ferrolearn-preprocess 0.5.0

Preprocessing transformers 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
//! Polynomial features: generate polynomial and interaction features.
//!
//! Given input features `[a, b]` and degree 2, this transformer generates:
//! - `[1, a, b, a², a·b, b²]` (default — `interaction_only = false`)
//! - `[1, a, b, a·b]` (with `interaction_only = true`)
//!
//! With `include_bias = false`, the constant column `1` is omitted.
//!
//! This transformer is stateless for value generation — call
//! [`Transform::transform`] directly. For scikit-learn API parity it ALSO
//! supports the stateful [`Fit`](ferrolearn_core::traits::Fit) →
//! [`FittedPolynomialFeatures`] path, which records `n_features_in_` /
//! `n_output_features_` / `powers_` and validates the input feature count in
//! `transform`; the fitted type's `transform` REUSES the very same value math as
//! the stateless path, so both paths are bit-identical to each other. That value
//! math (FIXED #2210) reproduces scikit-learn's INCREMENTAL degree-by-degree
//! column build (`_polynomial.py:525-564`) — each degree-`d` column is a stored
//! degree-`(d-1)` column times a single feature, in sklearn's `np.multiply` arg
//! order — so the polynomial VALUES are now bit-identical to the sklearn oracle
//! at EVERY degree, including degree≥3 (the prior independent per-column left-fold
//! diverged by 1 ULP on degree≥3 terms because FP multiplication is
//! non-associative).
//!
//! # `## REQ status`
//!
//! Binary (R-DEFER-2), translating `sklearn/preprocessing/_polynomial.py` (`class
//! PolynomialFeatures` `:99`). Design doc: `.design/preprocess/polynomial_features.md`. Expected
//! values from the live sklearn 1.5.2 oracle (R-CHAR-3). Consumers: `PipelineTransformer` impl +
//! crate re-export (`lib.rs`, grandfathered S5) + the `FittedPolynomialFeatures::transform` fitted
//! path. HONEST (R-HONEST-3): a dense int-degree transformer — the polynomial VALUES + column ORDER
//! match sklearn exactly, and the stateful `fit` → `FittedPolynomialFeatures` surface
//! (`n_features_in_`/`n_output_features_`/`powers_` + the transform feature-count check) SHIPS
//! (REQ-4/REQ-5), and the `ferrolearn-python` PyO3 binding (`_RsPolynomialFeatures` + the
//! `ferrolearn.PolynomialFeatures` stateful wrapper surfacing the fitted attributes) NOW SHIPS
//! (REQ-10); `get_feature_names_out`, degree-tuple, `order` (C/F memory layout), sparse, full-ctor,
//! ferray stay NOT-STARTED.
//!
//! | REQ | Status | Evidence |
//! |---|---|---|
//! | REQ-1 (int-degree dense values + column order) | SHIPPED | Column order: `fn feature_combinations` (bias=empty combo, DFS interaction_only strictly-increasing / else non-decreasing, sort by degree-then-lex) reproduces sklearn `_combinations` itertools order (`_polynomial.py:209-220`) column-for-column (this drives `powers_`/`n_output_features_`). Value math: `fn generate_poly_features` builds the columns DEGREE-BY-DEGREE via sklearn's incremental block recurrence (`_polynomial.py:525-564`) — each degree-`d` column = a stored degree-`(d-1)` column × a single feature in sklearn's `np.multiply` arg order — so values are bit-identical to the live oracle at EVERY degree. FIXED #2210: the prior independent per-column left-fold (`combo.iter().fold(1, |acc,j| acc*x[j])`) diverged by 1 ULP on degree≥3 terms (FP multiplication is non-associative); the incremental build now matches bit-for-bit (pin `divergence_polynomial_higher_degree_term_ulp_vs_sklearn`, x1*x2*x3 == `0x1.7b645a1cac082p+2`). Green guards (`tests/divergence_polynomial_features.rs`): 2-feat default `[1,2,3,4,6,9]`, 3-feat default `[1,2,3,5,4,6,10,9,15,25]`, 3-feat interaction `[1,2,3,5,6,10,15]`, deg-3 interaction no-bias `[2,3,5,6,10,15,30]`, multi-row. Consumers: `FittedPipelineTransformer` + re-export. |
//! | REQ-8 (input validation per check_array) | SHIPPED | FIXED #1180. `transform` guards (sklearn order) zero-samples → `InsufficientSamples`, zero-features → `InvalidParameter`, non-finite NaN/±inf → `InvalidParameter` — matching sklearn `transform` → `_validate_data` (`_polynomial.py:433-435`). Mirrors converged binarizer/normalizer. Critic two-round CLEAN: 13 tests incl. finite input (1e308) with overflowing product correctly ACCEPTED (input-only validation). |
//! | REQ-2 (degree tuple / min_degree) | NOT-STARTED | open prereq blocker #1181. Single `usize` degree, always starts at degree 1 (sklearn `:334-360`). |
//! | REQ-3 (order C/F param) | NOT-STARTED | open prereq blocker #1182. No `order` (sklearn `:201`,`:132-134`). |
//! | REQ-4 (stateful fit + n_features_in_/n_output_features_ + count check) | SHIPPED | FIXED #1183. `impl Fit<Array2<F>, ()> for PolynomialFeatures<F>` (`fit`): runs the SAME REQ-8 `validate_poly_input` guard (zero-samples → `InsufficientSamples`, zero-features/non-finite → `InvalidParameter`, sklearn `_validate_data` `_polynomial.py:323`), enumerates `feature_combinations(x.ncols())` (REUSED, not reimplemented), records `n_features_in_ = x.ncols()` and `n_output_features_ = combinations.len()` (== `_num_combinations`, sklearn `:362`), returns `FittedPolynomialFeatures`. `FittedPolynomialFeatures::transform` validates X (REQ-8) FIRST, generates via the SHARED `generate_poly_features` value math (delegated — byte-identical to the stateless path), THEN checks `x.ncols() != n_features_in_` → `ShapeMismatch` (sklearn `X has N features, but ... expecting M`, `:402-435`). The X-validation-before-n_features ORDER matches sklearn `_validate_data(reset=False)` → `check_array` before `_check_n_features` (#2207). Accessors `n_features_in()`/`n_output_features()`. Live-oracle tests: `fit_n_features_in_matches_ncols`, `fit_n_output_features_*`, `fitted_transform_matches_stateless_and_sklearn`, `fitted_transform_more/fewer_features_shape_mismatch`, `fitted_transform_nan_validation_before_n_features`. Consumers: `FittedPolynomialFeatures::transform` (the fitted path) + crate re-export `pub use polynomial_features::{PolynomialFeatures, FittedPolynomialFeatures}` (`lib.rs`). |
//! | REQ-5 (powers_ attribute) | SHIPPED | FIXED #1184. `FittedPolynomialFeatures<F>` carries `powers_: Array2<usize>` of shape `(n_output_features_, n_features_in_)` (matching `n_features_in_`'s `usize` storage idiom; sklearn `powers_` is int), built in `fit` from the SAME `feature_combinations` the value math uses (so row order == output column order): each combination → a `bincount` row counting occurrences of each feature index `0..n_features_in_`, the bias empty-combo → an all-zeros row (sklearn `np.vstack([np.bincount(c, minlength=n_features_in_) for c in combinations])`, `_polynomial.py:262-264`). Accessor `powers(&self) -> &Array2<usize>`. Live-oracle tests: `fit_powers_2feat/3feat_*` EXACT-match sklearn `.powers_` for default / `interaction_only` / `include_bias=False`, asserting shape + every entry. Consumer: the `powers_` accessor on the re-exported `FittedPolynomialFeatures`. |
//! | REQ-6 (get_feature_names_out) | NOT-STARTED | open prereq blocker #1185. None (sklearn `:266-303`). Depends on REQ-5. |
//! | REQ-7 (sparse CSR/CSC) | NOT-STARTED | open prereq blocker #1186. Dense-only (sklearn `:402-`,`:38-96`). |
//! | REQ-9 (full ctor + _parameter_constraints) | NOT-STARTED | open prereq blocker #1187. Positional `new`; validates only the `degree==0 && !include_bias` empty-output case (FIXED #2216: sklearn `_polynomial.py:326-330` rejects ONLY that case — `degree==0 && include_bias` is valid, emitting just the bias column). No keyword-only `*`-args, no `order`, no full `_parameter_constraints` (sklearn `:194-207`). |
//! | REQ-10 (PyO3 binding) | SHIPPED | FIXED #1188. `_RsPolynomialFeatures` (hand `#[pyclass]`, `ferrolearn-python/src/extras.rs`) over `ferrolearn_preprocess::{PolynomialFeatures, FittedPolynomialFeatures}<f64>` exposes `fit(x)` (ctor `PolynomialFeatures::new(degree, interaction_only, include_bias)`, `degree==0 && !include_bias`/NaN/±inf → `PyValueError`; FIXED #2216 `degree==0 && include_bias` is valid → bias-only column), `transform(x)` (`FittedPolynomialFeatures::transform` — wrong col count → `ShapeMismatch` → `PyValueError`), and `#[getter]`s `n_features_in_` (sklearn `_polynomial.py:323`), `n_output_features_` (`:362`), `powers_` (`:250-264`, `usize` `Array2` → numpy i64). Registered in `lib.rs`. The `_extras.py::PolynomialFeatures(_TransformerWrapper)` wrapper mirrors sklearn's ctor `(degree=2, *, interaction_only=False, include_bias=True, order="C")` (`_polynomial.py:201`, `degree` positional-or-keyword), is STATEFUL (overridden `fit` builds/fits `_rs`; pre-fit `transform`/attribute access → `NotFittedError` via `check_is_fitted`), does the FLOAT-only dtype cast-back (`check_array(dtype=FLOAT_DTYPES)`, `:432`: int→f64, float32→float32, float64→float64), and surfaces `n_features_in_`/`n_output_features_`/`powers_` as `@property`s reading from `_rs`. `order` is validated against sklearn's `StrOptions({"C","F"})` (bad string → `ValueError`); `'F'` is accepted-but-documented (the returned array is always C-contiguous — F memory-layout is REQ-3 #1182 NOT-STARTED — but the VALUES match sklearn's 'F' output, which is a layout-only knob). #2215-CLASS f64-ABI CAVEAT: for FLOAT32 input the polynomial products are computed in float64 (the f64 ABI) then cast back to float32 (sklearn computes them in float32), so the cast-back float32 VALUES can diverge from sklearn's by ~1e-6 on high-degree terms (the same reduced-precision caveat class as #2215/#2205) — float64 input is bit-exact; the float32 value pin uses a TOLERANT `atol=1e-5`. Exported in `__init__.py`. Non-test consumer: `ferrolearn.PolynomialFeatures(...).fit_transform(X)` + `lib.rs` `add_class` + `__init__.py` re-export. Live oracle: `tests/divergence_polynomial_features_py.py` (33 pass, sklearn 1.5.2) — fit_transform VALUES (d∈{2,3} × default/interaction_only/no-bias, float64 bit-exact), `powers_` int-exact, `n_output_features_`/`n_features_in_`, pre-fit `NotFittedError`, wrong-n_features `ValueError`, dtype int→f64/float32-tolerant/float64-bit-exact, get_params keys=={degree,interaction_only,include_bias,order}/defaults/clone/set_params, pipeline, order='F' values + bad-order `ValueError`. `get_feature_names_out` stays NOT-STARTED (REQ-6 #1185). |
//! | REQ-11 (ferray substrate) | NOT-STARTED | open prereq blocker #1189. `ndarray`+`num_traits`, not `ferray-core` (R-SUBSTRATE-1/2). |

use ferrolearn_core::error::FerroError;
use ferrolearn_core::pipeline::{FittedPipelineTransformer, PipelineTransformer};
use ferrolearn_core::traits::{Fit, Transform};
use ndarray::{Array1, Array2};
use num_traits::Float;

// ---------------------------------------------------------------------------
// PolynomialFeatures
// ---------------------------------------------------------------------------

/// A stateless polynomial feature generator.
///
/// Generates all polynomial combinations of the input features up to the
/// specified degree.
///
/// # Configuration
///
/// - `degree`: maximum polynomial degree (default `2`).
/// - `interaction_only`: if `true`, only cross-product terms are generated
///   (no pure powers like `a²`). Default `false`.
/// - `include_bias`: if `true`, a constant column of ones is prepended.
///   Default `true`.
///
/// # Examples
///
/// ```
/// use ferrolearn_preprocess::polynomial_features::PolynomialFeatures;
/// use ferrolearn_core::traits::Transform;
/// use ndarray::array;
///
/// let poly = PolynomialFeatures::<f64>::new(2, false, true).unwrap();
/// let x = array![[2.0, 3.0]];
/// let out = poly.transform(&x).unwrap();
/// // out = [[1, 2, 3, 4, 6, 9]]
/// ```
#[derive(Debug, Clone)]
pub struct PolynomialFeatures<F> {
    /// Maximum polynomial degree.
    pub(crate) degree: usize,
    /// If `true`, only interaction terms are produced (no pure powers).
    pub(crate) interaction_only: bool,
    /// If `true`, prepend a bias (constant ones) column.
    pub(crate) include_bias: bool,
    _marker: std::marker::PhantomData<F>,
}

impl<F: Float + Send + Sync + 'static> PolynomialFeatures<F> {
    /// Create a new `PolynomialFeatures` transformer.
    ///
    /// # Errors
    ///
    /// Returns [`FerroError::InvalidParameter`] if `degree == 0 && !include_bias`.
    /// `degree == 0` is ALLOWED when `include_bias == true` — the output is a
    /// single all-ones bias column (`powers_ == [[0,..,0]]`,
    /// `n_output_features_ == 1`), mirroring scikit-learn, which only rejects the
    /// `degree == 0 && include_bias == False` case
    /// (`sklearn/preprocessing/_polynomial.py:326-330`: "Setting degree to zero
    /// and include_bias to False would result in an empty output array.").
    pub fn new(
        degree: usize,
        interaction_only: bool,
        include_bias: bool,
    ) -> Result<Self, FerroError> {
        // sklearn `_polynomial.py:326-330`: ONLY reject degree==0 when
        // include_bias is False (an empty output array); degree==0 with
        // include_bias=True is valid and emits just the bias column.
        if degree == 0 && !include_bias {
            return Err(FerroError::InvalidParameter {
                name: "degree".into(),
                reason: "Setting degree to zero and include_bias to False would \
                         result in an empty output array."
                    .into(),
            });
        }
        Ok(Self {
            degree,
            interaction_only,
            include_bias,
            _marker: std::marker::PhantomData,
        })
    }

    /// Create a `PolynomialFeatures` with default settings:
    /// degree=2, interaction_only=false, include_bias=true.
    #[must_use]
    pub fn default_config() -> Self {
        Self {
            degree: 2,
            interaction_only: false,
            include_bias: true,
            _marker: std::marker::PhantomData,
        }
    }

    /// Return the configured degree.
    #[must_use]
    pub fn degree(&self) -> usize {
        self.degree
    }

    /// Return whether only interaction terms are generated.
    #[must_use]
    pub fn interaction_only(&self) -> bool {
        self.interaction_only
    }

    /// Return whether a bias column is included.
    #[must_use]
    pub fn include_bias(&self) -> bool {
        self.include_bias
    }

    /// Generate all combinations (with repetition unless `interaction_only`)
    /// of feature indices up to `degree`.
    ///
    /// Returns a list of index-tuples, where each tuple specifies which
    /// feature indices to multiply together to produce one output column.
    fn feature_combinations(&self, n_features: usize) -> Vec<Vec<usize>> {
        let mut combos: Vec<Vec<usize>> = Vec::new();

        // Bias term: empty product = 1
        if self.include_bias {
            combos.push(vec![]);
        }

        // Generate combinations of degrees 1..=self.degree
        let mut stack: Vec<(Vec<usize>, usize)> = Vec::new();

        // Start with each feature at degree 1. For `degree == 0` there are NO
        // degree>=1 combinations (sklearn `_combinations` ranges over
        // `range(max(1, min_degree), max_degree + 1)` = `range(1, 1)` = empty,
        // `_polynomial.py:215-217`), so only the bias empty-combo (if
        // `include_bias`) is emitted.
        if self.degree >= 1 {
            for i in 0..n_features {
                stack.push((vec![i], i));
            }
        }

        while let Some((combo, last_idx)) = stack.pop() {
            combos.push(combo.clone());

            if combo.len() < self.degree {
                // Extend with another feature
                let start = if self.interaction_only {
                    // Strictly increasing indices — no repeated features
                    last_idx + 1
                } else {
                    // Non-decreasing indices — allows repeated features (pure powers)
                    last_idx
                };
                for i in start..n_features {
                    let mut new_combo = combo.clone();
                    new_combo.push(i);
                    stack.push((new_combo, i));
                }
            }
        }

        // Sort: bias first, then by combo length, then lexicographically
        combos.sort_by(|a, b| a.len().cmp(&b.len()).then_with(|| a.cmp(b)));

        combos
    }
}

/// Run the shared `check_array` input validation (REQ-8) used by the stateless
/// [`Transform::transform`], the stateful [`Fit::fit`], and the
/// [`FittedPolynomialFeatures`] transform path, in sklearn's `check_array`
/// order: zero-samples → zero-features → non-finite
/// (`sklearn/utils/validation.py:1084`, `:1093`, `:1063`). Mirrors sklearn
/// `PolynomialFeatures.fit`/`.transform` → `_validate_data`
/// (`_polynomial.py:323`, `:433-435`), whose default `force_all_finite=True`
/// REJECTS NaN/±inf.
///
/// `context` names the calling site for diagnostics (e.g.
/// `"PolynomialFeatures::transform"` vs `"PolynomialFeatures::fit"`).
fn validate_poly_input<F: Float>(x: &Array2<F>, context: &str) -> Result<(), FerroError> {
    if x.nrows() == 0 {
        return Err(FerroError::InsufficientSamples {
            required: 1,
            actual: 0,
            context: context.into(),
        });
    }
    if x.ncols() == 0 {
        return Err(FerroError::InvalidParameter {
            name: "X".to_string(),
            reason: "Found array with 0 feature(s); a minimum of 1 is required \
                     by PolynomialFeatures"
                .to_string(),
        });
    }
    if x.iter().any(|v| !v.is_finite()) {
        return Err(FerroError::InvalidParameter {
            name: "X".to_string(),
            reason: "Input X contains non-finite values (NaN or infinity); \
                     PolynomialFeatures requires all-finite input"
                .to_string(),
        });
    }
    Ok(())
}

/// Generate the dense polynomial feature matrix (REQ-1 value math), shared by
/// the stateless [`Transform::transform`] and the stateful
/// [`FittedPolynomialFeatures`] transform path so the two are byte-identical.
///
/// Builds the output columns DEGREE-BY-DEGREE using scikit-learn's incremental
/// block recurrence (`_polynomial.py:525-564`), NOT an independent per-column
/// left-fold: the degree-1 block is `X` (copied in feature order); each degree-`d`
/// block is built feature-by-feature, where for `feature_idx` the new columns are
/// the just-built degree-`(d-1)` columns in slice `[start, end)` each multiplied
/// by `X[:, feature_idx]` — i.e. `previous_degree_column * X[:, feature_idx]` in
/// THAT multiply order, matching `np.multiply(XP[:, start:end], X[:, feature_idx],
/// casting="no")` bit-for-bit (FP multiplication is non-associative, so the
/// association order is load-bearing for the last ULP on degree≥3 terms). The
/// resulting column order is exactly sklearn's (bias, then degree-1, then each
/// degree-`d` block by `feature_idx`), which equals the [`feature_combinations`]
/// order the `powers_`/REQ-1 guards pin. The bias column (if `include_bias`) is
/// `1`. The input is assumed already validated by [`validate_poly_input`].
fn generate_poly_features<F: Float>(
    x: &Array2<F>,
    degree: usize,
    interaction_only: bool,
    include_bias: bool,
) -> Array2<F> {
    let n_samples = x.nrows();
    let n_features = x.ncols();
    let n_out = num_output_columns(n_features, degree, interaction_only, include_bias);
    let mut out = Array2::zeros((n_samples, n_out));

    // degree 0 (bias) term
    let mut current_col = 0usize;
    if include_bias {
        for i in 0..n_samples {
            out[[i, 0]] = F::one();
        }
        current_col = 1;
    }

    if n_features == 0 {
        return out;
    }

    if degree == 0 {
        // sklearn `_polynomial.py:532-533`: for `_max_degree == 0` the build
        // returns right after the bias column — NO degree-1 block, NO higher
        // blocks. degree==0 here implies include_bias (new() rejects
        // degree==0 && !include_bias), so `out` already holds the single
        // all-ones bias column.
        return out;
    }

    // degree 1 block: copy X column-for-column.
    // `index[j]` is the output column where degree-(d-1) block for feature `j`
    // begins; the trailing `index[n_features]` marks one past the block end.
    let mut index: Vec<usize> = Vec::with_capacity(n_features + 1);
    for j in 0..n_features {
        index.push(current_col);
        for i in 0..n_samples {
            out[[i, current_col]] = x[[i, j]];
        }
        current_col += 1;
    }
    index.push(current_col);

    // degree >= 2 blocks, built incrementally from the previous degree block.
    for _ in 2..=degree {
        let mut new_index: Vec<usize> = Vec::with_capacity(n_features + 1);
        // sklearn `end = index[-1]` (`_polynomial.py:544`): the LAST element of
        // `index`, NOT `index[n_features]`. When `interaction_only` runs out of
        // combinations at a high degree the inner loop `break`s early, leaving
        // `index` (the previous `new_index`) with FEWER than `n_features+1`
        // entries — so indexing `[n_features]` would panic (#2211, R-CODE-2). The
        // last element is the correct degree-(d-1) block end either way.
        let end = index.last().copied().unwrap_or(0);
        for feature_idx in 0..n_features {
            let mut start = index[feature_idx];
            new_index.push(current_col);
            if interaction_only {
                start += index[feature_idx + 1] - index[feature_idx];
            }
            if end <= start {
                // next_col <= current_col → no new columns for this feature.
                break;
            }
            // new column = (stored degree-(d-1) column) * X[:, feature_idx],
            // in that np.multiply arg order (prev FIRST, single feature SECOND).
            for src_col in start..end {
                for i in 0..n_samples {
                    out[[i, current_col]] = out[[i, src_col]] * x[[i, feature_idx]];
                }
                current_col += 1;
            }
        }
        new_index.push(current_col);
        index = new_index;
    }

    out
}

/// Number of output columns for the incremental build (== number of
/// [`feature_combinations`] entries / `n_output_features_`): the count of
/// degree-`1..=degree` combinations (with repetition unless `interaction_only`)
/// over `n_features` features, plus one for the bias if `include_bias`.
fn num_output_columns(
    n_features: usize,
    degree: usize,
    interaction_only: bool,
    include_bias: bool,
) -> usize {
    let mut total = usize::from(include_bias);
    if n_features == 0 {
        return total;
    }
    if degree == 0 {
        // sklearn `_combinations` emits no degree>=1 terms for degree==0
        // (`range(max(1,0), 0+1)` = `range(1,1)` = empty); the only output
        // column is the bias (`_polynomial.py:215-219`). degree==0 reaching here
        // implies include_bias (new() rejects degree==0 && !include_bias).
        return total;
    }
    // Count combinations per degree by mirroring the incremental block sizes.
    // degree-1 block has `n_features` columns; subsequent block sizes follow the
    // same start/end recurrence as the value build.
    let mut index: Vec<usize> = (0..=n_features).collect();
    total += n_features;
    let mut current = n_features;
    for _ in 2..=degree {
        let mut new_index: Vec<usize> = Vec::with_capacity(n_features + 1);
        // sklearn `index[-1]` (the last element), not `index[n_features]` —
        // robust to the `interaction_only` high-degree early break (#2211).
        let end = index.last().copied().unwrap_or(0);
        for feature_idx in 0..n_features {
            let mut start = index[feature_idx];
            new_index.push(current);
            if interaction_only {
                start += index[feature_idx + 1] - index[feature_idx];
            }
            if end <= start {
                break;
            }
            current += end - start;
            total += end - start;
        }
        new_index.push(current);
        index = new_index;
    }
    total
}

/// Build the `powers_` exponent matrix from a precomputed combination list
/// (REQ-5). Returns an `Array2<usize>` of shape `(combos.len(), n_features)`
/// where `powers_[k, j]` is the number of occurrences of input feature `j` in
/// combination `k` (its exponent) — sklearn's
/// `np.vstack([np.bincount(c, minlength=n_features_in_) for c in combinations])`
/// (`_polynomial.py:262-264`). The bias empty-combo → an all-zeros row. Built
/// from the SAME combinations the value math uses, so the row order equals the
/// output column order.
fn build_powers(combos: &[Vec<usize>], n_features: usize) -> Array2<usize> {
    let mut powers = Array2::<usize>::zeros((combos.len(), n_features));
    for (k, combo) in combos.iter().enumerate() {
        for &j in combo {
            powers[[k, j]] += 1;
        }
    }
    powers
}

impl<F: Float + Send + Sync + 'static> Default for PolynomialFeatures<F> {
    fn default() -> Self {
        Self::default_config()
    }
}

// ---------------------------------------------------------------------------
// Trait implementations
// ---------------------------------------------------------------------------

impl<F: Float + Send + Sync + 'static> Transform<Array2<F>> for PolynomialFeatures<F> {
    type Output = Array2<F>;
    type Error = FerroError;

    /// Generate polynomial and interaction features.
    ///
    /// # Errors
    ///
    /// Mirrors scikit-learn's `PolynomialFeatures.transform`
    /// (`sklearn/preprocessing/_polynomial.py:433`, `self._validate_data(...)` ->
    /// `check_array`), validating in `check_array` order
    /// (samples -> features -> finite):
    ///
    /// - Returns [`FerroError::InsufficientSamples`] if `x` has zero rows
    ///   (`check_array` `ensure_min_samples=1`; sklearn raises `ValueError:
    ///   Found array with 0 sample(s) ... a minimum of 1 is required`).
    /// - Returns [`FerroError::InvalidParameter`] if `x` has zero columns
    ///   (`check_array` `ensure_min_features=1`; sklearn raises `ValueError:
    ///   Found array with 0 feature(s) ... a minimum of 1 is required`).
    /// - Returns [`FerroError::InvalidParameter`] if `x` contains any non-finite
    ///   value (NaN, +inf, or -inf; `check_array` `force_all_finite=True`;
    ///   sklearn raises `ValueError: Input X contains NaN` / `infinity ...`).
    fn transform(&self, x: &Array2<F>) -> Result<Array2<F>, FerroError> {
        validate_poly_input(x, "PolynomialFeatures::transform")?;
        Ok(generate_poly_features(
            x,
            self.degree,
            self.interaction_only,
            self.include_bias,
        ))
    }
}

// ---------------------------------------------------------------------------
// FittedPolynomialFeatures (sklearn stateful `fit` -> fitted estimator path)
// ---------------------------------------------------------------------------

/// A fitted [`PolynomialFeatures`].
///
/// `PolynomialFeatures.fit` (sklearn `_polynomial.py:306-400`, "Compute number
/// of output features") learns NO numeric statistics — it records only the
/// shape metadata derivable from `n_features_in_` and the configured params:
/// `n_features_in_`, `n_output_features_` (the number of output columns), and
/// the `powers_` exponent matrix. The fitted type's [`Transform::transform`]
/// REUSES the very same combination enumeration ([`PolynomialFeatures::
/// feature_combinations`]) and value math ([`generate_poly_features`]) as the
/// stateless [`PolynomialFeatures`] path, so the two paths are bit-identical;
/// it additionally enforces the transform-time feature-count check against
/// `n_features_in_` (sklearn `X has N features, but ... expecting M`).
#[derive(Debug, Clone)]
pub struct FittedPolynomialFeatures<F> {
    /// Maximum polynomial degree (carried from the unfitted [`PolynomialFeatures`]).
    pub(crate) degree: usize,
    /// If `true`, only interaction terms are produced (no pure powers).
    pub(crate) interaction_only: bool,
    /// If `true`, a bias (constant ones) column is prepended.
    pub(crate) include_bias: bool,
    /// Number of features (columns) seen during [`Fit::fit`] — sklearn's
    /// `n_features_in_` (`_polynomial.py:323`, set by `_validate_data`).
    pub(crate) n_features_in_: usize,
    /// Total number of polynomial output columns — sklearn's
    /// `n_output_features_` (`_polynomial.py:362`, == `_num_combinations`).
    pub(crate) n_output_features_: usize,
    /// The exponent matrix of shape `(n_output_features_, n_features_in_)`:
    /// `powers_[i, j]` is the exponent of input feature `j` in output feature
    /// `i` — sklearn's `powers_` (`_polynomial.py:250-264`). The bias empty-combo
    /// row is all zeros.
    pub(crate) powers_: Array2<usize>,
    _marker: std::marker::PhantomData<F>,
}

impl<F: Float + Send + Sync + 'static> FittedPolynomialFeatures<F> {
    /// Return the number of features (columns) seen during [`Fit::fit`].
    ///
    /// Mirrors scikit-learn's `PolynomialFeatures.n_features_in_`
    /// (`_polynomial.py:323`).
    #[must_use]
    pub fn n_features_in(&self) -> usize {
        self.n_features_in_
    }

    /// Return the total number of polynomial output columns.
    ///
    /// Mirrors scikit-learn's `PolynomialFeatures.n_output_features_`
    /// (`_polynomial.py:362`).
    #[must_use]
    pub fn n_output_features(&self) -> usize {
        self.n_output_features_
    }

    /// Return the `powers_` exponent matrix of shape
    /// `(n_output_features_, n_features_in_)`.
    ///
    /// `powers_[i, j]` is the exponent of input feature `j` in output feature
    /// `i`, mirroring scikit-learn's `PolynomialFeatures.powers_`
    /// (`_polynomial.py:250-264`).
    #[must_use]
    pub fn powers(&self) -> &Array2<usize> {
        &self.powers_
    }

    /// Return the configured degree.
    #[must_use]
    pub fn degree(&self) -> usize {
        self.degree
    }

    /// Return whether only interaction terms are generated.
    #[must_use]
    pub fn interaction_only(&self) -> bool {
        self.interaction_only
    }

    /// Return whether a bias column is included.
    #[must_use]
    pub fn include_bias(&self) -> bool {
        self.include_bias
    }
}

impl<F: Float + Send + Sync + 'static> Fit<Array2<F>, ()> for PolynomialFeatures<F> {
    type Fitted = FittedPolynomialFeatures<F>;
    type Error = FerroError;

    /// Validate the input and record `n_features_in_` / `n_output_features_` /
    /// `powers_`, returning a [`FittedPolynomialFeatures`].
    ///
    /// `PolynomialFeatures.fit` (sklearn `_polynomial.py:306-400`, "Compute
    /// number of output features") learns NO numeric statistics. This runs the
    /// SAME REQ-8 `check_array` validation as [`Transform::transform`] (via the
    /// shared [`validate_poly_input`] helper; sklearn's `_validate_data` default
    /// `force_all_finite=True` REJECTS NaN/±inf), enumerates the polynomial
    /// combinations by REUSING [`PolynomialFeatures::feature_combinations`] (no
    /// reimplementation), and records `n_features_in_ = x.ncols()`,
    /// `n_output_features_ = combinations.len()` (== sklearn's
    /// `_num_combinations`, `:362`), and the `powers_` exponent matrix
    /// (`build_powers` = sklearn's `np.bincount` over the combinations, `:262`).
    /// `powers_` is built from the SAME combinations the transform value math
    /// uses, so its row order equals the output column order.
    ///
    /// # Errors
    ///
    /// Returns [`FerroError::InsufficientSamples`] for zero rows and
    /// [`FerroError::InvalidParameter`] for zero features or any non-finite
    /// value (NaN, +inf, -inf) — matching `check_array`
    /// (`sklearn/utils/validation.py:1084`, `:1093`, `:1063`) as routed through
    /// `PolynomialFeatures.fit` → `_validate_data` (`_polynomial.py:323`).
    fn fit(&self, x: &Array2<F>, _y: &()) -> Result<FittedPolynomialFeatures<F>, FerroError> {
        validate_poly_input(x, "PolynomialFeatures::fit")?;
        let n_features_in_ = x.ncols();
        // REUSE the SHIPPED combination enumeration (REQ-1) — do NOT reimplement.
        let combos = self.feature_combinations(n_features_in_);
        let n_output_features_ = combos.len();
        let powers_ = build_powers(&combos, n_features_in_);
        Ok(FittedPolynomialFeatures {
            degree: self.degree,
            interaction_only: self.interaction_only,
            include_bias: self.include_bias,
            n_features_in_,
            n_output_features_,
            powers_,
            _marker: std::marker::PhantomData,
        })
    }
}

impl<F: Float + Send + Sync + 'static> Transform<Array2<F>> for FittedPolynomialFeatures<F> {
    type Output = Array2<F>;
    type Error = FerroError;

    /// Generate polynomial features, delegating to the SAME combination
    /// enumeration + value math as the stateless [`PolynomialFeatures`] path.
    ///
    /// First applies the REQ-8 `check_array` guards (min-samples / min-features
    /// / finite) and generates the polynomial matrix, THEN validates that `x`
    /// has the same number of columns recorded during [`Fit::fit`]. This ORDER
    /// matches sklearn's `_validate_data(reset=False)`, which runs `check_array`
    /// BEFORE the `n_features_in_` consistency check (`_polynomial.py:433-435`,
    /// #2207): a NaN / ±inf / zero-sample / zero-feature input raises its
    /// `check_array` error EVEN when the column count is also wrong. Only after
    /// that does the feature-count comparison fire. Because the combinations are
    /// re-enumerated from the SAME params (degree / interaction_only /
    /// include_bias) and the value math is the shared [`generate_poly_features`],
    /// the output is byte-identical to `PolynomialFeatures::transform`.
    ///
    /// # Errors
    ///
    /// Returns [`FerroError::ShapeMismatch`] if the column count differs from
    /// `n_features_in_` (sklearn `ValueError: X has N features, but
    /// PolynomialFeatures is expecting M features as input.`,
    /// `_polynomial.py:402-435`). Returns [`FerroError::InsufficientSamples`]
    /// for zero rows and [`FerroError::InvalidParameter`] for zero features or
    /// any non-finite value (REQ-8, via [`validate_poly_input`]).
    fn transform(&self, x: &Array2<F>) -> Result<Array2<F>, FerroError> {
        // sklearn `_validate_data(reset=False)` runs `check_array` (finite /
        // min-samples / min-features) BEFORE the `n_features_in_` check (#2207).
        // So validate + generate FIRST; a NaN / ±inf / zero-sample / zero-feature
        // input must raise its check_array error EVEN when the column count is
        // also wrong. Only then does the feature-count comparison fire.
        validate_poly_input(x, "FittedPolynomialFeatures::transform")?;
        // REUSE the SHARED incremental value math — build from the configured
        // params so the column order matches the stateless path (and `powers_`)
        // exactly.
        let out = generate_poly_features(x, self.degree, self.interaction_only, self.include_bias);
        if x.ncols() != self.n_features_in_ {
            return Err(FerroError::ShapeMismatch {
                expected: vec![x.nrows(), self.n_features_in_],
                actual: vec![x.nrows(), x.ncols()],
                context: "FittedPolynomialFeatures::transform".into(),
            });
        }
        Ok(out)
    }
}

// ---------------------------------------------------------------------------
// Pipeline integration (generic)
// ---------------------------------------------------------------------------

impl<F: Float + Send + Sync + 'static> PipelineTransformer<F> for PolynomialFeatures<F> {
    /// Fit the polynomial features transformer using the pipeline interface.
    ///
    /// Because `PolynomialFeatures` is stateless, this simply boxes `self`
    /// as a [`FittedPipelineTransformer`].
    ///
    /// # Errors
    ///
    /// This implementation never returns an error.
    fn fit_pipeline(
        &self,
        _x: &Array2<F>,
        _y: &Array1<F>,
    ) -> Result<Box<dyn FittedPipelineTransformer<F>>, FerroError> {
        Ok(Box::new(self.clone()))
    }
}

impl<F: Float + Send + Sync + 'static> FittedPipelineTransformer<F> for PolynomialFeatures<F> {
    /// Transform data using the pipeline interface.
    ///
    /// # Errors
    ///
    /// Propagates errors from [`Transform::transform`].
    fn transform_pipeline(&self, x: &Array2<F>) -> Result<Array2<F>, FerroError> {
        self.transform(x)
    }
}

// ---------------------------------------------------------------------------
// Tests
// ---------------------------------------------------------------------------

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

    #[test]
    fn test_degree2_two_features_with_bias() {
        // degree=2, interaction_only=false, include_bias=true
        // Expected: [1, a, b, a², a·b, b²]
        let poly = PolynomialFeatures::<f64>::new(2, false, true).unwrap();
        let x = array![[2.0, 3.0]];
        let out = poly.transform(&x).unwrap();
        assert_eq!(out.shape()[0], 1);
        assert_eq!(out.shape()[1], 6); // 1 + 2 + 3 combinations
        assert_abs_diff_eq!(out[[0, 0]], 1.0, epsilon = 1e-10); // bias
        assert_abs_diff_eq!(out[[0, 1]], 2.0, epsilon = 1e-10); // a
        assert_abs_diff_eq!(out[[0, 2]], 3.0, epsilon = 1e-10); // b
        assert_abs_diff_eq!(out[[0, 3]], 4.0, epsilon = 1e-10); // a²
        assert_abs_diff_eq!(out[[0, 4]], 6.0, epsilon = 1e-10); // a·b
        assert_abs_diff_eq!(out[[0, 5]], 9.0, epsilon = 1e-10); // b²
    }

    #[test]
    fn test_degree2_interaction_only() {
        // degree=2, interaction_only=true, include_bias=true
        // Expected: [1, a, b, a·b]
        let poly = PolynomialFeatures::<f64>::new(2, true, true).unwrap();
        let x = array![[2.0, 3.0]];
        let out = poly.transform(&x).unwrap();
        assert_eq!(out.shape()[1], 4);
        assert_abs_diff_eq!(out[[0, 0]], 1.0, epsilon = 1e-10); // bias
        assert_abs_diff_eq!(out[[0, 1]], 2.0, epsilon = 1e-10); // a
        assert_abs_diff_eq!(out[[0, 2]], 3.0, epsilon = 1e-10); // b
        assert_abs_diff_eq!(out[[0, 3]], 6.0, epsilon = 1e-10); // a·b
    }

    #[test]
    fn test_no_bias() {
        // degree=2, interaction_only=false, include_bias=false
        // Expected: [a, b, a², a·b, b²]
        let poly = PolynomialFeatures::<f64>::new(2, false, false).unwrap();
        let x = array![[2.0, 3.0]];
        let out = poly.transform(&x).unwrap();
        assert_eq!(out.shape()[1], 5);
        assert_abs_diff_eq!(out[[0, 0]], 2.0, epsilon = 1e-10); // a
    }

    #[test]
    fn test_degree1_only_linear() {
        let poly = PolynomialFeatures::<f64>::new(1, false, true).unwrap();
        let x = array![[2.0, 3.0]];
        let out = poly.transform(&x).unwrap();
        // [1, a, b]
        assert_eq!(out.shape()[1], 3);
        assert_abs_diff_eq!(out[[0, 0]], 1.0, epsilon = 1e-10);
        assert_abs_diff_eq!(out[[0, 1]], 2.0, epsilon = 1e-10);
        assert_abs_diff_eq!(out[[0, 2]], 3.0, epsilon = 1e-10);
    }

    #[test]
    fn test_multiple_rows() {
        let poly = PolynomialFeatures::<f64>::new(2, false, false).unwrap();
        let x = array![[1.0, 2.0], [3.0, 4.0]];
        let out = poly.transform(&x).unwrap();
        assert_eq!(out.shape(), &[2, 5]);
        // Row 0: a=1, b=2 → [1, 2, 1, 2, 4]
        assert_abs_diff_eq!(out[[0, 0]], 1.0, epsilon = 1e-10);
        assert_abs_diff_eq!(out[[0, 1]], 2.0, epsilon = 1e-10);
        assert_abs_diff_eq!(out[[0, 2]], 1.0, epsilon = 1e-10);
        assert_abs_diff_eq!(out[[0, 3]], 2.0, epsilon = 1e-10);
        assert_abs_diff_eq!(out[[0, 4]], 4.0, epsilon = 1e-10);
    }

    #[test]
    fn test_single_feature_degree2() {
        // [a] → [1, a, a²]
        let poly = PolynomialFeatures::<f64>::new(2, false, true).unwrap();
        let x = array![[3.0]];
        let out = poly.transform(&x).unwrap();
        assert_eq!(out.shape()[1], 3);
        assert_abs_diff_eq!(out[[0, 0]], 1.0, epsilon = 1e-10);
        assert_abs_diff_eq!(out[[0, 1]], 3.0, epsilon = 1e-10);
        assert_abs_diff_eq!(out[[0, 2]], 9.0, epsilon = 1e-10);
    }

    #[test]
    fn test_degree0_include_bias_false_is_err() {
        // sklearn `_polynomial.py:326-330`: degree==0 with include_bias=False is
        // the ONLY degree==0 case that raises ("Setting degree to zero and
        // include_bias to False would result in an empty output array.").
        assert!(PolynomialFeatures::<f64>::new(0, false, false).is_err());
        assert!(PolynomialFeatures::<f64>::new(0, true, false).is_err());
    }

    #[test]
    fn test_degree0_include_bias_true_bias_only() -> Result<(), FerroError> {
        // sklearn ORACLE (degree=0, include_bias=True): a single all-ones bias
        // column, powers_ == [[0, 0]] (one all-zeros row), n_output_features_ == 1.
        //   python3 -c "import numpy as np; from sklearn.preprocessing import \
        //   PolynomialFeatures as P; m=P(degree=0, include_bias=True).fit( \
        //   np.array([[2.,3.]])); print(m.transform(np.array([[2.,3.]])).tolist(), \
        //   m.powers_.tolist(), m.n_output_features_)"
        //   -> [[1.0]] [[0, 0]] 1
        let poly = PolynomialFeatures::<f64>::new(0, false, true)?;
        let x = array![[2.0, 3.0]];

        // Stateless transform → single all-ones bias column.
        let out = poly.transform(&x)?;
        assert_eq!(out.shape(), &[1, 1]);
        assert_abs_diff_eq!(out[[0, 0]], 1.0, epsilon = 1e-10);

        // Fitted attributes: powers_ == [[0, 0]], n_output_features_ == 1.
        let fitted = poly.fit(&x, &())?;
        assert_eq!(fitted.n_output_features(), 1);
        assert_eq!(fitted.powers().shape(), &[1, 2]);
        assert_eq!(fitted.powers()[[0, 0]], 0);
        assert_eq!(fitted.powers()[[0, 1]], 0);

        // Fitted transform is byte-identical to the stateless path.
        let out_fitted = fitted.transform(&x)?;
        assert_eq!(out_fitted, out);
        Ok(())
    }

    #[test]
    fn test_default_config() {
        let poly = PolynomialFeatures::<f64>::default();
        assert_eq!(poly.degree(), 2);
        assert!(!poly.interaction_only());
        assert!(poly.include_bias());
    }

    #[test]
    fn test_pipeline_integration() {
        use ferrolearn_core::pipeline::PipelineTransformer;
        let poly = PolynomialFeatures::<f64>::new(2, false, true).unwrap();
        let x = array![[1.0, 2.0], [3.0, 4.0]];
        let y = Array1::zeros(2);
        let fitted = poly.fit_pipeline(&x, &y).unwrap();
        let result = fitted.transform_pipeline(&x).unwrap();
        assert_eq!(result.shape(), &[2, 6]);
    }

    #[test]
    fn test_degree3_single_feature() {
        // [a] with degree=3, no bias → [a, a², a³]
        let poly = PolynomialFeatures::<f64>::new(3, false, false).unwrap();
        let x = array![[2.0]];
        let out = poly.transform(&x).unwrap();
        assert_eq!(out.shape()[1], 3);
        assert_abs_diff_eq!(out[[0, 0]], 2.0, epsilon = 1e-10);
        assert_abs_diff_eq!(out[[0, 1]], 4.0, epsilon = 1e-10);
        assert_abs_diff_eq!(out[[0, 2]], 8.0, epsilon = 1e-10);
    }
}