :2045).
            let mut w: Vec<F> = singvals_sq
                .iter()
                .map(|&s2| one / (s2 + alpha) - inv_alpha)
                .collect();
            if let Some(d) = intercept_dim {
                // Cancel regularization for the intercept (sklearn :2051).
                w[d] = -inv_alpha;
            }

            // c = U · diag(w) · UT_y + alpha^-1 · y (sklearn :2052).
            // G_inverse_diag_i = sum_j w_j U[i,j]^2 + alpha^-1 (sklearn :2053).
            let mut sum_sq = <F as num_traits::Zero>::zero();
            for i in 0..n_samples {
                let mut c_i = <F as num_traits::Zero>::zero();
                let mut g_i = <F as num_traits::Zero>::zero();
                for j in 0..k {
                    let uij = u[(i, j)];
                    c_i += uij * (w[j] * ut_y[j]);
                    g_i += w[j] * uij * uij;
                }
                c_i += inv_alpha * y_c[i];
                g_i += inv_alpha;
                // looe_i = c_i / G_inverse_diag_i (sklearn :2149).
                let looe = c_i / g_i;
                sum_sq += looe * looe;
            }
            out.push(sum_sq / n_f);
        }
        Ok(out)
    }

    /// Eigen-mode GCV scores, used when `n_samples <= n_features` (sklearn
    /// `_eigen_decompose_gram` `_ridge.py:1900` then `_solve_eigen_gram`
    /// `_ridge.py:1914`). Returns the mean squared LOO error for each candidate
    /// alpha (lower is better).
    fn gcv_scores_eigen(&self, x_c: &Array2<F>, y_c: &Array1<F>) -> Result<Vec<F>, FerroError> {
        let n_samples = x_c.nrows();
        let one = <F as num_traits::One>::one();

        // Gram matrix K = X·Xᵀ on the centered design (sklearn `_compute_gram`,
        // dense → just `X X^T`, `_ridge.py:1799`).
        let mut k_mat = x_c.dot(&x_c.t());
        if self.fit_intercept {
            // Add outer(sqrt_sw, sqrt_sw): with uniform weights this is the
            // all-ones rank-1 matrix, emulating centering with the intercept
            // eigenvector (sklearn `_eigen_decompose_gram`, `_ridge.py:1909`).
            for i in 0..n_samples {
                for j in 0..n_samples {
                    k_mat[(i, j)] += one;
                }
            }
        }

        // Eigendecomposition K = Q diag(eigvals) Qᵀ (sklearn `linalg.eigh`,
        // `_ridge.py:1910`).
        let (eigvals, q) = eigh_sym(&k_mat)?;
        let m = eigvals.len();

        // QT_y[j] = sum_i Q[i,j] * y[i] (sklearn :1911).
        let mut qt_y = vec![<F as num_traits::Zero>::zero(); m];
        for j in 0..m {
            let mut acc = <F as num_traits::Zero>::zero();
            for i in 0..n_samples {
                acc += q[(i, j)] * y_c[i];
            }
            qt_y[j] = acc;
        }

        // Intercept eigenvector: the column of Q most aligned with the
        // normalized sqrt_sw (uniform → ones/√n) (sklearn :1926-1927).
        let intercept_dim = if self.fit_intercept {
            Some(find_intercept_dim(&q, n_samples, m))
        } else {
            None
        };

        let n_f = F::from(n_samples).unwrap_or(one);
        let mut out = Vec::with_capacity(self.alphas.len());
        for &alpha in &self.alphas {
            // w_j = 1 / (eigvals_j + alpha) (sklearn :1919).
            let mut w: Vec<F> = eigvals.iter().map(|&ev| one / (ev + alpha)).collect();
            if let Some(d) = intercept_dim {
                // Cancel regularization for the intercept (sklearn :1928).
                w[d] = <F as num_traits::Zero>::zero();
            }

            // c = Q · diag(w) · QT_y (sklearn :1930).
            // G_inverse_diag_i = sum_j w_j Q[i,j]^2 (sklearn :1931).
            let mut sum_sq = <F as num_traits::Zero>::zero();
            for i in 0..n_samples {
                let mut c_i = <F as num_traits::Zero>::zero();
                let mut g_i = <F as num_traits::Zero>::zero();
                for j in 0..m {
                    let qij = q[(i, j)];
                    c_i += qij * (w[j] * qt_y[j]);
                    g_i += w[j] * qij * qij;
                }
                // looe_i = c_i / G_inverse_diag_i (sklearn :2149).
                let looe = c_i / g_i;
                sum_sq += looe * looe;
            }
            out.push(sum_sq / n_f);
        }
        Ok(out)
    }
}

/// Find the column index of an orthonormal factor (`U` or `Q`, both
/// `(n_samples, k)`) that is most aligned with the normalized uniform-weight
/// vector `ones/√n`. Mirrors sklearn `_find_smallest_angle` (`_ridge.py:1579`):
/// the query and columns are unit vectors, so the most-aligned column maximises
/// `|query · column|`. With `query = ones/√n` the per-column dot product is
/// proportional to the column sum, so `|column-sum|` is the discriminant.
fn find_intercept_dim<F: Float + std::ops::AddAssign + 'static>(
    u: &Array2<F>,
    n_samples: usize,
    k: usize,
) -> usize {
    let mut best_idx = 0usize;
    let mut best_abs = F::neg_infinity();
    for j in 0..k {
        let mut col_sum = <F as num_traits::Zero>::zero();
        for i in 0..n_samples {
            col_sum += u[(i, j)];
        }
        let a = col_sum.abs();
        if a > best_abs {
            best_abs = a;
            best_idx = j;
        }
    }
    best_idx
}

/// Thin SVD via the ferray substrate, returning `(U, S)` as ndarray types.
///
/// Bridges `ndarray → ferray` for the decomposition and back (R-SUBSTRATE-4),
/// routing through [`ferray::linalg::svd`] (`ferray-linalg/src/decomp/svd.rs`).
fn svd_u_s<F: LinalgFloat>(a: &Array2<F>) -> Result<(Array2<F>, Array1<F>), FerroError> {
    let (rows, cols) = a.dim();
    let flat: Vec<F> = a.iter().copied().collect();
    let fa = FerrayArray::<F, Ix2>::from_vec(Ix2::new([rows, cols]), flat).map_err(|e| {
        FerroError::NumericalInstability {
            message: format!("RidgeCV GCV: failed to build design matrix for SVD: {e}"),
        }
    })?;
    let (u, s, _vt) =
        ferray::linalg::svd(&fa, false).map_err(|e| FerroError::NumericalInstability {
            message: format!("RidgeCV GCV: SVD failed: {e}"),
        })?;
    let u_nd = ferray_to_ndarray2(&u)?;
    let s_nd = ferray_to_ndarray1(&s)?;
    Ok((u_nd, s_nd))
}

/// Symmetric eigendecomposition via the ferray substrate, returning
/// `(eigvals, Q)` as ndarray types (ascending eigenvalues).
///
/// Bridges `ndarray → ferray` and back (R-SUBSTRATE-4), routing through
/// [`ferray::linalg::eigh`] (`ferray-linalg/src/decomp/eigen.rs`).
fn eigh_sym<F: LinalgFloat>(a: &Array2<F>) -> Result<(Array1<F>, Array2<F>), FerroError> {
    let (rows, cols) = a.dim();
    let flat: Vec<F> = a.iter().copied().collect();
    let fa = FerrayArray::<F, Ix2>::from_vec(Ix2::new([rows, cols]), flat).map_err(|e| {
        FerroError::NumericalInstability {
            message: format!("RidgeCV GCV: failed to build Gram matrix for eigh: {e}"),
        }
    })?;
    let (vals, q) = ferray::linalg::eigh(&fa).map_err(|e| FerroError::NumericalInstability {
        message: format!("RidgeCV GCV: eigendecomposition failed: {e}"),
    })?;
    let vals_nd = ferray_to_ndarray1(&vals)?;
    let q_nd = ferray_to_ndarray2(&q)?;
    Ok((vals_nd, q_nd))
}

/// Bridge a ferray 2-D array back to `ndarray::Array2` (R-SUBSTRATE-4).
fn ferray_to_ndarray2<F: LinalgFloat>(a: &FerrayArray<F, Ix2>) -> Result<Array2<F>, FerroError> {
    let shape = a.shape();
    let (rows, cols) = (shape[0], shape[1]);
    let nd = a.clone().into_ndarray();
    let flat: Vec<F> = nd.iter().copied().collect();
    Array2::from_shape_vec((rows, cols), flat).map_err(|e| FerroError::NumericalInstability {
        message: format!("RidgeCV GCV: ferray→ndarray (2-D) bridge failed: {e}"),
    })
}

/// Bridge a ferray 1-D array back to `ndarray::Array1` (R-SUBSTRATE-4).
fn ferray_to_ndarray1<F: LinalgFloat>(
    a: &FerrayArray<F, ferray::Ix1>,
) -> Result<Array1<F>, FerroError> {
    let nd = a.clone().into_ndarray();
    let flat: Vec<F> = nd.iter().copied().collect();
    Ok(Array1::from_vec(flat))
}

impl<F: Float + Send + Sync + ScalarOperand + 'static> Predict<Array2<F>> for FittedRidgeCV<F> {
    type Output = Array1<F>;
    type Error = FerroError;

    /// Predict target values for the given feature matrix.
    ///
    /// Computes `X @ coefficients + intercept`.
    ///
    /// # Errors
    ///
    /// Returns [`FerroError::ShapeMismatch`] if the number of features
    /// does not match the fitted model.
    fn predict(&self, x: &Array2<F>) -> Result<Array1<F>, FerroError> {
        let n_features = x.ncols();
        if n_features != self.coefficients.len() {
            return Err(FerroError::ShapeMismatch {
                expected: vec![self.coefficients.len()],
                actual: vec![n_features],
                context: "number of features must match fitted model".into(),
            });
        }

        let preds = x.dot(&self.coefficients) + self.intercept;
        Ok(preds)
    }
}

impl<F: Float + Send + Sync + ScalarOperand + 'static> HasCoefficients<F> for FittedRidgeCV<F> {
    fn coefficients(&self) -> &Array1<F> {
        &self.coefficients
    }

    fn intercept(&self) -> F {
        self.intercept
    }
}

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

    #[test]
    fn test_ridge_cv_default_builder() {
        let m = RidgeCV::<f64>::new();
        assert_eq!(m.alphas.len(), 3);
        // Default mirrors sklearn `RidgeCV(cv=None)` → LOO-GCV path.
        assert_eq!(m.cv, None);
        assert!(m.fit_intercept);
    }

    #[test]
    fn test_ridge_cv_with_cv_sets_kfold() {
        let m = RidgeCV::<f64>::new().with_cv(5);
        assert_eq!(m.cv, Some(5));
    }

    #[test]
    fn test_ridge_cv_custom_alphas() {
        let m = RidgeCV::<f64>::new().with_alphas(vec![0.01, 0.1, 1.0, 10.0, 100.0]);
        assert_eq!(m.alphas.len(), 5);
    }

    #[test]
    fn test_ridge_cv_fit_selects_alpha() {
        let x = Array2::from_shape_vec((20, 1), (1..=20).map(f64::from).collect()).unwrap();
        let y = Array1::from_iter((1..=20).map(|i| 2.0 * f64::from(i) + 1.0));

        let model = RidgeCV::<f64>::new()
            .with_alphas(vec![0.001, 0.01, 0.1, 1.0, 10.0, 100.0])
            .with_cv(5);

        let fitted = model.fit(&x, &y).unwrap();

        // For a clean linear relationship, a small alpha should win.
        assert!(fitted.best_alpha() <= 1.0);
    }

    #[test]
    fn test_ridge_cv_predict() {
        let x = Array2::from_shape_vec((10, 1), (1..=10).map(f64::from).collect()).unwrap();
        let y = Array1::from_iter((1..=10).map(|i| 2.0 * f64::from(i) + 1.0));

        let model = RidgeCV::<f64>::new().with_cv(3);
        let fitted = model.fit(&x, &y).unwrap();

        let preds = fitted.predict(&x).unwrap();
        assert_eq!(preds.len(), 10);

        // Predictions should be close to the true values.
        for i in 0..10 {
            assert_relative_eq!(preds[i], y[i], epsilon = 1.0);
        }
    }

    #[test]
    fn test_ridge_cv_has_coefficients() {
        let x = Array2::from_shape_vec((10, 2), (0..20).map(f64::from).collect()).unwrap();
        let y = Array1::from_iter((0..10).map(f64::from));

        let model = RidgeCV::<f64>::new().with_cv(3);
        let fitted = model.fit(&x, &y).unwrap();

        assert_eq!(fitted.coefficients().len(), 2);
    }

    #[test]
    fn test_ridge_cv_empty_alphas_error() {
        let x = Array2::from_shape_vec((3, 1), vec![1.0, 2.0, 3.0]).unwrap();
        let y = array![1.0, 2.0, 3.0];

        let model = RidgeCV::<f64>::new().with_alphas(vec![]);
        let result = model.fit(&x, &y);
        assert!(result.is_err());
    }

    #[test]
    fn test_ridge_cv_negative_alpha_error() {
        let x = Array2::from_shape_vec((3, 1), vec![1.0, 2.0, 3.0]).unwrap();
        let y = array![1.0, 2.0, 3.0];

        let model = RidgeCV::<f64>::new().with_alphas(vec![1.0, -0.5]);
        let result = model.fit(&x, &y);
        assert!(result.is_err());
    }

    #[test]
    fn test_ridge_cv_shape_mismatch() {
        let x = Array2::from_shape_vec((3, 1), vec![1.0, 2.0, 3.0]).unwrap();
        let y = array![1.0, 2.0];

        let model = RidgeCV::<f64>::new();
        let result = model.fit(&x, &y);
        assert!(result.is_err());
    }

    #[test]
    fn test_ridge_cv_insufficient_samples() {
        let x = Array2::from_shape_vec((2, 1), vec![1.0, 2.0]).unwrap();
        let y = array![1.0, 2.0];

        let model = RidgeCV::<f64>::new().with_cv(5);
        let result = model.fit(&x, &y);
        assert!(result.is_err());
    }

    #[test]
    fn test_ridge_cv_cv_too_small() {
        let x = Array2::from_shape_vec((10, 1), (1..=10).map(f64::from).collect()).unwrap();
        let y = Array1::from_iter((1..=10).map(f64::from));

        let model = RidgeCV::<f64>::new().with_cv(1);
        let result = model.fit(&x, &y);
        assert!(result.is_err());
    }

    #[test]
    fn test_ridge_cv_no_intercept() {
        let x = Array2::from_shape_vec((10, 1), (1..=10).map(f64::from).collect()).unwrap();
        let y = Array1::from_iter((1..=10).map(|i| 2.0 * f64::from(i)));

        let model = RidgeCV::<f64>::new().with_cv(3).with_fit_intercept(false);
        let fitted = model.fit(&x, &y).unwrap();

        // With no intercept and origin-passing data, predictions should be close.
        let preds = fitted.predict(&x).unwrap();
        assert_eq!(preds.len(), 10);
    }

    #[test]
    fn test_ridge_cv_predict_feature_mismatch() {
        let x_train = Array2::from_shape_vec((10, 2), (0..20).map(f64::from).collect()).unwrap();
        let y = Array1::from_iter((0..10).map(f64::from));

        let fitted = RidgeCV::<f64>::new().with_cv(3).fit(&x_train, &y).unwrap();

        let x_bad = Array2::from_shape_vec((3, 1), vec![1.0, 2.0, 3.0]).unwrap();
        let result = fitted.predict(&x_bad);
        assert!(result.is_err());
    }

    #[test]
    fn test_kfold_indices_coverage() {
        let folds = kfold_indices(10, 3);
        assert_eq!(folds.len(), 3);

        // Every index 0..10 should appear exactly once.
        let mut all: Vec<usize> = folds.into_iter().flatten().collect();
        all.sort();
        assert_eq!(all, (0..10).collect::<Vec<_>>());
    }

    /// `kfold_indices` must produce CONTIGUOUS folds matching sklearn
    /// `KFold(shuffle=False)` (`_split.py`, `_BaseKFold`), NOT round-robin.
    /// Oracle: `KFold(n_splits=k).split(arange(n))` test folds (live
    /// sklearn 1.5.2). The first `n % k` folds get one extra element.
    #[test]
    fn kfold_indices_contiguous_matches_sklearn() {
        // n=7, k=3 → first 1 fold gets the extra: [0,1,2],[3,4],[5,6].
        assert_eq!(
            kfold_indices(7, 3),
            vec![vec![0, 1, 2], vec![3, 4], vec![5, 6]]
        );
        // n=8, k=4 → n % k == 0, all equal: [0,1],[2,3],[4,5],[6,7].
        assert_eq!(
            kfold_indices(8, 4),
            vec![vec![0, 1], vec![2, 3], vec![4, 5], vec![6, 7]]
        );
        // n=10, k=3 → first 1 fold gets the extra: [0,1,2,3],[4,5,6],[7,8,9].
        assert_eq!(
            kfold_indices(10, 3),
            vec![vec![0, 1, 2, 3], vec![4, 5, 6], vec![7, 8, 9]]
        );
    }

    /// End-to-end `RidgeCV` with explicit `cv = Some(4)` on a fixed fixture
    /// (no RNG), verifying the sklearn-DEFAULT scoring contract: with
    /// `scoring=None` sklearn `RidgeCV(cv=k)` routes through
    /// `GridSearchCV(Ridge(), {'alpha': alphas}, cv=k, scoring=None)`
    /// (`_ridge.py:2429-2434`), whose default scorer is the regressor's R²
    /// (`RegressorMixin.score`), averaged per fold and MAXIMIZED. Oracle (live
    /// sklearn 1.5.2): `RidgeCV(alphas=[0.1,1.0,10.0], cv=4).fit(X, y)` →
    /// per-alpha averaged R² `{0.1: 0.279276, 1.0: 0.283149, 10.0: -0.593329}`
    /// → `alpha_ = 1.0`, `coef_ = [1.054289, 0.387522, 0.231554]`,
    /// `intercept_ = 0.44811`. The contiguous folds (`_ridge.py:2429`
    /// `GridSearchCV(cv=k)` → `KFold(shuffle=False)`, fixed in #429) drive the
    /// train/test partition; the R² maximization (NOT neg-MSE) drives the
    /// selection.
    #[test]
    fn ridge_cv_explicit_cv4_selects_alpha_matches_sklearn() -> Result<(), FerroError> {
        let x = Array2::from_shape_vec(
            (8, 3),
            vec![
                1.0, 0.0, 2.0, //
                0.0, 1.0, 1.0, //
                2.0, 1.0, 0.0, //
                1.0, 2.0, 1.0, //
                0.0, 0.0, 3.0, //
                3.0, 1.0, 2.0, //
                1.0, 1.0, 1.0, //
                2.0, 0.0, 1.0,
            ],
        )
        .map_err(|e| FerroError::NumericalInstability {
            message: format!("test fixture build failed: {e}"),
        })?;
        let y = array![2.0, 1.0, 3.0, 2.5, 1.0, 5.0, 2.0, 2.5];

        let model = RidgeCV::<f64>::new()
            .with_alphas(vec![0.1, 1.0, 10.0])
            .with_cv(4);
        let fitted = model.fit(&x, &y)?;

        assert_relative_eq!(fitted.best_alpha(), 1.0, epsilon = 1e-9);

        let coef = fitted.coefficients();
        assert_relative_eq!(coef[0], 1.054_289, epsilon = 1e-5);
        assert_relative_eq!(coef[1], 0.387_522, epsilon = 1e-5);
        assert_relative_eq!(coef[2], 0.231_554, epsilon = 1e-5);
        assert_relative_eq!(fitted.intercept(), 0.448_11, epsilon = 1e-5);

        Ok(())
    }

    /// `r2_score` must match sklearn `r2_score`
    /// (`sklearn/metrics/_regression.py:1220-1228` + edge cases at
    /// `_regression.py:872-891`) for the single-output case, including the
    /// `SS_tot == 0` edge: perfect prediction of a constant target → `1.0`,
    /// imperfect prediction of a constant target → `0.0`. Oracle (live sklearn
    /// 1.5.2): `r2_score([5,5],[5,5]) == 1.0`, `r2_score([5,5],[5,6]) == 0.0`,
    /// `r2_score([1,2,3],[1,2,3]) == 1.0`.
    #[test]
    fn r2_score_matches_sklearn_edge_cases() {
        // SS_tot == 0, SS_res == 0 → 1.0 (perfect prediction of constant y).
        assert_eq!(r2_score(&array![5.0, 5.0], &array![5.0, 5.0]), 1.0);
        // SS_tot == 0, SS_res != 0 → 0.0 (imperfect prediction of constant y).
        assert_eq!(r2_score(&array![5.0, 5.0], &array![5.0, 6.0]), 0.0);
        // Normal case: perfect prediction → 1.0.
        assert_eq!(
            r2_score(&array![1.0, 2.0, 3.0], &array![1.0, 2.0, 3.0]),
            1.0
        );
    }
}