ferrolearn-cluster 0.5.0

Clustering algorithms 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
//! Spectral Clustering via graph Laplacian eigenmaps.
//!
//! This module provides [`SpectralClustering`], a graph-based clustering
//! algorithm that embeds the data into a low-dimensional space defined by the
//! eigenvectors of the graph Laplacian and then clusters the embedded points
//! with K-Means.
//!
//! # Algorithm
//!
//! 1. **Affinity matrix**: compute the pairwise RBF (Gaussian) kernel
//!    `A[i,j] = exp(-gamma * ||x_i - x_j||^2)`.
//! 2. **Normalized graph Laplacian**: `L_sym = D^{-1/2} A D^{-1/2}`, where
//!    `D[i,i] = sum_j A[i,j]` is the degree matrix.
//! 3. **Eigendecomposition**: compute the `n_clusters` eigenvectors of
//!    `L_sym` that correspond to the **largest** eigenvalues (these are the
//!    smoothest graph signals).
//! 4. **Row-normalize** the embedding matrix so each row has unit L2 norm.
//! 5. **K-Means** clustering on the embedded points with `n_init` restarts.
//!
//! Spectral Clustering does **not** implement [`ferrolearn_core::Predict`]
//! because there is no simple way to embed new points into the learned eigenspace
//! without refitting.
//!
//! # Notes
//!
//! The eigendecomposition is performed in `f64` via `ferrolearn_core::Backend`
//! (`NdarrayFaerBackend`) regardless of the input float type `F`, because
//! `faer`'s solver only supports `f64`.  The results are then cast back to `F`.
//!
//! # Examples
//!
//! ```
//! use ferrolearn_cluster::SpectralClustering;
//! use ferrolearn_core::Fit;
//! use ndarray::Array2;
//!
//! let x = Array2::from_shape_vec((6, 2), vec![
//!     1.0_f64, 1.0,  1.1, 1.0,  1.0, 1.1,
//!     9.0, 9.0,  9.1, 9.0,  9.0, 9.1,
//! ]).unwrap();
//!
//! let model = SpectralClustering::<f64>::new(2).with_random_state(42);
//! let fitted = model.fit(&x, &()).unwrap();
//! assert_eq!(fitted.labels().len(), 6);
//! ```
//!
//! # `## REQ status`
//!
//! Binary classification (R-DEFER-2): two states only — SHIPPED needs impl + a
//! non-test production consumer + green verification; NOT-STARTED carries the
//! open prereq blocker. **`SpectralClustering` has NO PyO3 binding** — there is
//! no `_RsSpectralClustering` and no `ferrolearn.SpectralClustering` (confirmed
//! by grep: zero `SpectralClustering`/`RsSpectral` hits under
//! `ferrolearn-python/`). The non-test production consumer is therefore the
//! crate re-export at the crate root (`pub use
//! spectral::{FittedSpectralClustering, SpectralClustering}` in
//! `ferrolearn-cluster/src/lib.rs`), exposing `fit` / `fit_predict` /
//! `labels()`. **Honest underclaim (R-HONEST-3): this unit is a SIMPLIFIED
//! spectral-clustering VARIANT and does NOT achieve `SpectralClustering` label
//! parity** — the embedding diverges (ferrolearn row-L2-normalized top-k of
//! `D^{-1/2}AD^{-1/2}` vs sklearn `_spectral_embedding` bottom-k of
//! `I − D^{-1/2}AD^{-1/2}` scaled by `1/dd`), affinity/assign_labels modes and
//! most params are missing, and there is no binding. The ONLY contract that
//! VALUE-matches the live oracle with a real consumer is gamma/`n_clusters`
//! parameter validation (the gamma `[0, inf)` interval just landed via #930).
//! Green verification = the in-tree `spectral` lib tests plus the live-sklearn
//! pins / guards (`ferrolearn-cluster/tests/divergence_spectral.rs`): the RED
//! pin `divergence_spectral_gamma_zero_allowed` (#930), now PASSING after the
//! over-rejection fix; plus the green guards
//! `green_spectral_gamma_negative_rejected` (#930),
//! `green_spectral_n_clusters_zero_rejected`,
//! `green_spectral_insufficient_samples`. Cites use symbol anchors (ferrolearn)
//! / `file:line` (sklearn 1.5.2, commit 156ef14). Live oracle = installed
//! sklearn 1.5.2. (REQ numbering follows `.design/cluster/spectral.md`.)
//!
//! | REQ | Status | Evidence |
//! |---|---|---|
//! | REQ-7 (gamma `[0, inf)` validation: `gamma >= 0` accepted / `gamma < 0` rejected; `n_clusters >= 1`; `n_samples >= n_clusters`) | SHIPPED (validation contract only) | `fn fit` for `SpectralClustering` rejects ONLY `self.gamma < F::zero()` with `FerroError::InvalidParameter { name: "gamma", reason: "gamma must be >= 0 (sklearn Interval[0, inf))" }`, mirroring `SpectralClustering._parameter_constraints["gamma"] = Interval(Real, 0, None, closed="left")` (`_spectral.py:612`) — `gamma = 0.0` is INSIDE the closed-left interval so it is ACCEPTED (RBF collapses to all-ones affinity); `n_clusters == 0` rejected per `Interval(Integral, 1, None, closed="left")` (`_spectral.py:607`); `n_samples < n_clusters` rejected with `FerroError::InsufficientSamples`. Non-test consumer: crate re-export of `fit` / `fit_predict` / `labels()` (`lib.rs`). Verified: green pin `divergence_spectral_gamma_zero_allowed` (#930, now PASSING after the over-rejection fix): `with_gamma(0.0).fit(X, &())` returns `Ok` (sklearn `SpectralClustering(gamma=0.0).fit(X)` runs, NO error); green guards `green_spectral_gamma_negative_rejected` (`with_gamma(-1.0).fit` → `Err`), `green_spectral_n_clusters_zero_rejected` (`new(0).fit` → `Err`), `green_spectral_insufficient_samples` (`new(3).fit(X_1row)` → `Err`). GAP (NOT-STARTED, REQ-6): `n_clusters=8` default and the sklearn `InvalidParameterError`/`ValueError` error ABI are absent — ferrolearn requires `n_clusters` and uses `FerroError`. |
//! | REQ-1 (RBF affinity VALUE) | NOT-STARTED | open prereq blocker **#934**. `fn affinity_matrix` computes `exp(-gamma*||xi-xj||^2)` (diagonal forced `1.0`), VALUE-matching `pairwise_kernels(X, metric='rbf')` (`_spectral.py:730`) to full f64 precision (`rbf_kernel(blobs, gamma=0.1)[0,1] = 0.9950124791926823`). BUT `fn affinity_matrix` is a PRIVATE helper with no public accessor and no non-test consumer (no `affinity_matrix_`, REQ-8) — cannot be SHIPPED standalone (R-HONEST-2/R-DEFER-1). |
//! | REQ-2 (`labels_` VALUE parity) | NOT-STARTED | open prereq blocker **#929** (depends on #934 KMeans). `fit` / `fit_predict` produce `labels_` via a DIFFERENT pipeline (`fn normalized_laplacian` + `fn top_k_eigenvectors` + `fn row_normalize` + ferrolearn `KMeans`) than sklearn `_spectral_embedding` + `k_means` (`_spectral.py:756-766`). On `make_circles(…, random_state=0)` at `gamma=10`, sklearn vs ferrolearn `fit_predict` differ at indices 10,19 (2/30). Agreement on blobs / circles@0.1 is COINCIDENTAL (shared subspace) — NOT value parity. |
//! | REQ-3 (spectral embedding algorithm — `I − D^{-1/2}AD^{-1/2}` + `embedding/dd` + sign-flip) | NOT-STARTED | open prereq blocker **#929**. sklearn `_spectral_embedding` (`_spectral_embedding.py:300-469`): `L = csgraph_laplacian(A, normed=True) = I − D^{-1/2}AD^{-1/2}` (`:333`), SMALLEST-eigenvalue eigenvectors, `embedding = embedding / dd` (`:378`/`:443`), `_deterministic_vector_sign_flip` (`:465`). ferrolearn `fn normalized_laplacian` builds `D^{-1/2}AD^{-1/2}` (NO `I−`), `fn top_k_eigenvectors` takes the LARGEST, `fn row_normalize` does unit-L2 per row (NOT `/dd`), no sign flip — blobs `gamma=0.1` sklearn embedding row-magnitude ≈ `0.158` (`=1/dd`) vs ferrolearn `1.0`. The root-cause divergence. |
//! | REQ-4 (affinity modes nearest_neighbors/precomputed/poly/sigmoid) | NOT-STARTED | open prereq blocker **#931**. sklearn builds affinity via `kneighbors_graph` (`'nearest_neighbors'`, `:709-713`), passthrough (`'precomputed'`, `:720-721`), or any `pairwise_kernels` metric (`:730`). ferrolearn `fn affinity_matrix` is hard-coded RBF only; no `affinity` parameter. |
//! | REQ-5 (assign_labels discretize/cluster_qr) | NOT-STARTED | open prereq blocker **#932**. sklearn `assign_labels ∈ {'kmeans','discretize','cluster_qr'}` (`_spectral.py:625`, branch `:755-766`). ferrolearn `fn fit` hard-codes `KMeans`; no `assign_labels` parameter. |
//! | REQ-6 (missing params eigen_solver/n_components/eigen_tol/n_neighbors/degree/coef0/kernel_params/n_jobs/verbose/affinity + `n_clusters=8` default) | NOT-STARTED | open prereq blocker **#933**. sklearn `__init__` (`_spectral.py:633-666`) takes 16 params with `n_clusters=8` default (`:635`). `SpectralClustering<F>` (`fn new` + builders) has only `n_clusters` (REQUIRED, no default) / `gamma` / `n_init` / `random_state`. |
//! | REQ-8 (fitted attrs `affinity_matrix_`/`n_features_in_`) | NOT-STARTED | open prereq blocker **#934**. sklearn exposes `affinity_matrix_` + `n_features_in_` (`_spectral.py:524-538`). `FittedSpectralClustering<F>` stores only private `labels_` (+ `PhantomData`), exposes only `labels()`. No `affinity_matrix_` accessor — so the value-matching affinity (REQ-1) has no consumer. |
//! | REQ-9 (KMeans assign-labels parity) | NOT-STARTED | open prereq blocker **#934** (kmeans.rs unit). sklearn `k_means(maps, n_clusters, n_init, random_state)` uses k-means++ init with NumPy RNG (`_spectral.py:756`). ferrolearn `KMeans::new(k).with_n_init` is a separate unit with its own init/RNG — even a matched embedding would not guarantee identical labels. |
//! | REQ-10 (PyO3 binding) | NOT-STARTED | open prereq blocker **#935**. `grep -rln SpectralClustering ferrolearn-python/` is EMPTY — no `_RsSpectralClustering`, so `import ferrolearn` cannot reach `SpectralClustering`. The only non-test consumer of `fit` / `fit_predict` / `labels()` is the crate re-export (`lib.rs`). |
//! | REQ-11 (ferray substrate) | NOT-STARTED | open prereq blocker **#935**. `spectral.rs` imports `ndarray::{Array1, Array2}` + `num_traits::Float` + `ferrolearn_core::NdarrayFaerBackend` (`eigh`); not migrated to `ferray-core` / `ferray::linalg` (R-SUBSTRATE-1/2). |
//! | REQ-12 (reject non-finite input) | SHIPPED | `fn reject_non_finite` called at the top of `fn fit` (after the param/sample checks, before the RBF affinity / eigendecomposition) rejects NaN AND infinity with `FerroError::InvalidParameter{name:"X"}`, mirroring sklearn's `SpectralClustering.fit` → `_validate_data(force_all_finite=True)` default (`_spectral.py:691`), which raises `ValueError` (`validation.py:147-154`). ferrolearn previously rejected NaN (NaN affinities fail the eigendecomposition) but SILENTLY ACCEPTED +Inf; now both reject. Consumer: the existing `fit` entry — crate re-export `pub use spectral::{FittedSpectralClustering, SpectralClustering}` (`lib.rs`). Pinned by `divergence_nonfinite_reject_spillover.rs` (`divergence_spectral_fit_rejects_inf`) — live sklearn 1.5.2 raises, ferrolearn now `Err`. Finite input byte-identical (the module's oracle pins stay green). Closes #2286 for this estimator. |

use crate::kmeans::KMeans;
use ferrolearn_core::NdarrayFaerBackend;
use ferrolearn_core::backend::Backend;
use ferrolearn_core::error::FerroError;
use ferrolearn_core::traits::{Fit, Predict};
use ndarray::{Array1, Array2};
use num_traits::Float;

// ─────────────────────────────────────────────────────────────────────────────
// Configuration struct
// ─────────────────────────────────────────────────────────────────────────────

/// Spectral Clustering configuration (unfitted).
///
/// Holds hyperparameters.  Call [`Fit::fit`] to run the algorithm and produce a
/// [`FittedSpectralClustering`].
///
/// # Type Parameters
///
/// - `F`: The floating-point type (`f32` or `f64`).
#[derive(Debug, Clone)]
pub struct SpectralClustering<F> {
    /// Number of clusters.
    pub n_clusters: usize,
    /// RBF kernel parameter: `A[i,j] = exp(-gamma * ||x_i - x_j||^2)`.
    /// Defaults to `1.0`.
    pub gamma: F,
    /// Number of K-Means restarts in the final clustering step.
    pub n_init: usize,
    /// Optional random seed for the K-Means restarts.
    pub random_state: Option<u64>,
}

impl<F: Float> SpectralClustering<F> {
    /// Create a new `SpectralClustering` with the given number of clusters.
    ///
    /// Defaults: `gamma = 1.0`, `n_init = 10`, `random_state = None`.
    #[must_use]
    pub fn new(n_clusters: usize) -> Self {
        Self {
            n_clusters,
            gamma: F::one(),
            n_init: 10,
            random_state: None,
        }
    }

    /// Set the RBF kernel parameter `gamma`.
    ///
    /// Must be non-negative (`>= 0`).
    #[must_use]
    pub fn with_gamma(mut self, gamma: F) -> Self {
        self.gamma = gamma;
        self
    }

    /// Set the number of K-Means restarts.
    #[must_use]
    pub fn with_n_init(mut self, n_init: usize) -> Self {
        self.n_init = n_init;
        self
    }

    /// Set the random seed for reproducibility.
    #[must_use]
    pub fn with_random_state(mut self, seed: u64) -> Self {
        self.random_state = Some(seed);
        self
    }
}

// ─────────────────────────────────────────────────────────────────────────────
// Fitted struct
// ─────────────────────────────────────────────────────────────────────────────

/// Fitted Spectral Clustering model.
///
/// Stores the cluster labels for the training data.
///
/// Spectral Clustering does **not** implement [`ferrolearn_core::Predict`].
#[derive(Debug, Clone)]
pub struct FittedSpectralClustering<F> {
    /// Cluster label for each training sample (0-indexed).
    labels_: Array1<usize>,
    /// Phantom for the float type.
    _marker: std::marker::PhantomData<F>,
}

impl<F: Float> FittedSpectralClustering<F> {
    /// Return the cluster labels for the training data.
    #[must_use]
    pub fn labels(&self) -> &Array1<usize> {
        &self.labels_
    }
}

// ─────────────────────────────────────────────────────────────────────────────
// Internal helpers
// ─────────────────────────────────────────────────────────────────────────────

/// Reject `X` containing any non-finite value (NaN or infinity).
///
/// Mirrors sklearn's `SpectralClustering.fit` →
/// `self._validate_data(X, accept_sparse=["csr","csc","coo"], dtype=np.float64, ensure_min_samples=2)`
/// (`sklearn/cluster/_spectral.py:691`), which keeps the `force_all_finite=True`
/// default and raises `ValueError("Input X contains NaN.")` /
/// `"... contains infinity ..."` (`sklearn/utils/validation.py:147-154`). NaN AND
/// infinity are both rejected up front so neither reaches the RBF affinity /
/// eigendecomposition. Never panics (R-CODE-2).
fn reject_non_finite<F: Float>(x: &Array2<F>) -> Result<(), FerroError> {
    if x.iter().any(|v| !v.is_finite()) {
        return Err(FerroError::InvalidParameter {
            name: "X".into(),
            reason: "Input X contains NaN or infinity.".into(),
        });
    }
    Ok(())
}

/// Build the RBF affinity matrix in `f64`.
fn affinity_matrix<F: Float>(x: &Array2<F>, gamma: f64) -> Array2<f64> {
    let n = x.nrows();
    Array2::from_shape_fn((n, n), |(i, j)| {
        if i == j {
            1.0_f64
        } else {
            let sq: F = x
                .row(i)
                .iter()
                .zip(x.row(j).iter())
                .fold(F::zero(), |acc, (&a, &b)| acc + (a - b) * (a - b));
            let sq64 = sq.to_f64().unwrap_or(0.0);
            (-gamma * sq64).exp()
        }
    })
}

/// Compute the normalized symmetric graph Laplacian `D^{-1/2} A D^{-1/2}`.
fn normalized_laplacian(a: &Array2<f64>) -> Array2<f64> {
    let n = a.nrows();
    // Degree vector: d[i] = sum_j A[i,j]
    let d: Vec<f64> = (0..n).map(|i| a.row(i).iter().sum()).collect();

    // D^{-1/2}: avoid division by zero.
    let d_inv_sqrt: Vec<f64> = d
        .iter()
        .map(|&di| if di > 0.0 { 1.0 / di.sqrt() } else { 0.0 })
        .collect();

    // L_sym[i,j] = d_inv_sqrt[i] * A[i,j] * d_inv_sqrt[j]
    Array2::from_shape_fn((n, n), |(i, j)| d_inv_sqrt[i] * a[[i, j]] * d_inv_sqrt[j])
}

/// Extract the top-`k` eigenvectors (by eigenvalue magnitude) of a symmetric
/// matrix.  Returns an `(n, k)` matrix.
fn top_k_eigenvectors(sym: &Array2<f64>, k: usize) -> Result<Array2<f64>, FerroError> {
    let (eigenvalues, eigenvectors) = NdarrayFaerBackend::eigh(sym)?;

    // `eigh` returns eigenvalues in non-decreasing order; the top-k are at the end.
    let n = eigenvalues.len();
    let start = n.saturating_sub(k);

    let n_rows = eigenvectors.nrows();
    let mut result = Array2::<f64>::zeros((n_rows, k));
    for (new_col, old_col) in (start..n).enumerate() {
        for row in 0..n_rows {
            result[[row, new_col]] = eigenvectors[[row, old_col]];
        }
    }

    Ok(result)
}

/// Row-normalize a matrix so each row has unit L2 norm.
/// Rows with zero norm are left as-is.
fn row_normalize(m: &Array2<f64>) -> Array2<f64> {
    let (n, d) = m.dim();
    Array2::from_shape_fn((n, d), |(i, j)| {
        let norm: f64 = m.row(i).iter().map(|&v| v * v).sum::<f64>().sqrt();
        if norm > 0.0 {
            m[[i, j]] / norm
        } else {
            m[[i, j]]
        }
    })
}

// ─────────────────────────────────────────────────────────────────────────────
// Fit implementation
// ─────────────────────────────────────────────────────────────────────────────

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

    /// Fit the Spectral Clustering model to the data.
    ///
    /// # Errors
    ///
    /// - [`FerroError::InvalidParameter`] if `n_clusters == 0`, `gamma < 0`,
    ///   or `n_init == 0`.
    /// - [`FerroError::InsufficientSamples`] if `n_samples < n_clusters`.
    /// - [`FerroError::NumericalInstability`] if the eigendecomposition fails.
    fn fit(&self, x: &Array2<F>, _y: &()) -> Result<FittedSpectralClustering<F>, FerroError> {
        let n_samples = x.nrows();

        // Validate parameters.
        if self.n_clusters == 0 {
            return Err(FerroError::InvalidParameter {
                name: "n_clusters".into(),
                reason: "must be at least 1".into(),
            });
        }
        if self.gamma < F::zero() {
            return Err(FerroError::InvalidParameter {
                name: "gamma".into(),
                reason: "gamma must be >= 0 (sklearn Interval[0, inf))".into(),
            });
        }
        if n_samples == 0 {
            return Err(FerroError::InsufficientSamples {
                required: self.n_clusters,
                actual: 0,
                context: "SpectralClustering requires at least n_clusters samples".into(),
            });
        }
        if n_samples < self.n_clusters {
            return Err(FerroError::InsufficientSamples {
                required: self.n_clusters,
                actual: n_samples,
                context: "SpectralClustering requires at least n_clusters samples".into(),
            });
        }

        // Reject non-finite X up front (NaN AND Inf), mirroring sklearn's
        // `_validate_data(force_all_finite=True)` reached from
        // `SpectralClustering.fit` (`_spectral.py:691`), which raises
        // `ValueError` (R-DEV-1, R-CODE-2). ferrolearn previously rejected NaN
        // (NaN affinities propagate to a failed eigendecomposition) but
        // SILENTLY ACCEPTED +Inf — this rejects both.
        reject_non_finite(x)?;

        let gamma64 = self.gamma.to_f64().unwrap_or(1.0);

        // Step 1: affinity matrix.
        let aff = affinity_matrix(x, gamma64);

        // Step 2: normalized Laplacian.
        let lap = normalized_laplacian(&aff);

        // Step 3: top-k eigenvectors.
        let k = self.n_clusters;
        let embed = top_k_eigenvectors(&lap, k)?;

        // Step 4: row-normalize.
        let embed_norm = row_normalize(&embed);

        // Step 5: K-Means on the embedded points.
        // Convert embedding to F.
        let embed_f: Array2<F> = Array2::from_shape_fn(embed_norm.dim(), |(i, j)| {
            F::from(embed_norm[[i, j]]).unwrap_or_else(F::zero)
        });

        let mut km = KMeans::<F>::new(k).with_n_init(self.n_init);
        if let Some(seed) = self.random_state {
            km = km.with_random_state(seed);
        }

        let fitted_km = km.fit(&embed_f, &())?;
        let labels = fitted_km.predict(&embed_f)?;

        Ok(FittedSpectralClustering {
            labels_: labels,
            _marker: std::marker::PhantomData,
        })
    }
}

impl<F: Float + Send + Sync + 'static> SpectralClustering<F> {
    /// Fit on `x` and return the cluster labels for those samples in one
    /// call. Equivalent to sklearn `ClusterMixin.fit_predict`.
    ///
    /// # Errors
    ///
    /// Forwards any error from [`Fit::fit`].
    pub fn fit_predict(&self, x: &Array2<F>) -> Result<Array1<usize>, FerroError> {
        let fitted = self.fit(x, &())?;
        Ok(fitted.labels().clone())
    }
}

// ─────────────────────────────────────────────────────────────────────────────
// Tests
// ─────────────────────────────────────────────────────────────────────────────

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

    /// Two well-separated 2-D blobs.
    fn two_blobs() -> Array2<f64> {
        Array2::from_shape_vec(
            (10, 2),
            vec![
                0.0, 0.0, 0.2, 0.1, -0.1, 0.2, 0.1, -0.1, 0.0, 0.1, 10.0, 10.0, 10.2, 10.1, 9.9,
                10.2, 10.1, 9.9, 10.0, 10.1,
            ],
        )
        .unwrap()
    }

    #[test]
    fn test_two_blobs_two_clusters() {
        let x = two_blobs();
        let model = SpectralClustering::<f64>::new(2)
            .with_gamma(0.1)
            .with_random_state(42);
        let fitted = model.fit(&x, &()).unwrap();

        let labels = fitted.labels();
        assert_eq!(labels.len(), 10);

        // Points 0-4 should share a label.
        assert_eq!(labels[0], labels[1]);
        assert_eq!(labels[0], labels[2]);
        assert_eq!(labels[0], labels[3]);
        assert_eq!(labels[0], labels[4]);

        // Points 5-9 should share a different label.
        assert_eq!(labels[5], labels[6]);
        assert_eq!(labels[5], labels[7]);
        assert_eq!(labels[5], labels[8]);
        assert_eq!(labels[5], labels[9]);

        assert_ne!(labels[0], labels[5]);
    }

    #[test]
    fn test_labels_length_matches_n_samples() {
        let x = two_blobs();
        let fitted = SpectralClustering::<f64>::new(2)
            .with_random_state(0)
            .fit(&x, &())
            .unwrap();
        assert_eq!(fitted.labels().len(), x.nrows());
    }

    #[test]
    fn test_labels_in_valid_range() {
        let x = two_blobs();
        let k = 2usize;
        let fitted = SpectralClustering::<f64>::new(k)
            .with_random_state(1)
            .fit(&x, &())
            .unwrap();
        for &l in fitted.labels() {
            assert!(l < k, "label {l} >= n_clusters {k}");
        }
    }

    #[test]
    fn test_single_cluster() {
        let x = two_blobs();
        let fitted = SpectralClustering::<f64>::new(1)
            .with_random_state(0)
            .fit(&x, &())
            .unwrap();
        for &l in fitted.labels() {
            assert_eq!(l, 0);
        }
    }

    #[test]
    fn test_invalid_n_clusters_zero() {
        let x = two_blobs();
        let result = SpectralClustering::<f64>::new(0).fit(&x, &());
        assert!(result.is_err());
    }

    #[test]
    fn test_gamma_zero_allowed() {
        // sklearn `gamma: Interval(Real, 0, None, closed="left")`
        // (_spectral.py:612) → [0.0, inf), so gamma=0.0 is ACCEPTED (RBF
        // collapses to an all-ones affinity). Mirror that contract (R-HONEST-4):
        // the prior assertion `is_err()` pinned a non-sklearn over-rejection.
        let x = two_blobs();
        let result = SpectralClustering::<f64>::new(2)
            .with_gamma(0.0)
            .fit(&x, &());
        assert!(result.is_ok());
    }

    #[test]
    fn test_invalid_gamma_negative() {
        let x = two_blobs();
        let result = SpectralClustering::<f64>::new(2)
            .with_gamma(-1.0)
            .fit(&x, &());
        assert!(result.is_err());
    }

    #[test]
    fn test_empty_data_error() {
        let x = Array2::<f64>::zeros((0, 2));
        let result = SpectralClustering::<f64>::new(2).fit(&x, &());
        assert!(result.is_err());
    }

    #[test]
    fn test_insufficient_samples_error() {
        let x = Array2::from_shape_vec((1, 2), vec![0.0, 0.0]).unwrap();
        let result = SpectralClustering::<f64>::new(3).fit(&x, &());
        assert!(result.is_err());
    }

    #[test]
    fn test_n_clusters_equals_n_samples() {
        // k == n should be valid (each point its own cluster).
        let x = Array2::from_shape_vec((3, 2), vec![0.0, 0.0, 5.0, 5.0, 10.0, 10.0]).unwrap();
        let fitted = SpectralClustering::<f64>::new(3)
            .with_random_state(0)
            .fit(&x, &())
            .unwrap();
        assert_eq!(fitted.labels().len(), 3);
    }

    #[test]
    fn test_f32_support() {
        let x = Array2::from_shape_vec(
            (6, 2),
            vec![
                0.0f32, 0.0, 0.1, 0.1, -0.1, 0.1, 10.0, 10.0, 10.1, 10.1, 9.9, 10.1,
            ],
        )
        .unwrap();

        let fitted = SpectralClustering::<f32>::new(2)
            .with_gamma(0.1)
            .with_random_state(42)
            .fit(&x, &())
            .unwrap();
        assert_eq!(fitted.labels().len(), 6);
    }

    #[test]
    fn test_reproducibility_with_seed() {
        let x = two_blobs();
        let model = SpectralClustering::<f64>::new(2)
            .with_gamma(0.1)
            .with_random_state(7);

        let fitted1 = model.fit(&x, &()).unwrap();
        let fitted2 = model.fit(&x, &()).unwrap();
        assert_eq!(fitted1.labels(), fitted2.labels());
    }
}