section-properties 0.3.0

Structural cross-section properties: geometric/principal moments, section moduli, warping/torsion FEM, plastic, cold-formed, and fire analysis.
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
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
//! Solver backends. See module docs in [`crate::fea`].

use super::{CgResult, CgStatus, SkylineLdlt, SparseMatrix, cg_solve};
use crate::fea::solver::SolverError;

/// Selectable solver backend.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub enum SolverKind {
    /// General sparse LU with partial pivoting.
    SparseLu,
    /// Profile LDL^T (symmetric systems).
    SkylineLdlt,
    /// Jacobi-preconditioned conjugate gradient.
    PcG,
    /// Incomplete-Cholesky (IC(0)) preconditioned conjugate gradient.
    Iccg,
    /// Intel MKL PARDISO via FFI (requires the `pardiso` feature and MKL).
    Pardiso,
}

/// Direct solver types (factor once, solve many).
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub enum DirectSolver {
    SparseLu,
    SkylineLdlt,
    #[cfg(feature = "pardiso")]
    Pardiso,
}

/// Iterative solver types.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub enum IterativeSolver {
    PcG,
    Iccg,
}

/// A factored direct solver able to solve A x = b for multiple right-hand sides.
pub enum FactoredDirectSolver {
    Lu(SparseLu),
    Ldlt(SkylineLdlt),
    #[cfg(feature = "pardiso")]
    Pardiso(crate::fea::solvers::pardiso::PardisoSolver),
}

impl FactoredDirectSolver {
    /// Factor `matrix` with the selected direct backend.
    pub fn factor(kind: DirectSolver, matrix: &SparseMatrix) -> Result<Self, SolverError> {
        match kind {
            DirectSolver::SparseLu => Ok(FactoredDirectSolver::Lu(
                SparseLu::factor(matrix).map_err(|e| SolverError::FactorizationFailed(e))?,
            )),
            DirectSolver::SkylineLdlt => Ok(FactoredDirectSolver::Ldlt(
                SkylineLdlt::factor(matrix)
                    .map_err(|e| SolverError::FactorizationFailed(e.to_string()))?,
            )),
            #[cfg(feature = "pardiso")]
            DirectSolver::Pardiso => {
                // PARDISO requires the constraint vector for the augmented system
                // This is a limitation - we'll need to handle this separately
                Err(SolverError::NotImplemented(
                    "PARDISO factor requires constraint vector. Use DirectLagrangeSolver instead."
                        .to_string(),
                ))
            }
        }
    }

    /// Direct solve (LU / LDLT). Returns error if solver fails.
    pub fn solve(&self, b: &[f64]) -> Result<Vec<f64>, SolverError> {
        match self {
            FactoredDirectSolver::Lu(lu) => lu.solve(b).map_err(SolverError::SolveFailed),
            FactoredDirectSolver::Ldlt(l) => l
                .solve(b)
                .map_err(|e| SolverError::SolveFailed(e.to_string())),
            #[cfg(feature = "pardiso")]
            FactoredDirectSolver::Pardiso(_) => Err(SolverError::NotImplemented(
                "PARDISO solve not implemented in this context".to_string(),
            )),
        }
    }
}

/// Iterative solver for multiple right-hand sides.
pub struct IterativeSolverInstance {
    kind: IterativeSolver,
    matrix: SparseMatrix,
    max_iter: usize,
    tol: f64,
}

impl IterativeSolverInstance {
    /// Create a new iterative solver instance.
    pub fn new(kind: IterativeSolver, matrix: SparseMatrix, max_iter: usize, tol: f64) -> Self {
        Self {
            kind,
            matrix,
            max_iter,
            tol,
        }
    }

    /// Solve A x = b. Returns error if solver fails or doesn't converge.
    pub fn solve(&self, b: &[f64]) -> Result<Vec<f64>, SolverError> {
        match self.kind {
            IterativeSolver::PcG => {
                let result = cg_solve(&self.matrix, b, self.max_iter, self.tol);
                match result.status {
                    CgStatus::Converged => Ok(result.x),
                    CgStatus::NotPositiveDefinite => Err(SolverError::SolveFailed(
                        "matrix not positive definite".to_string(),
                    )),
                    CgStatus::Breakdown => {
                        Err(SolverError::SolveFailed("solver breakdown".to_string()))
                    }
                    CgStatus::InvalidInput => {
                        Err(SolverError::InvalidInput("invalid input".to_string()))
                    }
                    CgStatus::MaxIterations => Err(SolverError::SolveFailed(
                        "max iterations reached".to_string(),
                    )),
                }
            }
            IterativeSolver::Iccg => {
                let result = iccg_solve(&self.matrix, b, self.max_iter, self.tol);
                match result.status {
                    CgStatus::Converged => Ok(result.x),
                    CgStatus::NotPositiveDefinite => Err(SolverError::SolveFailed(
                        "matrix not positive definite".to_string(),
                    )),
                    CgStatus::Breakdown => {
                        Err(SolverError::SolveFailed("solver breakdown".to_string()))
                    }
                    CgStatus::InvalidInput => {
                        Err(SolverError::InvalidInput("invalid input".to_string()))
                    }
                    CgStatus::MaxIterations => Err(SolverError::SolveFailed(
                        "max iterations reached".to_string(),
                    )),
                }
            }
        }
    }
}

/// Legacy SparseSolver for backward compatibility.
pub struct SparseSolver {
    direct: FactoredDirectSolver,
}

impl SparseSolver {
    /// Factor `matrix` with the selected backend.
    pub fn factor(kind: SolverKind, matrix: &SparseMatrix) -> Result<Self, String> {
        match kind {
            SolverKind::SparseLu => Ok(SparseSolver { direct: FactoredDirectSolver::factor(DirectSolver::SparseLu, matrix).map_err(|e| e.to_string())? }),
            SolverKind::SkylineLdlt => Ok(SparseSolver { direct: FactoredDirectSolver::factor(DirectSolver::SkylineLdlt, matrix).map_err(|e| e.to_string())? }),
            SolverKind::Pardiso => Err("PARDISO factor requires constraint vector. Use DirectLagrangeSolver instead.".to_string()),
            SolverKind::PcG | SolverKind::Iccg => Err("iterative solvers (PCG, ICCG) do not support factor(). Use solve_iterative() or IterativeSolverInstance.".to_string()),
        }
    }

    /// Direct solve (LU / LDLT). Returns error if solver fails.
    pub fn solve(&self, b: &[f64]) -> Result<Vec<f64>, SolverError> {
        self.direct.solve(b)
    }

    /// Iterative solve used for PCG/ICCG kinds.
    pub fn solve_iterative(
        kind: SolverKind,
        matrix: &SparseMatrix,
        b: &[f64],
        max_iter: usize,
        tol: f64,
    ) -> Result<Vec<f64>, SolverError> {
        match kind {
            SolverKind::PcG => {
                let result = cg_solve(matrix, b, max_iter, tol);
                match result.status {
                    CgStatus::Converged => Ok(result.x),
                    CgStatus::NotPositiveDefinite => Err(SolverError::SolveFailed(
                        "matrix not positive definite".to_string(),
                    )),
                    CgStatus::Breakdown => {
                        Err(SolverError::SolveFailed("solver breakdown".to_string()))
                    }
                    CgStatus::InvalidInput => {
                        Err(SolverError::InvalidInput("invalid input".to_string()))
                    }
                    CgStatus::MaxIterations => Err(SolverError::SolveFailed(
                        "max iterations reached".to_string(),
                    )),
                }
            }
            SolverKind::Iccg => {
                let result = iccg_solve(matrix, b, max_iter, tol);
                match result.status {
                    CgStatus::Converged => Ok(result.x),
                    CgStatus::NotPositiveDefinite => Err(SolverError::SolveFailed(
                        "matrix not positive definite".to_string(),
                    )),
                    CgStatus::Breakdown => {
                        Err(SolverError::SolveFailed("solver breakdown".to_string()))
                    }
                    CgStatus::InvalidInput => {
                        Err(SolverError::InvalidInput("invalid input".to_string()))
                    }
                    CgStatus::MaxIterations => Err(SolverError::SolveFailed(
                        "max iterations reached".to_string(),
                    )),
                }
            }
            _ => Err(SolverError::NotImplemented(
                "solver kind not supported for iterative solve".to_string(),
            )),
        }
    }
}

// ---------------------------------------------------------------------------
// Sparse LU: left-looking, column-oriented with true row partial pivoting
// (PA = L U). No static pivoting / diagonal perturbation - returns error
// on near-singular matrices.
// ---------------------------------------------------------------------------
//
// IMPORTANT: This implementation is DENSE-BACKED.
// The input sparse matrix is converted to a dense n×n matrix for Gaussian
// elimination with partial pivoting. Time complexity: O(n³), Space: O(n²).
// Suitable for n up to ~2000-3000 on typical hardware.
// For larger problems, a true sparse LU (e.g. SuiteSparse, PARDISO) should be used.
//
/// Sparse LU factorisation P A = L U with true row partial pivoting.
pub struct SparseLu {
    n: usize,
    /// Row-wise unit-lower triangle (strict part), sorted columns per row.
    l_rows: Vec<Vec<(usize, f64)>>,
    /// Row-wise upper triangle including diagonal, sorted columns per row.
    u_rows: Vec<Vec<(usize, f64)>>,
    /// Row permutation: perm[final_row] = orig_row. P * A = L * U.
    perm: Vec<usize>,
}

impl SparseLu {
    /// Get the permutation vector (perm\[i\] = original row index at position i).
    pub fn perm(&self) -> &[usize] {
        &self.perm
    }

    /// Get L factor rows (strict lower triangle, unit diagonal implicit).
    pub fn l_rows(&self) -> &[Vec<(usize, f64)>] {
        &self.l_rows
    }

    /// Get U factor rows (upper triangle including diagonal).
    pub fn u_rows(&self) -> &[Vec<(usize, f64)>] {
        &self.u_rows
    }

    /// Get matrix size.
    pub fn n(&self) -> usize {
        self.n
    }

    /// Verify PA = LU for debugging purposes.
    /// Returns max absolute difference between PA and LU.
    pub fn verify_pa_eq_lu(&self, a: &SparseMatrix) -> f64 {
        let n = self.n;
        // Build A dense
        let mut a_dense = vec![vec![0.0f64; n]; n];
        for i in 0..n {
            for k in a.row_ptr()[i]..a.row_ptr()[i + 1] {
                let j = a.csr_cols()[k];
                a_dense[i][j] = a.csr_vals()[k];
            }
        }

        // Build PA
        let mut pa = vec![vec![0.0f64; n]; n];
        for i in 0..n {
            let orig_i = self.perm[i];
            for j in 0..n {
                pa[i][j] = a_dense[orig_i][j];
            }
        }

        // Build LU
        let mut lu = vec![vec![0.0f64; n]; n];
        for i in 0..n {
            for j in 0..n {
                let mut sum = 0.0;
                for k in 0..n {
                    let l_ik = if i == k {
                        1.0
                    } else if k < i {
                        self.l_rows[i]
                            .iter()
                            .find(|(c, _)| *c == k)
                            .map(|(_, v)| *v)
                            .unwrap_or(0.0)
                    } else {
                        0.0
                    };
                    let u_kj = if k <= j {
                        self.u_rows[k]
                            .iter()
                            .find(|(c, _)| *c == j)
                            .map(|(_, v)| *v)
                            .unwrap_or(0.0)
                    } else {
                        0.0
                    };
                    sum += l_ik * u_kj;
                }
                lu[i][j] = sum;
            }
        }

        // Max absolute difference
        let mut max_diff = 0.0;
        for i in 0..n {
            for j in 0..n {
                let diff = (pa[i][j] - lu[i][j]).abs();
                if diff > max_diff {
                    max_diff = diff;
                }
            }
        }
        max_diff
    }

    /// Factor A with true row partial pivoting (PA = LU).
    /// Returns error if matrix is singular or nearly singular.
    pub fn factor(a: &SparseMatrix) -> Result<SparseLu, String> {
        let n = a.n;
        let ac = if a.compressed {
            None
        } else {
            let mut m = a.clone();
            m.compress();
            Some(m)
        };
        let a_ref = ac.as_ref().unwrap_or(a);
        let (row_ptr, cols, vals) = a_ref.csr_data();

        // Convert to dense row-major working matrix for pivoting operations
        // This is memory-intensive but necessary for true partial pivoting
        // on a sparse matrix. We use a dense column-major working array.
        let mut a_dense = vec![vec![0.0f64; n]; n];
        for i in 0..n {
            for k in row_ptr[i]..row_ptr[i + 1] {
                let j = cols[k];
                a_dense[i][j] = vals[k];
            }
        }

        // Permutation vector: perm[i] = original row index that ended up at position i
        let mut perm = (0..n).collect::<Vec<usize>>();

        // LU factors stored in dense format during factorization, then compressed
        let mut l = vec![vec![0.0f64; n]; n]; // unit diagonal implicit
        let mut u = vec![vec![0.0f64; n]; n];

        // Compute per-column scales for scale-invariant pivot tolerance
        // Using column-wise infinity norm scaled by machine epsilon
        let mut col_scales = vec![0.0f64; n];
        for j in 0..n {
            let mut max_abs = 0.0f64;
            for i in 0..n {
                max_abs = max_abs.max(a_dense[i][j].abs());
            }
            col_scales[j] = max_abs;
        }
        // Global matrix scale (for reference)
        let _matrix_scale = col_scales.iter().fold(0.0f64, |a, &v| a.max(v));

        // Scale-invariant pivot tolerance:
        // pivot_tol = EPS * n * max(|A[:,k]|_inf) * safety_factor
        // This ensures that small-scale matrices aren't incorrectly flagged as singular
        // and large-scale matrices don't reject legitimate small pivots
        const PIVOT_SAFETY_FACTOR: f64 = 100.0;

        // LU factorization with partial pivoting (Gaussian elimination with row swaps)
        for k in 0..n {
            // Find pivot row: max |A[i][k]| for i >= k
            let mut pivot_row = k;
            let mut max_val = a_dense[k][k].abs();
            for i in (k + 1)..n {
                let v = a_dense[i][k].abs();
                if v > max_val {
                    max_val = v;
                    pivot_row = i;
                }
            }

            // Check for singularity using scale-aware tolerance (per-column)
            // True scale invariance: use actual column scale without artificial floor.
            // If col_scale == 0 or non-finite, the column is all zeros or invalid -> singular.
            let col_scale = col_scales[k];
            if !col_scale.is_finite() || col_scale == 0.0 {
                return Err(format!(
                    "Singular matrix at column {}: column is all zeros or contains NaN/Inf (col_scale = {:.2e})",
                    k, col_scale
                ));
            }
            let pivot_tol = f64::EPSILON * n as f64 * col_scale * PIVOT_SAFETY_FACTOR;

            if max_val <= pivot_tol {
                return Err(format!(
                    "Singular or near-singular matrix at column {}: max pivot = {:.2e}, tolerance = {:.2e} (col_scale = {:.2e})",
                    k, max_val, pivot_tol, col_scale
                ));
            }

            // Swap rows k and pivot_row in A and L, and permutation
            if pivot_row != k {
                a_dense.swap(k, pivot_row);
                // Swap corresponding rows in L (only columns < k are filled)
                // Use split_at_mut to get two mutable references
                let (l_first, l_second) = if k < pivot_row {
                    l.split_at_mut(pivot_row)
                } else {
                    l.split_at_mut(k)
                };
                let (row_k, row_pivot) = if k < pivot_row {
                    (&mut l_first[k], &mut l_second[0])
                } else {
                    (&mut l_second[0], &mut l_first[k])
                };
                for j in 0..k {
                    std::mem::swap(&mut row_k[j], &mut row_pivot[j]);
                }
                perm.swap(k, pivot_row);
            }

            // U[k][k] = A[k][k] (after potential swap)
            u[k][k] = a_dense[k][k];

            // Compute U[k][j] for j > k
            for j in (k + 1)..n {
                u[k][j] = a_dense[k][j];
            }

            // Compute L[i][k] for i > k
            let pivot = u[k][k];
            for i in (k + 1)..n {
                l[i][k] = a_dense[i][k] / pivot;
            }

            // Update trailing submatrix: A[i][j] -= L[i][k] * U[k][j] for i,j > k
            for i in (k + 1)..n {
                let lik = l[i][k];
                if lik != 0.0 {
                    for j in (k + 1)..n {
                        a_dense[i][j] -= lik * u[k][j];
                    }
                    a_dense[i][k] = 0.0; // Below diagonal becomes zero
                }
            }
        }

        // Build sparse L and U factors (only non-zero entries)
        // Use relative drop tolerance: keep entries where |v| > max_row_abs * EPS * n
        // This avoids deleting legitimate small entries in scaled matrices
        let mut l_rows: Vec<Vec<(usize, f64)>> = vec![Vec::new(); n];
        let mut u_rows: Vec<Vec<(usize, f64)>> = vec![Vec::new(); n];

        // Compute per-row scale for relative drop tolerance
        let mut l_row_scale = vec![0.0f64; n];
        let mut u_row_scale = vec![0.0f64; n];
        for i in 0..n {
            for j in 0..i {
                l_row_scale[i] = l_row_scale[i].max(l[i][j].abs());
            }
            for j in i..n {
                u_row_scale[i] = u_row_scale[i].max(u[i][j].abs());
            }
        }

        for i in 0..n {
            // L has unit diagonal (not stored), strictly lower part
            let l_tol = l_row_scale[i] * f64::EPSILON * n as f64;
            for j in 0..i {
                if l[i][j].abs() > l_tol {
                    l_rows[i].push((j, l[i][j]));
                }
            }
            // U has diagonal and upper part
            let u_tol = u_row_scale[i] * f64::EPSILON * n as f64;
            for j in i..n {
                if u[i][j].abs() > u_tol {
                    u_rows[i].push((j, u[i][j]));
                }
            }
            l_rows[i].sort_unstable_by_key(|e| e.0);
            u_rows[i].sort_unstable_by_key(|e| e.0);
        }

        Ok(SparseLu {
            n,
            l_rows,
            u_rows,
            perm,
        })
    }

    /// Solve A x = b via forward/back substitution using P A = L U.
    /// Returns x = A^{-1} b or error if diagonal is near-zero.
    pub fn solve(&self, b: &[f64]) -> Result<Vec<f64>, String> {
        let n = self.n;
        if b.len() != n {
            return Err(format!(
                "Solve failed: RHS length {} does not match matrix size {}",
                b.len(),
                n
            ));
        }
        // Check for NaN/Inf in RHS
        for (i, &val) in b.iter().enumerate() {
            if !val.is_finite() {
                return Err(format!(
                    "Solve failed: RHS contains non-finite value at index {}: {}",
                    i, val
                ));
            }
        }
        // Apply permutation: Pb
        let mut y = vec![0.0f64; n];
        for i in 0..n {
            y[i] = b[self.perm[i]];
        }

        // Forward substitution: L y = Pb (L has unit diagonal)
        for i in 0..n {
            let mut sum = 0.0;
            for &(c, v) in &self.l_rows[i] {
                sum += v * y[c];
            }
            y[i] -= sum;
        }

        // Backward substitution: U x = y
        let mut x = vec![0.0f64; n];
        for i in (0..n).rev() {
            let mut sum = 0.0;
            // Find diagonal and sum U[i][j] * x[j] for j > i
            let mut diag = 1.0;
            for &(c, v) in &self.u_rows[i] {
                if c == i {
                    diag = v;
                } else if c > i {
                    sum += v * x[c];
                }
            }
            if !diag.is_finite() || diag == 0.0 {
                return Err(format!(
                    "Solve failed: near-zero or invalid diagonal at row {}: diag = {:.2e}",
                    i, diag
                ));
            }
            x[i] = (y[i] - sum) / diag;
        }
        Ok(x)
    }
}

// ---------------------------------------------------------------------------
// IC(0)-preconditioned conjugate gradient.
// ---------------------------------------------------------------------------

/// IC(0) incomplete Cholesky factorisation (same sparsity pattern as the
/// lower triangle of A), stored row-wise: A ~ L L^T.
pub struct Ic0Factor {
    n: usize,
    /// Row offsets into flattened storage for fast application.
    ptr: Vec<usize>,
    flat_cols: Vec<usize>,
    flat_vals: Vec<f64>,
}

impl Ic0Factor {
    pub fn factor(a: &SparseMatrix) -> Result<Ic0Factor, String> {
        let n = a.n;
        let ac = if a.compressed {
            None
        } else {
            let mut m = a.clone();
            m.compress();
            Some(m)
        };
        let a_ref = ac.as_ref().unwrap_or(a);
        let (row_ptr, cols, vals) = a_ref.csr_data();

        // Lower triangle row-wise with diagonal.
        let mut l_rows: Vec<Vec<(usize, f64)>> = vec![Vec::new(); n];
        for i in 0..n {
            for k in row_ptr[i]..row_ptr[i + 1] {
                if cols[k] <= i {
                    l_rows[i].push((cols[k], vals[k]));
                }
            }
            l_rows[i].sort_unstable_by_key(|e| e.0);
        }

        // IC(0): row-wise algorithm. For each row i, walk the strictly-lower
        // entries left-to-right: divide by the pivot of their column, then
        // apply the rank-1 correction to remaining entries (pattern
        // intersection only). Finally scale the diagonal.
        for i in 0..n {
            let len_i = l_rows[i].len();
            for m in 0..len_i {
                let (k, _) = l_rows[i][m];
                if k >= i {
                    break;
                }
                let dk = l_rows[k].last().map(|&(_, v)| v).unwrap_or(1.0);
                if dk.abs() < 1e-300 {
                    return Err(format!("IC(0): zero pivot at row {k}"));
                }
                l_rows[i][m].1 /= dk;
                let l_ik = l_rows[i][m].1;
                if l_ik == 0.0 {
                    continue;
                }
                let krow: Vec<(usize, f64)> = l_rows[k].clone();
                for &(cj, lj) in krow.iter() {
                    if cj <= k || cj > i {
                        continue;
                    }
                    // Includes the diagonal (cj == i): a_ii -= L_ik * L_ki.
                    if let Some(pos) = l_rows[i].iter().position(|&(cc, _)| cc == cj) {
                        l_rows[i][pos].1 -= l_ik * lj;
                    }
                }
            }

            // Diagonal: take corrected a_ii and scale to sqrt.
            let diag_val = l_rows[i].last().map(|&(_, v)| v).unwrap_or(0.0);
            let d = diag_val;
            if !(d > 0.0) || !d.is_finite() {
                return Err(format!("IC(0) breakdown at row {i}: non-positive diagonal"));
            }
            let last = l_rows[i].len() - 1;
            debug_assert_eq!(l_rows[i][last].0, i);
            l_rows[i][last].1 = d.sqrt();
        }

        // Flatten for fast forward/back substitution.
        let mut ptr = vec![0usize; n + 1];
        let mut flat_cols = Vec::new();
        let mut flat_vals = Vec::new();
        for i in 0..n {
            ptr[i + 1] = ptr[i] + l_rows[i].len();
            for &(c, v) in &l_rows[i] {
                flat_cols.push(c);
                flat_vals.push(v);
            }
        }

        Ok(Ic0Factor {
            n,
            ptr,
            flat_cols,
            flat_vals,
        })
    }

    /// Solve (L L^T) x = b.
    pub fn solve(&self, b: &[f64]) -> Vec<f64> {
        let n = self.n;
        let mut y = b.to_vec();
        // Forward: L y' = b
        for i in 0..n {
            let mut s = y[i];
            for k in self.ptr[i]..self.ptr[i + 1] {
                let c = self.flat_cols[k];
                if c < i {
                    s -= self.flat_vals[k] * y[c];
                }
            }
            let d = self.flat_vals[self.ptr[i + 1] - 1];
            y[i] = s / d;
        }
        // Backward: L^T x = y' via scattered updates.
        let mut x = y;
        for i in (0..n).rev() {
            let d = self.flat_vals[self.ptr[i + 1] - 1];
            let xi = x[i] / d;
            x[i] = xi;
            for k in self.ptr[i]..self.ptr[i + 1] - 1 {
                let c = self.flat_cols[k];
                x[c] -= self.flat_vals[k] * xi;
            }
        }
        x
    }
}

/// IC(0)-preconditioned conjugate gradient.
pub fn iccg_solve(a: &SparseMatrix, b: &[f64], max_iter: usize, tol: f64) -> CgResult {
    let ic = match Ic0Factor::factor(a) {
        Ok(f) => f,
        Err(_) => return super::cg_solve(a, b, max_iter, tol),
    };

    let n = b.len();
    let ac = if a.compressed {
        None
    } else {
        let mut m = a.clone();
        m.compress();
        Some(m)
    };
    let a_ref = ac.as_ref().unwrap_or(a);
    let (rptr, ccols, cvals) = a_ref.csr_data();
    let matvec = |p: &[f64], out: &mut [f64]| {
        for row in 0..n {
            let mut sum = 0.0;
            for k in rptr[row]..rptr[row + 1] {
                sum += cvals[k] * p[ccols[k]];
            }
            out[row] = sum;
        }
    };

    let b_norm = b.iter().map(|v| v * v).sum::<f64>().sqrt();
    let mut x = vec![0.0f64; n];
    if b_norm == 0.0 {
        return CgResult {
            x,
            iterations: 0,
            residual: 0.0,
            status: CgStatus::InvalidInput,
        };
    }

    let mut r = b.to_vec();
    let mut z = ic.solve(&r);
    let mut p = z.clone();
    let mut rz: f64 = r.iter().zip(z.iter()).map(|(&a, &b2)| a * b2).sum();
    let mut iterations = 0;

    let mut ap = vec![0.0f64; n];
    for iter in 0..max_iter {
        matvec(&p, &mut ap);
        let pap: f64 = p.iter().zip(ap.iter()).map(|(&a, &b2)| a * b2).sum();
        if pap.abs() <= 0.0 {
            return CgResult {
                x,
                iterations,
                residual: r.iter().map(|v| v * v).sum::<f64>().sqrt(),
                status: CgStatus::Breakdown,
            };
        }
        let alpha = rz / pap;
        if !alpha.is_finite() {
            return CgResult {
                x,
                iterations,
                residual: r.iter().map(|v| v * v).sum::<f64>().sqrt(),
                status: CgStatus::Breakdown,
            };
        }
        for i in 0..n {
            x[i] += alpha * p[i];
            r[i] -= alpha * ap[i];
        }

        let rn = r.iter().map(|v| v * v).sum::<f64>().sqrt();
        iterations = iter + 1;
        if rn < tol * b_norm {
            return CgResult {
                x,
                iterations,
                residual: rn,
                status: CgStatus::Converged,
            };
        }

        z = ic.solve(&r);
        let rz_new: f64 = r.iter().zip(z.iter()).map(|(&a, &b2)| a * b2).sum();
        if rz == 0.0 || !rz.is_finite() {
            return CgResult {
                x,
                iterations,
                residual: r.iter().map(|v| v * v).sum::<f64>().sqrt(),
                status: CgStatus::Breakdown,
            };
        }
        let beta = rz_new / rz;
        if !beta.is_finite() {
            return CgResult {
                x,
                iterations,
                residual: r.iter().map(|v| v * v).sum::<f64>().sqrt(),
                status: CgStatus::Breakdown,
            };
        }
        for i in 0..n {
            p[i] = z[i] + beta * p[i];
        }
        rz = rz_new;
    }

    CgResult {
        x,
        iterations,
        residual: r.iter().map(|v| v * v).sum::<f64>().sqrt(),
        status: CgStatus::MaxIterations,
    }
}

// ---------------------------------------------------------------------------
// PARDISO (Intel MKL) optional FFI binding.
// ---------------------------------------------------------------------------
#[cfg(feature = "pardiso")]
pub mod pardiso {
    //! Direct solve via Intel MKL PARDISO (`mkl_rt`). Enable with
    //! `--features pardiso`; requires `mkl_rt.3.dll` (or equivalent) on PATH.
    use super::super::{CscMatrix, NEAR_ZERO_TOL_BASE, SkylineLdlt, SparseMatrix};
    use std::os::raw::c_void;

    #[cfg_attr(target_os = "windows", link(name = "mkl_rt.2", kind = "raw-dylib"))]
    #[cfg_attr(not(target_os = "windows"), link(name = "mkl_rt.2"))]
    unsafe extern "C" {
        fn pardisoinit(
            pt: *mut c_void,
            maxfct: *const i32,
            mnum: *const i32,
            iparm: *mut i32,
            msglvl: *const i32,
            error: *mut i32,
        );
        fn pardiso(
            pt: *mut c_void,
            maxfct: *const i32,
            mnum: *const i32,
            mtype: *const i32,
            phase: *const i32,
            n: *const i32,
            a: *const f64,
            ia: *const i32,
            ja: *const i32,
            perm: *const i32,
            nrhs: *const i32,
            iparm: *mut i32,
            msglvl: *const i32,
            b: *const f64,
            x: *mut f64,
            error: *mut i32,
        );
    }

    const MAXFCT: i32 = 1;
    const MNUM: i32 = 1;
    /// Symmetric indefinite matrix.
    const MTYPE: i32 = -2;

    /// PARDISO-backed direct solver for the augmented Lagrangian system.
    pub struct PardisoSolver {
        /// Leading block size (K is leading_n x leading_n)
        leading_n: usize,
        /// Augmented matrix size (leading_n + 1)
        n: usize,
        pt: Vec<u64>,
        iparm: Vec<i32>,
        csc: CscMatrix,
        ia: Vec<i32>,
        ja: Vec<i32>,
        vals: Vec<f64>,
        factorised: bool,
        /// Constraint vector c for Lagrange multiplier computation.
        c: Vec<f64>,
        /// Scale factor for scale-invariant tolerances (max diagonal entry of leading block).
        scale: f64,
    }

    unsafe impl Send for PardisoSolver {}

    impl PardisoSolver {
        pub fn new(k: &SparseMatrix, c: &[f64]) -> Result<Self, String> {
            let csc = super::super::DirectLagrangeSolver::assemble_torsion_lagrange(k, c);
            let leading_n = csc.n_rows - 1; // leading block size
            let n = leading_n + 1; // augmented matrix size

            // Compute scale from leading block K (max diagonal)
            let mut scale = 0.0f64;
            let mut k_compressed = k.clone();
            k_compressed.compress();
            for i in 0..leading_n {
                scale = scale.max(k_compressed.matvec_diag(i).abs());
            }
            let scale = scale.max(1.0);

            let mut s = Self {
                leading_n,
                n,
                pt: vec![0u64; 64],
                iparm: vec![0i32; 64],
                csc,
                ia: Vec::new(),
                ja: Vec::new(),
                vals: Vec::new(),
                factorised: false,
                c: c.to_vec(),
                scale,
            };
            unsafe {
                let mut err: i32 = 0;
                let msglvl: i32 = 0;
                pardisoinit(
                    s.pt.as_mut_ptr() as *mut c_void,
                    &MAXFCT,
                    &MNUM,
                    s.iparm.as_mut_ptr(),
                    &msglvl,
                    &mut err,
                );
                if err != 0 {
                    return Err(format!("pardisoinit failed: {err}"));
                }
                // Zero-based indexing for ia/ja arrays.
                s.iparm[34] = 1;
            }
            Ok(s)
        }

        fn ensure_factorised(&mut self) -> Result<(), String> {
            if self.factorised {
                return Ok(());
            }
            self.ia = self.csc.col_ptr.iter().map(|&v| v as i32).collect();
            self.ja = self.csc.rows.iter().map(|&v| v as i32).collect();
            self.vals = self.csc.vals.clone();

            unsafe {
                let mut err: i32 = 0;
                let phase: i32 = 13; // analyse + numerical factorisation
                let n = self.n as i32;
                let nrhs: i32 = 0;
                let msglvl: i32 = 0;
                let dummy_b: f64 = 0.0;
                let mut dummy_x: f64 = 0.0;
                pardiso(
                    self.pt.as_mut_ptr() as *mut c_void,
                    &MAXFCT,
                    &MNUM,
                    &MTYPE,
                    &phase,
                    &n,
                    self.vals.as_ptr(),
                    self.ia.as_ptr(),
                    self.ja.as_ptr(),
                    std::ptr::null(),
                    &nrhs,
                    self.iparm.as_mut_ptr(),
                    &msglvl,
                    &dummy_b,
                    &mut dummy_x,
                    &mut err,
                );
                if err != 0 {
                    return Err(format!("pardiso factorisation failed: {err}"));
                }
            }
            self.factorised = true;
            Ok(())
        }

        /// Solve [K c; c^T 0] [u; lam] = [f; 0]; returns (u, lam).
        pub fn solve_with_multiplier(&mut self, f: &[f64]) -> Result<(Vec<f64>, f64), String> {
            self.ensure_factorised()?;
            let n = self.n; // augmented matrix size
            let mut b = f.to_vec();
            b.push(0.0); // size becomes n
            let mut x = vec![0.0f64; n];

            unsafe {
                let mut err: i32 = 0;
                let phase: i32 = 33; // solve + iterative refinement
                let nn = n as i32; // pass augmented size directly
                let nrhs: i32 = 1;
                let msglvl: i32 = 0;
                pardiso(
                    self.pt.as_mut_ptr() as *mut c_void,
                    &MAXFCT,
                    &MNUM,
                    &MTYPE,
                    &phase,
                    &nn,
                    self.vals.as_ptr(),
                    self.ia.as_ptr(),
                    self.ja.as_ptr(),
                    std::ptr::null(),
                    &nrhs,
                    self.iparm.as_mut_ptr(),
                    &msglvl,
                    b.as_ptr(),
                    x.as_mut_ptr(),
                    &mut err,
                );
                if err != 0 {
                    return Err(format!("pardiso solve failed: {err}"));
                }
            }

            let lam = x.pop().unwrap_or(0.0);
            Ok((x, lam))
        }

        /// Solve [K c; c^T 0] [u; lam] = [f; 0]; returns u.
        pub fn solve_direct_lagrange(&mut self, f: &[f64]) -> Result<Vec<f64>, String> {
            self.ensure_factorised()?;
            let n = self.n; // augmented matrix size
            let mut b = f.to_vec();
            b.push(0.0); // size becomes n
            let mut x = vec![0.0f64; n];

            unsafe {
                let mut err: i32 = 0;
                let phase: i32 = 33; // solve + iterative refinement
                let nn = n as i32;
                let nrhs: i32 = 1;
                let msglvl: i32 = 0;
                pardiso(
                    self.pt.as_mut_ptr() as *mut c_void,
                    &MAXFCT,
                    &MNUM,
                    &MTYPE,
                    &phase,
                    &nn,
                    self.vals.as_ptr(),
                    self.ia.as_ptr(),
                    self.ja.as_ptr(),
                    std::ptr::null(),
                    &nrhs,
                    self.iparm.as_mut_ptr(),
                    &msglvl,
                    b.as_ptr(),
                    x.as_mut_ptr(),
                    &mut err,
                );
                if err != 0 {
                    return Err(format!("pardiso solve failed: {err}"));
                }
            }

            x.pop();
            Ok(x)
        }

        /// Multiplier error metric |lam| / max|u| (Python's u[-1]/max|u|).
        pub fn multiplier_error(&self, c: &[f64], f: &[f64], u: &[f64]) -> f64 {
            // lam can be recovered from the augmented solve; approximate via
            // residual of the constraint row is out of scope here, so mirror
            // the skyline implementation by refactoring K once more.
            let ldlt = match SkylineLdlt::factor(&{
                let m = {
                    // leading block from CSC columns 0..leading_n
                    let leading_n = self.leading_n;
                    let mut sm = SparseMatrix::new(leading_n);
                    for col in 0..leading_n {
                        for k in self.csc.col_ptr[col]..self.csc.col_ptr[col + 1] {
                            let r = self.csc.rows[k];
                            if r < leading_n && col < leading_n {
                                sm.add(r, col, self.csc.vals[k]);
                            }
                        }
                    }
                    sm.compress();
                    sm
                };
                m
            }) {
                Ok(l) => l,
                Err(_) => return f64::INFINITY,
            };
            let w1 = ldlt.solve(f).unwrap_or_else(|_| vec![0.0; c.len()]);
            let w2 = ldlt.solve(c).unwrap_or_else(|_| vec![0.0; c.len()]);
            let mut ct_w2 = 0.0f64;
            for (ci, wi) in c.iter().zip(w2.iter()) {
                ct_w2 += ci * wi;
            }
            let mut ct_w1 = 0.0f64;
            for (ci, wi) in c.iter().zip(w1.iter()) {
                ct_w1 += ci * wi;
            }
            // Scale-invariant tolerance for Lagrange multiplier
            let near_zero_tol = NEAR_ZERO_TOL_BASE * self.scale.max(1.0);
            // u = w1 - lambda*w2 with constraint c^T u = 0 gives
            // lambda = (c^T w1) / (c^T w2) = ct_w1 / ct_w2.
            let lambda = if ct_w2.abs() > near_zero_tol {
                ct_w1 / ct_w2
            } else {
                0.0
            };
            let max_u = u.iter().fold(0.0f64, |a, &v| a.max(v.abs()));
            if max_u > 0.0 {
                lambda.abs() / max_u
            } else {
                f64::INFINITY
            }
        }
    }

    impl Drop for PardisoSolver {
        fn drop(&mut self) {
            if !self.factorised {
                return;
            }
            unsafe {
                let mut err: i32 = 0;
                let phase: i32 = -1; // release memory
                let n = self.n as i32;
                let nrhs: i32 = 0;
                let msglvl: i32 = 0;
                let dummy: f64 = 0.0;
                let mut x_dummy: f64 = 0.0;
                pardiso(
                    self.pt.as_mut_ptr() as *mut c_void,
                    &MAXFCT,
                    &MNUM,
                    &MTYPE,
                    &phase,
                    &n,
                    self.vals.as_ptr(),
                    self.ia.as_ptr(),
                    self.ja.as_ptr(),
                    std::ptr::null(),
                    &nrhs,
                    self.iparm.as_mut_ptr(),
                    &msglvl,
                    &dummy,
                    &mut x_dummy,
                    &mut err,
                );
            }
        }
    }
}
#[cfg(all(feature = "pardiso", test))]
mod pardiso_tests {
    use super::pardiso::PardisoSolver;
    use crate::fea::SparseMatrix;

    #[test]
    fn pardiso_laplacian_smoke() {
        // Requires mkl_rt.3.dll on PATH at runtime.
        let n = 20;
        let mut k = SparseMatrix::new(n);
        for i in 0..n {
            k.add(i, i, 2.0);
            if i + 1 < n {
                k.add(i, i + 1, -1.0);
                k.add(i + 1, i, -1.0);
            }
        }
        let c: Vec<f64> = vec![1.0; n];
        let f: Vec<f64> = (0..n).map(|i| (i as f64) * 0.05).collect();

        let mut s = PardisoSolver::new(&k, &c).expect("pardiso init");
        let u = s.solve_direct_lagrange(&f).expect("pardiso solve");
        assert_eq!(u.len(), n);
        // residual check
        let mut kmat = SparseMatrix::new(n);
        for i in 0..n {
            kmat.add(i, i, 2.0);
            if i + 1 < n {
                kmat.add(i, i + 1, -1.0);
                kmat.add(i + 1, i, -1.0);
            }
        }
        kmat.compress();
        let prod = kmat.matvec(&u);
        for i in 0..n {
            assert!(
                (prod[i] - f[i] + c[i]).abs() < 1e-6 || (prod[i] - f[i]).abs() < 1e-6,
                "row {} residual too large",
                i
            );
        }
    }
}