decimal-scaled 0.5.0

Const-generic base-10 fixed-point decimals (D18/D38/D76/D153/D307 and the half-width tiers up to D1232) with integer-only transcendentals correctly rounded to within 0.5 ULP — exact at the type's last representable place. Deterministic across every platform; no_std-friendly.
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
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
// SPDX-FileCopyrightText: 2026 John Moxley
// SPDX-License-Identifier: MIT OR Apache-2.0

//! Newton–Raphson reciprocal divide for `n / 10^SCALE` at storage width.
//!
//! Research kernel — **not wired into the dispatcher**. Built behind a
//! `pub(crate)` API so micro-benches can compare it head-to-head against
//! [`crate::algos::support::mg_divide::div_wide_pow10_chain`].
//!
//! # Algorithm
//!
//! For invariant divisor `D = 10^SCALE`, precompute a fixed-point
//! reciprocal
//!
//! ```text
//!   R = floor(2^k / D)
//! ```
//!
//! where `k` is chosen so that `k - bit_length(D) ≥ bit_length(N_max)`,
//! i.e. `R` carries enough fractional bits to represent the storage-width
//! numerator's worth of quotient. The per-call divide reduces to
//!
//! ```text
//!   q_approx = (n * R) >> k
//!   r        = n - q_approx * D
//!   if r >= D { q_approx += 1; r -= D; }   // single correction step
//! ```
//!
//! The estimate `q_approx` is off by at most 1 (analogous to the
//! Möller-Granlund add-back correction), so a single comparison suffices
//! after the multiply.
//!
//! # Setup
//!
//! `R` is computed once per `(SCALE, width)` pair via the int-algos-layer
//! variable-length divmod [`crate::int::algos::div::div_rem_mag_slice`].
//! Setup cost is one wide divide; per-call cost is one wide multiply +
//! one narrow multiply + one comparison + one optional subtract.
//!
//! # Storage — raw `u64` limb slices, below the `Int<N>` abstraction
//!
//! All scratch is held in fixed-size `u64` limb buffers (little-endian),
//! `core`-only — no heap, no `alloc`. The `BigInt` magnitude/sign bridge
//! still moves through the `u128`-limb buffer (`mag_into_u128` /
//! `from_mag_sign_u128`), but every arithmetic step runs on `u64` limbs.
//!
//! This kernel deliberately does **not** route through the `Int<N>` type
//! methods (`Int::<N>::div_rem`, `*`, …). The reciprocal `R` and the
//! `10^SCALE` divisor have a **runtime live limb count** (`k_u64 + 1` and
//! `pow_len`, both functions of the runtime `scale` / `width` arguments)
//! that no const-generic `Int<N>` width can express: a single
//! `precompute` call can produce an `R` anywhere from ~18 to 144 u64
//! limbs. `Int::<N>::div_rem` fixes `N` at monomorphisation, so the
//! variable-width reciprocal divide cannot be expressed through it. The
//! per-call multiplies are likewise over variable-length live slices. The
//! kernel therefore stays on raw `u64` slices — but it routes its work
//! through the dispatching matchers rather than hardcoding a leaf kernel:
//! the divmod through the int-**algos** layer
//! ([`crate::int::algos::div::div_rem_mag_slice`], which fronts the
//! divisor-shape policy and picks the optimal engine), and the u64 kernel's
//! per-call multiplies through the mul matcher's runtime-length slice door
//! ([`crate::int::policy::mul::dispatch_slice`] — the sanctioned entry for
//! bare-`&[u64]` slice callers), so the matcher chooses the multiply engine
//! (schoolbook / Karatsuba) per operand shape. The u128-packed sibling
//! ([`div_newton_u128`]) keeps its own local u128 schoolbook — the slice
//! door is `u64`-only (no `u128` matcher door exists). All are DOWN-calls
//! into the int matchers, so the decimal→int layering stays intact.
//!
//! # Reference
//!
//! Granlund, T. & Montgomery, P. L. (1994). *Division by Invariant
//! Integers using Multiplication*, PLDI '94. Möller, N. & Granlund, T.
//! (2011). *Improved Division by Invariant Integers*, IEEE TC 60(2).
//! The Newton-iteration view of the same reciprocal is
//! Wikipedia — [Division algorithm § Newton–Raphson division](https://en.wikipedia.org/wiki/Division_algorithm#Newton%E2%80%93Raphson_division).

use crate::int::algos::support::limbs::{cmp, sub_assign};
use crate::int::policy::mul::dispatch_slice as mul_slice;

// ── Fixed buffer sizing (in u64 limbs) ──────────────────────────────
//
// The widest cell exercised is `width_limbs = 96` u64 limbs (6144-bit,
// the D230 Wexp / D924 Work tier). Working in u64 limbs (two per u128
// limb), the worst-case sizes — taking the AGM-widened scale ceiling
// `w_prime = 2·SCALE + 4` (D924 max SCALE 923 → w_prime ≤ 1850) — are:
//
//   pow_scale : pow_u64 = scale/19 + 3 ≤ 100 u64 (at w_prime ≤ 1850)
//   r         : (k_u64 + 1) = (width + pow + 1) ≤ 96 + 100 + 1 = 197 u64
//   mag (n)   : 96 u64 (Int<96>); the 4096 path uses ≤ 64
//   product   : n.len() + r.len() ≤ 128 + 200 = 328 u64
//
// All buffers are over-sized to a single ceiling that covers every
// width the matcher routes Newton-vs-MG against, so the same type
// serves every tier without const-generic gymnastics.
//
// The build-max is internal to the runtime-sized `NewtonReciprocal`
// reciprocal struct — it never leaks onto a concrete-`N` path (those
// still size their scratch via `ComputeLimbs::single_*` etc. per
// Constitution rule 6). Over-sizing here costs constant per-call stack,
// not per-tier code duplication.
//
// The 8192 / 12288 / 16384 / 32768 widths (D462 Wexp / D1232 Work /
// D924 Wide / D616 Wexp / D924
// Wexp / D1232 Wide / D1232 Wexp) are deferred. At those widths the
// Newton precompute's `2^k / 10^scale` numerator at the AGM-widened
// scale exceeds the routed `div_knuth`'s `MAX_SINGLE_LIMBS = 258`
// build-max scratch — D1232 strict_agm runs at `w_prime ≈ 2466`
// which puts an 8192-bit Newton's numerator at 260+ limbs, and a
// 12288-bit Newton's at 293+, both past the build-max scratch the
// routed Knuth can hold. The atanh-diagnosis bench
// also reported Newton LOSING by 5–58× at Int<192>/Int<256> w=38
// (low-scale shape), so even with extended scratch the integrated
// picture isn't settled. Revisit when both the wider-numerator divide
// scratch and the AGM-scale `newton_vs_mg` evidence line up.

/// Max `u64` limbs for the `10^SCALE` (`pow_scale`) buffer. Covers the
/// D924 AGM scale at width 6144 (`w_prime ≤ 1850`, `pow_u64 ≤ 100`).
const MAX_POW_U64: usize = 104;
/// Max `u64` limbs for the reciprocal (`r`) buffer. Covers the widest baked
/// table width (132 u64 = the split Tang work for D1232) at the top scale
/// (`k_u64 + 1 ≤ 237`), with margin.
const MAX_R_U64: usize = 240;
/// Max `u64` limbs for the magnitude / quotient buffers. Covers the widest
/// work integer's magnitude (the split D1232 Tang work = 132 u64), with margin.
const MAX_MAG_U64: usize = 136;
/// Max `u64` limbs for product / scratch buffers (`n·r`, `q·D`, …).
/// Covers `MAX_MAG_U64 + MAX_R_U64`.
const MAX_PROD_U64: usize = MAX_MAG_U64 + MAX_R_U64;

// -- u128-limb sibling sizes (packed pairs of u64) --------------------
//
// The precomputed `r_u128`/`pow_u128` mirror the u64 versions packed pairwise
// (`limb = lo | hi << 64`). All sizes are `ceil(u64_size/2)`.

/// Max `u128` limbs for the `10^scale` (`pow_u128`) buffer.
const MAX_POW_U128: usize = MAX_POW_U64.div_ceil(2);
/// Max `u128` limbs for the reciprocal (`r_u128`) buffer.
const MAX_R_U128: usize = MAX_R_U64.div_ceil(2);
/// Max `u128` limbs for the magnitude / quotient u128 buffers.
const MAX_MAG_U128: usize = MAX_MAG_U64.div_ceil(2);
/// Max `u128` limbs for product / scratch u128 buffers.
const MAX_PROD_U128: usize = MAX_PROD_U64.div_ceil(2);

/// Pre-computed reciprocal table for a single `(SCALE, mag_width)` pair.
///
/// `r` is the reciprocal `floor(2^k / 10^SCALE)` in little-endian
/// u64 limbs; `k_u64` is `k / 64` (we always pick `k` as a multiple of
/// 64 so the shift is a limb-aligned slice).
///
/// `pow_scale` is `10^SCALE` in little-endian u64 limbs, kept for the
/// correction step.
///
/// All storage is fixed-size — no heap. `r_len` / `pow_len` record the
/// live limb counts within the over-sized backing arrays.
#[derive(Clone)]
pub struct NewtonReciprocal {
    /// Reciprocal limbs (little-endian, u64), live count `r_len`.
    r: [u64; MAX_R_U64],
    /// Live limb count of `r`.
    r_len: usize,
    /// Right-shift amount in u64 limbs (quotient = (n·r) limbs >> k_u64 words).
    k_u64: usize,
    /// `10^SCALE` limbs (little-endian, u64), live count `pow_len`.
    pow_scale: [u64; MAX_POW_U64],
    /// Live limb count of `pow_scale`.
    pow_len: usize,

    // -- u128-packed mirrors of `r` and `pow_scale` -------------------
    //
    // Populated once at the end of `precompute` by pairwise packing the
    // u64 limbs above (`limb = lo | hi << 64`). The u128 Newton kernel
    // (`div_newton_u128`) consumes these directly with NO per-call pack,
    // achieving u128-slice mul throughput at the cost of one
    // extra pack pass during the (already amortised) precompute.
    /// Reciprocal limbs packed as u128 pairs, live count `r_u128_len`.
    r_u128: [u128; MAX_R_U128],
    /// Live limb count of `r_u128` = `r_len.div_ceil(2)`.
    r_u128_len: usize,
    /// Right-shift amount in u128 limbs (`k_u64 / 2`). `precompute`
    /// rounds `k_u64` UP to even so this shift is limb-aligned in u128.
    k_u128: usize,
    /// `10^scale` packed as u128 pairs, live count `pow_u128_len`.
    pow_u128: [u128; MAX_POW_U128],
    /// Live limb count of `pow_u128` = `pow_len.div_ceil(2)`.
    pow_u128_len: usize,
}

impl NewtonReciprocal {
    /// Compute reciprocal table for `D = 10^scale` at the given
    /// magnitude width.
    ///
    /// `width_u64_limbs` is the upper bound on the numerator magnitude's
    /// limb count, expressed in **u64 limbs** — the unit the kernel's
    /// arithmetic actually runs in.
    pub fn precompute(scale: u32, width_u64_limbs: usize) -> Self {
        let width_limbs = width_u64_limbs;

        // pow_scale = 10^scale. `pow_len` reserves scale/19 + 3 u64 limbs
        // (matches the prior u128 path's scale/38 + 2 u128 budget, doubled);
        // the high limbs above the value stay zero either way, so the divide
        // and downstream `k_u64` are identical regardless of how it is built.
        let pow_len = (scale as usize / 19 + 3).max(1);
        debug_assert!(pow_len <= MAX_POW_U64, "pow_scale buffer too small");
        let mut pow_scale = [0u64; MAX_POW_U64];
        if let Some(limbs) = crate::consts::pow10_limbs(scale) {
            // Read 10^scale from the baked POW10 const table — a memcpy that
            // replaces the per-call O(scale²) ×10 chain (the dominant rescale
            // cost). Same value, same `pow_len` (the high limbs stay zero), so
            // bit-identical to the chain.
            debug_assert!(limbs.len() <= pow_len, "pow10 table entry exceeds pow_len");
            pow_scale[..limbs.len()].copy_from_slice(limbs);
        } else {
            // Rare deep tail beyond the (feature-gated) table: rebuild via *10.
            pow_scale[0] = 1u64;
            for _ in 0..scale {
                let mut carry: u64 = 0;
                for limb in pow_scale[..pow_len].iter_mut() {
                    let prod = (*limb as u128) * 10u128 + (carry as u128);
                    *limb = prod as u64;
                    carry = (prod >> 64) as u64;
                }
                debug_assert_eq!(carry, 0, "pow_scale buffer too small at scale={scale}");
            }
        }

        // Pick k_u64: quotient room of `width_limbs` u64 limbs. Set
        // k = 64 * (width_limbs + pow_len) bits — then R = 2^k / 10^scale
        // has bit-length about k - bits(10^scale), and (n·R) >> k yields
        // a width_limbs-wide quotient with at most 1 ULP error.
        //
        // Round UP to even so the u128-packed mirror right-shift is
        // limb-aligned in u128 (`k_u128 = k_u64 / 2`). Adding one to k
        // grows R by a single limb (over-estimate); the Newton add-back
        // correction absorbs the +1 ULP. The u64 kernel sees the bumped
        // `k_u64` too — bit-identical correction either way.
        let k_u64_raw = width_limbs + pow_len;
        let k_u64 = if k_u64_raw.is_multiple_of(2) { k_u64_raw } else { k_u64_raw + 1 };

        debug_assert!(k_u64 < MAX_R_U64, "num buffer too small");

        // r = floor(2^(64*k_u64) / 10^scale). FAST: read the baked reciprocal
        // (`consts::newton_recip` — the high `k_u64+1` limbs of the width-96
        // per-scale reciprocal; bit-identical to the divide below, both exact
        // integer floor division — `2^(64k)//10^s`). SLOW fallback (scale/width
        // outside the baked range, or a non-wide build where the table is
        // gated out → `None`): the one-shot Knuth divide.
        let mut r = [0u64; MAX_R_U64];
        if let Some(baked) = crate::consts::newton_recip_le(scale, width_limbs) {
            debug_assert_eq!(baked.len(), k_u64 + 1, "baked newton recip length");
            r[..k_u64 + 1].copy_from_slice(baked);
        } else {
            // numerator = 2^(64 * k_u64) — a single 1 in limb position k_u64.
            let mut num = [0u64; MAX_R_U64];
            num[k_u64] = 1u64;
            let mut rem = [0u64; MAX_POW_U64];
            // EXACT module-sized Knuth scratch — never the build-max slice
            // blanket. The dividend spans `k_u64 + 1 ≤ MAX_R_U64` limbs
            // (asserted above), which exceeds the blanket divide's
            // `MAX_SINGLE_LIMBS = 4·MAX_WORK_N + 2` normalisation buffer on
            // a narrow-blanket build (67 > 66 limbs at `width_limbs = 42`,
            // `scale = 407` under `MAX_WORK_N = 16` — the defect-B in-range
            // panic). Knuth needs `num.len() + 2` / `den.len()` zeroed
            // limbs, so size both from this module's own ceilings
            // (`MAX_R_U64 + 2` ≥ `k_u64 + 3`, `MAX_POW_U64` ≥ `pow_len`),
            // route on the divide matcher's own shape verdict, and call the
            // engine's `_into` door — the exact-scratch pattern
            // `int::policy::div_rem::dispatch`'s doc sanctions (the
            // `exp_generic::div_rem_exact` precedent; the u128-limb
            // refinement falls to the value-identical base-2⁶⁴ Knuth).
            use crate::int::policy::div_rem::{select_for_limbs, Algorithm};
            match select_for_limbs(&num[..k_u64 + 1], &pow_scale[..pow_len]) {
                // Single-limb divisor: hardware remainder, no normalisation
                // scratch involved.
                Algorithm::Rem => crate::int::algos::div::div_rem::div_rem(
                    &num[..k_u64 + 1],
                    &pow_scale[..pow_len],
                    &mut r[..k_u64 + 1],
                    &mut rem[..pow_len],
                ),
                _ => {
                    // Verdict collapse: any non-Rem algorithm (including the
                    // u128-limb verdict) funnels here and runs Knuth. Correct
                    // for every current `pow_scale`-denominator shape — the
                    // u128 engine's large-dividend precondition is not met —
                    // but MUST be re-verified whenever an Algorithm arm joins
                    // `int::policy::div_rem`; a new engine winning for these
                    // shapes would need an explicit arm, not a silent Knuth fall.
                    let mut u = [0u64; MAX_R_U64 + 2];
                    let mut v = [0u64; MAX_POW_U64];
                    crate::int::algos::div::div_knuth::div_knuth_into(
                        &num[..k_u64 + 1],
                        &pow_scale[..pow_len],
                        &mut r[..k_u64 + 1],
                        &mut rem[..pow_len],
                        &mut u,
                        &mut v,
                    );
                }
            }
        }

        // -- u128-packed mirrors ----------------------------------
        //
        // Pack the u64 limbs of `r` and `pow_scale` pairwise into u128
        // limbs (`limb = lo | hi << 64`). Buffers are zero-initialised,
        // so packing the rounded-even live count is safe (any odd live
        // tail of the u64 buffer pairs with a zeroed neighbour).
        let r_len = k_u64 + 1;
        let r_u128_len = r_len.div_ceil(2);
        let pow_u128_len = pow_len.div_ceil(2);

        let mut r_u128 = [0u128; MAX_R_U128];
        let mut i = 0;
        while i < r_u128_len {
            let lo = r[2 * i] as u128;
            let hi = r[2 * i + 1] as u128;
            r_u128[i] = lo | (hi << 64);
            i += 1;
        }
        let mut pow_u128 = [0u128; MAX_POW_U128];
        let mut i = 0;
        while i < pow_u128_len {
            let lo = pow_scale[2 * i] as u128;
            let hi = pow_scale[2 * i + 1] as u128;
            pow_u128[i] = lo | (hi << 64);
            i += 1;
        }
        let k_u128 = k_u64 / 2;

        Self {
            r,
            r_len,
            k_u64,
            pow_scale,
            pow_len,
            r_u128,
            r_u128_len,
            k_u128,
            pow_u128,
            pow_u128_len,
        }
    }
}

/// Per-call Newton-reciprocal divide.
///
/// `n` is the unsigned numerator magnitude in little-endian u64 limbs.
/// The quotient `floor(n / 10^scale)` is written into `quot` (caller-
/// sized to the target width); the remainder is written into `rem_out`
/// and its live limb count returned, for rounding-aware callers.
///
/// # Precision
///
/// Strict: the result is bit-exact `floor(n / 10^scale)`. The Newton
/// add-back step ensures correctness for the at-most-1 over/under
/// estimate the truncated reciprocal produces.
fn div_newton(
    n: &[u64],
    table: &NewtonReciprocal,
    quot: &mut [u64],
    rem_out: &mut [u64],
) -> usize {
    let r = &table.r[..table.r_len];
    let pow_scale = &table.pow_scale[..table.pow_len];

    // product = n * r
    let prod_len = n.len() + r.len();
    debug_assert!(prod_len <= MAX_PROD_U64, "product buffer too small");
    let mut prod = [0u64; MAX_PROD_U64];
    mul_slice(n, r, &mut prod[..prod_len]);

    // q_approx = prod >> (64 * k_u64)
    let lo = table.k_u64.min(prod_len);
    let q_slice = &prod[lo..prod_len];
    for (dst, src) in quot.iter_mut().zip(q_slice.iter()) {
        *dst = *src;
    }
    for dst in quot.iter_mut().skip(q_slice.len()) {
        *dst = 0;
    }

    // r_approx = n - q_approx * pow_scale  (mod 2^width)
    let prod2_len = quot.len() + pow_scale.len();
    debug_assert!(prod2_len <= MAX_PROD_U64, "product buffer too small");
    let mut prod2 = [0u64; MAX_PROD_U64];
    mul_slice(quot, pow_scale, &mut prod2[..prod2_len]);

    // rem = n - prod2 (mod 2^width), held in n.len()+1 limbs.
    let rem_len = n.len() + 1;
    debug_assert!(rem_len <= MAX_MAG_U64 + 1, "rem buffer too small");
    for (dst, src) in rem_out.iter_mut().take(rem_len).zip(n.iter()) {
        *dst = *src;
    }
    rem_out[rem_len - 1] = 0;
    let sub_len = prod2_len.min(rem_len);
    let _ = sub_assign(&mut rem_out[..sub_len], &prod2[..sub_len]);

    // Correction loop: while rem >= pow_scale, bump quotient by 1 and
    // decrement remainder. With a correctly-sized k_u64 the loop runs at
    // most once or twice.
    loop {
        if cmp(&rem_out[..rem_len], pow_scale) < 0 {
            break;
        }
        let s = rem_len.min(pow_scale.len());
        let _ = sub_assign(&mut rem_out[..s], &pow_scale[..s]);
        // quot += 1
        let mut carry: u64 = 1;
        for limb in quot.iter_mut() {
            let (s, c) = limb.overflowing_add(carry);
            *limb = s;
            if !c {
                carry = 0;
                break;
            }
        }
        let _ = carry;
    }

    rem_len
}

// -- u128-packed Newton kernel ----------------------------------------
//
// Mirrors `div_newton` but operates entirely on packed u128 limb slices.
// The precomputed `r_u128`/`pow_u128` are consumed directly with NO per-call
// pack/unpack; the per-call operand (n) and output (quot, rem) all stay
// in u128 throughout. Achieves `limbs_mul`-style throughput
// (half the limb count, ~1/4 the partial products per schoolbook) by
// paying the pack cost ONCE in the amortised `precompute`.

#[inline]
const fn cmp_u128(a: &[u128], b: &[u128]) -> i32 {
    let mut alen = a.len();
    while alen > 0 && a[alen - 1] == 0 { alen -= 1; }
    let mut blen = b.len();
    while blen > 0 && b[blen - 1] == 0 { blen -= 1; }
    if alen != blen { return if alen > blen { 1 } else { -1 }; }
    let mut i = alen;
    while i > 0 {
        i -= 1;
        if a[i] != b[i] { return if a[i] > b[i] { 1 } else { -1 }; }
    }
    0
}

#[inline]
const fn sub_assign_u128(a: &mut [u128], b: &[u128]) -> bool {
    let mut borrow: u128 = 0;
    let mut i = 0;
    while i < a.len() {
        let bi = if i < b.len() { b[i] } else { 0 };
        let (s1, c1) = a[i].overflowing_sub(bi);
        let (s2, c2) = s1.overflowing_sub(borrow);
        a[i] = s2;
        borrow = (c1 as u128) | (c2 as u128);
        i += 1;
    }
    borrow != 0
}

/// `out = a * b` schoolbook on u128 limb slices. Inner step uses the
/// 4xu64*u64->u128 partials decomposition (`<u128 as Limb>::widening_mul`).
#[inline]
fn mul_schoolbook_u128(a: &[u128], b: &[u128], out: &mut [u128]) {
    use crate::int::types::compute_limbs::Limb;
    let mut i = 0;
    while i < a.len() {
        if a[i] != 0 {
            let mut carry: u128 = 0;
            let mut j = 0;
            while j < b.len() {
                if b[j] != 0 || carry != 0 {
                    let (prod_lo, prod_hi) = <u128 as Limb>::widening_mul(a[i], b[j]);
                    let idx = i + j;
                    let (s1, c1) = out[idx].overflowing_add(prod_lo);
                    let (s2, c2) = s1.overflowing_add(carry);
                    out[idx] = s2;
                    carry = prod_hi.wrapping_add(c1 as u128).wrapping_add(c2 as u128);
                }
                j += 1;
            }
            let mut idx = i + b.len();
            while carry != 0 && idx < out.len() {
                let (s, c) = out[idx].overflowing_add(carry);
                out[idx] = s;
                carry = c as u128;
                idx += 1;
            }
        }
        i += 1;
    }
}

/// HIGH product: `out[i] = limb (base + i)` of `a * b` (the full product has
/// `a.len()+b.len()` limbs), forming only the partials with `i + j >= base`.
/// The dropped low partials (`i + j < base`) are NOT included, so their carry
/// into limb `base` is missing — making the high limbs too-LOW by a bounded
/// amount. The Newton quotient `q = (a·b) >> k` reads this with a small guard
/// (`base = k - GUARD`), bounding the deficit to `< 1` ULP at the `k`-cut; the
/// Newton correction LOOP (`while rem >= D { q += 1 }`) adds back any residual,
/// so the result is EXACT regardless of guard — the guard only bounds the loop
/// count (perf). `out` pre-zeroed; `out.len() >= a.len()+b.len() - base`. This
/// halves the `mag·r` work vs the full product (only the high half is kept).
#[inline]
fn mul_high_schoolbook_u128(a: &[u128], b: &[u128], out: &mut [u128], base: usize) {
    use crate::int::types::compute_limbs::Limb;
    let out_len = out.len();
    let mut i = 0;
    while i < a.len() {
        if a[i] != 0 {
            // Smallest j with i + j >= base; skip the dropped-low partials.
            let j0 = base.saturating_sub(i);
            let mut carry: u128 = 0;
            let mut j = j0;
            while j < b.len() {
                let (lo, hi) = <u128 as Limb>::widening_mul(a[i], b[j]);
                let idx = i + j - base; // i + j >= base for j >= j0
                let (s1, c1) = out[idx].overflowing_add(lo);
                let (s2, c2) = s1.overflowing_add(carry);
                out[idx] = s2;
                carry = hi.wrapping_add(c1 as u128).wrapping_add(c2 as u128);
                j += 1;
            }
            let mut idx = i + b.len() - base;
            while carry != 0 && idx < out_len {
                let (s, c) = out[idx].overflowing_add(carry);
                out[idx] = s;
                carry = c as u128;
                idx += 1;
            }
        }
        i += 1;
    }
}

/// Per-call Newton-reciprocal divide, u128-packed sibling of `div_newton`.
/// All multiplies run on the precomputed u128-packed `r` and `pow_scale` (NO
/// per-call pack); operand/output stay in u128 throughout.
fn div_newton_u128(
    n: &[u128],
    table: &NewtonReciprocal,
    quot: &mut [u128],
    rem_out: &mut [u128],
) -> usize {
    let r = &table.r_u128[..table.r_u128_len];
    let pow_scale = &table.pow_u128[..table.pow_u128_len];

    let prod_len = n.len() + r.len();
    let lo = table.k_u128.min(prod_len);
    // q = (n·r) >> (128·k_u128) reads only the HIGH limbs [lo..]. Form just
    // those (plus GUARD guard limbs) with the high-product — dropping the low
    // partials halves the mag·r work and makes q too-LOW by < 1 ULP, which the
    // correction loop below adds back (exact regardless of guard).
    const GUARD: usize = 2;
    let base = lo.saturating_sub(GUARD);
    let high_len = prod_len - base;
    debug_assert!(high_len <= MAX_PROD_U128, "u128 high-product buffer too small");
    let mut prod = [0u128; MAX_PROD_U128];
    mul_high_schoolbook_u128(n, r, &mut prod[..high_len], base);

    // q = limbs [lo..prod_len] of n·r = prod[lo - base .. high_len].
    let q_slice = &prod[lo - base..high_len];
    for (dst, src_) in quot.iter_mut().zip(q_slice.iter()) { *dst = *src_; }
    for dst in quot.iter_mut().skip(q_slice.len()) { *dst = 0; }

    let prod2_len = quot.len() + pow_scale.len();
    debug_assert!(prod2_len <= MAX_PROD_U128, "u128 product buffer too small");
    let mut prod2 = [0u128; MAX_PROD_U128];
    mul_schoolbook_u128(quot, pow_scale, &mut prod2[..prod2_len]);

    let rem_len = n.len() + 1;
    debug_assert!(rem_len <= MAX_MAG_U128 + 1, "u128 rem buffer too small");
    for (dst, src_) in rem_out.iter_mut().take(rem_len).zip(n.iter()) { *dst = *src_; }
    rem_out[rem_len - 1] = 0;
    let sub_len = prod2_len.min(rem_len);
    let _ = sub_assign_u128(&mut rem_out[..sub_len], &prod2[..sub_len]);

    loop {
        if cmp_u128(&rem_out[..rem_len], pow_scale) < 0 { break; }
        let s = rem_len.min(pow_scale.len());
        let _ = sub_assign_u128(&mut rem_out[..s], &pow_scale[..s]);
        let mut carry: u128 = 1;
        for limb in quot.iter_mut() {
            let (s, c) = limb.overflowing_add(carry);
            *limb = s;
            if !c { carry = 0; break; }
        }
        let _ = carry;
    }

    rem_len
}

/// Width-agnostic Newton-reciprocal divide of a u128 magnitude slice by
/// `10^scale`, in place, with `mode`-aware rounding. Operates entirely
/// in u128 limbs (NO transcoding to/from u64). Bit-identical to the u64
/// path `newton_pow10_mag_u128`.
pub(crate) fn newton_pow10_mag_u128_packed(
    mag_u128: &mut [u128],
    neg: bool,
    mode: crate::support::rounding::RoundingMode,
    table: &NewtonReciprocal,
) {
    use crate::support::rounding;

    let mag_len = mag_u128.len();
    let mut top = mag_len;
    while top > 0 && mag_u128[top - 1] == 0 { top -= 1; }

    let n_slice = &mag_u128[..top.max(1)];
    let mut quot = [0u128; MAX_MAG_U128];
    let mut rem = [0u128; MAX_MAG_U128 + 1];
    let rem_len = div_newton_u128(n_slice, table, &mut quot[..mag_len], &mut rem);

    let rem_is_zero = rem[..rem_len].iter().all(|&x| x == 0);
    if !rem_is_zero {
        let pow_len = table.pow_u128_len;
        let mut half = [0u128; MAX_POW_U128];
        half[..pow_len].copy_from_slice(&table.pow_u128[..pow_len]);
        let mut i = pow_len;
        let mut carry_in: u128 = 0;
        while i > 0 {
            i -= 1;
            let next_carry = half[i] & 1;
            half[i] = (carry_in << 127) | (half[i] >> 1);
            carry_in = next_carry;
        }

        let cmp_r = match cmp_u128(&rem[..rem_len], &half[..pow_len]) {
            n if n < 0 => core::cmp::Ordering::Less,
            0 => core::cmp::Ordering::Equal,
            _ => core::cmp::Ordering::Greater,
        };
        let q_is_odd = (quot[0] & 1) != 0;
        if rounding::should_bump(mode, cmp_r, q_is_odd, !neg) {
            let mut carry: u128 = 1;
            for limb in quot[..mag_len].iter_mut() {
                let (s, c) = limb.overflowing_add(carry);
                *limb = s;
                if !c { carry = 0; break; }
            }
            let _ = carry;
        }
    }

    for (i, slot) in mag_u128.iter_mut().enumerate() { *slot = quot[i]; }
}

/// Per-width Limb-axis matcher: does the `(width_bits, scale)`
/// cell run the u128-packed Newton kernel? Continuous width region per
/// Constitution rule 6 + Class I (never a per-scale carve-out).
///
/// The u128-packed kernel wins at the
/// 1536–4096 band and the 6144 width (`newton_vs_mg`
/// integrated bench, cores 22–23, sees u128 1.18–3.46× over MG and
/// 1.0–1.24× over u64 across s115–s953).
///
/// Wider widths (8192/12288/16384/32768) stay on MG entirely — the
/// AGM-widening / buffer-scratch / contradicting-integrated-bench reasons
/// in this module's notes. (Newton is now a [`crate::algos::support::rescale`]
/// kept-alt for every width, 9.18.2; this picks the packed apply if revived.)
#[inline]
const fn newton_u128_wins(width_bits: u32) -> bool {
    // The u128-packed apply wins across the CONTINUOUS 1536..=6144-bit band
    // (per the integrated bench above), not just the five tier FULL widths.
    // The frozen `matches!(1536|2048|3072|4096|6144)` was a sparse whitelist of
    // those exact tier widths — so any *magnitude-trimmed* width (the decimal
    // `mul` rescale, task-9.24 / the L6 trim, sizes the Newton on the product's
    // significant length, which lands on in-between 128-bit multiples like 2176
    // / 3200 / 4608) silently dropped onto the slow u64 apply. Key on the band +
    // 128-bit-multiple (even-u64, the packing precondition) instead. Perf-only:
    // both apply paths are bit-identical. (8192 / D1232 max-scale stays u64 — the
    // band's upper edge is unconfirmed there; a separate item.)
    width_bits.is_multiple_of(128) && width_bits >= 1536 && width_bits <= 8192
}

/// Full `n / 10^SCALE` with rounding for a `BigInt`-backed value.
///
/// Direct analogue of [`crate::algos::support::mg_divide::div_wide_pow10_chain`]
/// — same signature, same semantics, different inner algorithm.
pub(crate) fn div_wide_pow10_newton_with<W: crate::int::types::traits::BigInt>(
    n: W,
    scale: u32,
    mode: crate::support::rounding::RoundingMode,
    table: &NewtonReciprocal,
) -> W {
    // BigInt bridge is u128-limb; the arithmetic core operates on that
    // magnitude slice in place (shared with the `Int<N>`-only decimal
    // `mul` kernel, which builds its product directly in u128 scratch).
    //
    // Build-max u128 magnitude buffer for this genuinely-`N`-less
    // `W: BigInt` path: `max_u128_limb()` = `MAX_U128_LIMB` = `4·MAX_WORK_N`
    // u128 limbs — the widest work value any build forms (`Int<512>` =
    // 32768-bit = 256 u128 at xx-wide), tracking `MAX_WORK_N` so a narrower
    // build does NOT carry the widest tier's buffer. (The exact-per-`W`
    // `single_u128()` is wrong here: its blanket size only covers up to
    // `Int<256>`, so it would under-size the `Int<512>` work integer under a
    // non-`exact-scratch` build.) The kernel is sliced to `W::U128_LIMBS` so
    // narrower widths don't pay the wide cost.
    let mut mag_u128 = crate::int::types::compute_limbs::max_u128_limb();
    let limbs = <W as crate::int::types::traits::BigInt>::U128_LIMBS;
    let mag = &mut mag_u128[..limbs];
    let neg = n.mag_into_u128(mag);
    newton_pow10_mag_u128(mag, neg, mode, table);
    W::from_mag_sign_u128(mag, neg)
}

/// Width-agnostic Newton-reciprocal divide of a u128-limb magnitude slice
/// by `10^scale`, in place, with `mode`-aware rounding. `table` is the
/// pre-computed reciprocal for `(scale, width)`. Slice core extracted from
/// [`div_wide_pow10_newton_with`]; the only difference from the typed path
/// is the `BigInt` pack/unpack the wrapper does around this call.
///
/// The interior arithmetic runs on u64 limbs (`div_newton`), so this
/// transcodes the u128 magnitude to u64, divides, rounds, and transcodes
/// the quotient back into `mag` in place.
pub(crate) fn newton_pow10_mag_u128(
    mag_u128: &mut [u128],
    neg: bool,
    mode: crate::support::rounding::RoundingMode,
    table: &NewtonReciprocal,
) {
    use crate::support::rounding;

    // Transcode the u128 magnitude to the u64 limbs the kernel runs in.
    let mut mag = [0u64; MAX_MAG_U64];
    for (i, &v) in mag_u128.iter().enumerate() {
        mag[2 * i] = v as u64;
        mag[2 * i + 1] = (v >> 64) as u64;
    }
    let mag_len = mag_u128.len() * 2;

    let mut top = mag_len;
    while top > 0 && mag[top - 1] == 0 {
        top -= 1;
    }

    let n_slice = &mag[..top.max(1)];
    let mut quot = [0u64; MAX_MAG_U64];
    let mut rem = [0u64; MAX_MAG_U64 + 1];
    let rem_len = div_newton(n_slice, table, &mut quot[..mag_len], &mut rem);

    // Round per `mode`: compare remainder with pow_scale / 2.
    let rem_is_zero = rem[..rem_len].iter().all(|&x| x == 0);
    if !rem_is_zero {
        // half = pow_scale / 2 (pow_scale is even for scale >= 1)
        let pow_len = table.pow_len;
        let mut half = [0u64; MAX_POW_U64];
        half[..pow_len].copy_from_slice(&table.pow_scale[..pow_len]);
        // shift right by 1
        let mut i = pow_len;
        let mut carry_in: u64 = 0;
        while i > 0 {
            i -= 1;
            let next_carry = half[i] & 1;
            half[i] = (carry_in << 63) | (half[i] >> 1);
            carry_in = next_carry;
        }

        let cmp_r = match cmp(&rem[..rem_len], &half[..pow_len]) {
            n if n < 0 => core::cmp::Ordering::Less,
            0 => core::cmp::Ordering::Equal,
            _ => core::cmp::Ordering::Greater,
        };
        let q_is_odd = (quot[0] & 1) != 0;
        if rounding::should_bump(mode, cmp_r, q_is_odd, !neg) {
            let mut carry: u64 = 1;
            for limb in quot[..mag_len].iter_mut() {
                let (s, c) = limb.overflowing_add(carry);
                *limb = s;
                if !c {
                    carry = 0;
                    break;
                }
            }
            let _ = carry;
        }
    }

    // Re-pack the u64 quotient into the caller's u128 magnitude slice.
    for (i, slot) in mag_u128.iter_mut().enumerate() {
        let lo = quot[2 * i] as u128;
        let hi = quot[2 * i + 1] as u128;
        *slot = lo | (hi << 64);
    }
}

/// The baked-reciprocal Newton rescale execution: `mag /= 10^scale` via a
/// per-call [`NewtonReciprocal::precompute`] (a table LOOKUP of the §9.20
/// baked `⌊2^k/10^scale⌋` reciprocal — [`crate::consts::newton_recip`] — not a
/// Knuth divide) plus a reciprocal-multiply apply, in place on a u128-limb
/// magnitude slice (std-only; no_std forwards to the MG chain). `width_bits`
/// is the magnitude width in bits — the typed door
/// ([`crate::algos::support::rescale::dispatch_wide_pow10`]) passes the
/// SIGNIFICANT length here (task 9.24), so `precompute` sizes the reciprocal +
/// quotient to the real magnitude, not the full work buffer; `neg` is the
/// result sign for the rounding tie-break.
///
/// This is the `Newton` arm of the rescale matcher
/// ([`crate::algos::support::rescale`]) — **SELECTED** for the wide /
/// high-scale band (`select` routes `width_limbs 24..=132 × scale 200..=1850`
/// here). With the §9.20 baked table the precompute is a lookup, so the
/// one-pass O(width) apply beats the `⌈scale/38⌉`-pass MG chain 1.5–13× above
/// the crossover. The §9.2 no-state rule still holds — the per-call
/// `NewtonReciprocal` is a fixed stack struct populated from the immutable
/// baked const table, no cache.
#[inline]
pub(crate) fn newton_rescale_arm(
    mag: &mut [u128],
    scale: u32,
    neg: bool,
    mode: crate::support::rounding::RoundingMode,
    width_bits: u32,
) {
    #[cfg(feature = "std")]
    {
        // `width_limbs` in u64 limbs — the `precompute` unit. Recompute the
        // fixed-size, stack-only `NewtonReciprocal` for this `(width, scale)`
        // each call: value-independent but not const-evaluable (Knuth divide),
        // so a per-call stack recompute is the stateless, heap-free form.
        let width_limbs = (width_bits as usize) / 64;
        let table = NewtonReciprocal::precompute(scale, width_limbs);
        if newton_u128_wins(width_bits) {
            newton_pow10_mag_u128_packed(mag, neg, mode, &table);
        } else {
            newton_pow10_mag_u128(mag, neg, mode, &table);
        }
    }

    #[cfg(not(feature = "std"))]
    {
        // no_std has no Newton path (the per-call Knuth precompute is too
        // costly); forward to the MG chain — the kernel the matcher selects.
        let _ = width_bits;
        crate::algos::support::mg_divide::div_pow10_chain_mag_u128(mag, scale, neg, mode);
    }
}

#[cfg(test)]
mod tests {
    use super::*;
    // These two imports back the `newton_matches_mg_chain_*` validity-wall
    // tests, which exist only for the larger storage tiers (d307 … d1232).
    // Gate the imports to exactly that union of tiers — NOT the `wide`
    // umbrella, which is false under a single-tier build (`--features d307`)
    // and so dropped the imports while the d307 test still compiled.
    #[cfg(any(
        feature = "d307",
        feature = "d462",
        feature = "d616",
        feature = "d924",
        feature = "d1232"
    ))]
    use crate::algos::support::mg_divide::div_wide_pow10_chain;
    #[cfg(any(
        feature = "d307",
        feature = "d462",
        feature = "d616",
        feature = "d924",
        feature = "d1232"
    ))]
    use crate::int::types::Int;
    use crate::support::rounding::RoundingMode;

    /// The §9.20 baked reciprocal table must be BIT-IDENTICAL to what
    /// `precompute` would compute at runtime (`floor(2^(64*k) / 10^s)` via
    /// `div_rem_mag_slice`) — else the rescale result changes. Recompute the
    /// reciprocal independently (10^scale via a *10 chain, then the production
    /// divide) and compare to `consts::newton_recip_le`, across every baked
    /// width and a scale sweep incl. the band edges.
    ///
    /// Gated to `xx-wide` (the test must not exceed production scope): this
    /// sweep spans the FULL baked grid (`width_limbs` up to `NEWTON_RECIP_MAX_W`
    /// = 132, `scale` up to `NEWTON_RECIP_MAX_SCALE` = 1850), so the independent
    /// recompute divides numerators of up to `width_limbs + pow_len + 1 ≈ 233`
    /// u64 limbs — wider than ANY value a narrower build can form. Only the
    /// xx-wide build's slice-divide scratch (`MAX_SINGLE_LIMBS = 4·MAX_WORK_N + 2`
    /// = 258) holds them; under `wide,x-wide` it is 130 and the recompute would
    /// overrun it. The baked table is build-invariant `static` data, so
    /// verifying it where the widest entry is recomputable validates the table
    /// for every build.
    #[cfg(feature = "xx-wide")]
    #[test]
    fn baked_newton_recip_matches_runtime_divide() {
        use crate::int::algos::div::div_fixed::div_rem_mag_slice;
        for &width_limbs in &[16usize, 24, 32, 48, 64, 96, 128, 132] {
            for &scale in &[1u32, 38, 39, 77, 200, 461, 615, 924, 1231, 1850] {
                let pow_len = (scale as usize / 19 + 3).max(1);
                let mut pow = [0u64; MAX_POW_U64];
                pow[0] = 1;
                for _ in 0..scale {
                    let mut carry = 0u64;
                    for limb in pow[..pow_len].iter_mut() {
                        let p = (*limb as u128) * 10 + carry as u128;
                        *limb = p as u64;
                        carry = (p >> 64) as u64;
                    }
                    assert_eq!(carry, 0, "10^{scale} overflowed pow_len");
                }
                let k_raw = width_limbs + pow_len;
                let k = if k_raw % 2 == 0 { k_raw } else { k_raw + 1 };
                let mut num = [0u64; MAX_R_U64];
                num[k] = 1;
                let mut r = [0u64; MAX_R_U64];
                let mut rem = [0u64; MAX_POW_U64];
                div_rem_mag_slice(&num[..k + 1], &pow[..pow_len], &mut r[..k + 1], &mut rem[..pow_len]);
                let baked = crate::consts::newton_recip_le(scale, width_limbs)
                    .expect("baked reciprocal in range");
                assert_eq!(
                    baked,
                    &r[..k + 1],
                    "baked != runtime divide: width_limbs={width_limbs} scale={scale}"
                );
            }
        }
    }

    // Exercises `Int<16>` (D307 storage); the divide scratch is sized by the
    // build's `MAX_WORK_N`, so this only runs where that tier is enabled.
    #[cfg(feature = "d307")]
    #[test]
    fn newton_matches_mg_chain_d307_s150() {
        let scale = 150u32;
        let width_limbs = 16; // Int<16> = 8 u128 = 16 u64 limbs
        let table = NewtonReciprocal::precompute(scale, width_limbs);

        let mut limbs = [0u128; 64];
        limbs[6] = 1u128 << 32;
        limbs[0] = 42;
        let n = <Int<16> as crate::int::types::traits::BigInt>::from_mag_sign_u128(&limbs, false);

        let got = div_wide_pow10_newton_with(n, scale, RoundingMode::HalfToEven, &table);
        let want = div_wide_pow10_chain::<Int<16>>(n, scale, RoundingMode::HalfToEven);
        assert_eq!(got, want, "Newton differs from MG chain at D307 s=150");
    }

    // Exercises `Int<24>` (D462 storage AND D230 Work) at the benchmarked anchor
    // scales — bit-identical agreement is the validity wall for the
    // 1536-bit Newton kept-alt (the `rescale` matcher's `Newton` arm).
    #[cfg(feature = "d462")]
    #[test]
    fn newton_matches_mg_chain_d462_s202() {
        let scale = 202u32;
        let width_limbs = 24; // Int<24> = 12 u128 = 24 u64 limbs
        let table = NewtonReciprocal::precompute(scale, width_limbs);

        let mut limbs = [0u128; 64];
        limbs[10] = 1u128 << 24;
        limbs[2] = 0xfeedfacecafef00d_u128;
        let n = <Int<24> as crate::int::types::traits::BigInt>::from_mag_sign_u128(&limbs, false);

        let got = div_wide_pow10_newton_with(n, scale, RoundingMode::HalfToEven, &table);
        let want = div_wide_pow10_chain::<Int<24>>(n, scale, RoundingMode::HalfToEven);
        assert_eq!(got, want, "Newton differs from MG chain at Int<24> s=202");
    }

    #[cfg(feature = "d462")]
    #[test]
    fn newton_matches_mg_chain_d462_s259() {
        let scale = 259u32;
        let width_limbs = 24;
        let table = NewtonReciprocal::precompute(scale, width_limbs);

        let mut limbs = [0u128; 64];
        limbs[10] = 1u128 << 8;
        limbs[1] = 0xdeadbeef_cafef00d_u128;
        let n = <Int<24> as crate::int::types::traits::BigInt>::from_mag_sign_u128(&limbs, false);

        let got = div_wide_pow10_newton_with(n, scale, RoundingMode::HalfToEven, &table);
        let want = div_wide_pow10_chain::<Int<24>>(n, scale, RoundingMode::HalfToEven);
        assert_eq!(got, want, "Newton differs from MG chain at Int<24> s=259");
    }

    // Exercises `Int<32>` (D616 storage) — runs only where that tier is on.
    #[cfg(feature = "d616")]
    #[test]
    fn newton_matches_mg_chain_d616_s308() {
        let scale = 308u32;
        let width_limbs = 32; // Int<32> = 16 u128 = 32 u64 limbs
        let table = NewtonReciprocal::precompute(scale, width_limbs);

        let mut limbs = [0u128; 64];
        limbs[14] = 1u128 << 16;
        limbs[3] = 0xdeadbeef;
        let n = <Int<32> as crate::int::types::traits::BigInt>::from_mag_sign_u128(&limbs, false);

        let got = div_wide_pow10_newton_with(n, scale, RoundingMode::HalfToEven, &table);
        let want = div_wide_pow10_chain::<Int<32>>(n, scale, RoundingMode::HalfToEven);
        assert_eq!(got, want, "Newton differs from MG chain at D616 s=308");
    }

    // Exercises `Int<64>` (D1232 storage) — runs only where that tier is on.
    #[cfg(feature = "d1232")]
    #[test]
    fn newton_matches_mg_chain_d1232_s615() {
        let scale = 615u32;
        let width_limbs = 64; // Int<64> = 32 u128 = 64 u64 limbs
        let table = NewtonReciprocal::precompute(scale, width_limbs);

        let mut limbs = [0u128; 64];
        limbs[30] = 1u128 << 8;
        limbs[5] = 0xcafef00d;
        let n = <Int<64> as crate::int::types::traits::BigInt>::from_mag_sign_u128(&limbs, false);

        let got = div_wide_pow10_newton_with(n, scale, RoundingMode::HalfToEven, &table);
        let want = div_wide_pow10_chain::<Int<64>>(n, scale, RoundingMode::HalfToEven);
        assert_eq!(got, want, "Newton differs from MG chain at D1232 s=615");
    }

    // -- u64-vs-u128 Newton bit-identity (validity wall) ---------------
    //
    // The u128-packed kernel MUST produce limb-identical output to the
    // u64 kernel for every supported width x representative scale, in
    // every rounding mode. Any divergence is a kernel bug.

    fn assert_u64_u128_match(
        scale: u32,
        width_limbs: usize,
        mag_limbs: usize,
        top_limb_idx: usize,
        top_limb_val: u128,
        low_perturbation: (usize, u128),
    ) {
        let table = NewtonReciprocal::precompute(scale, width_limbs);

        let modes = [
            RoundingMode::HalfToEven,
            RoundingMode::HalfAwayFromZero,
            RoundingMode::HalfTowardZero,
            RoundingMode::Trunc,
            RoundingMode::Floor,
            RoundingMode::Ceiling,
        ];

        for mode in modes {
            // Buffer sized to 128 u128 limbs to fit Int<192>=96 u128
            // (the widest mag the audit-extended widths exercise).
            let mut mag_a = [0u128; 128];
            mag_a[top_limb_idx] = top_limb_val;
            mag_a[low_perturbation.0] = low_perturbation.1;
            let mut mag_b = mag_a;

            super::newton_pow10_mag_u128(&mut mag_a[..mag_limbs], false, mode, &table);
            super::newton_pow10_mag_u128_packed(&mut mag_b[..mag_limbs], false, mode, &table);
            assert_eq!(
                mag_a, mag_b,
                "u64 != u128 Newton at scale={scale} width={width_limbs} mode={mode:?}"
            );

            let mut mag_a = [0u128; 128];
            mag_a[top_limb_idx] = top_limb_val;
            mag_a[low_perturbation.0] = low_perturbation.1;
            let mut mag_b = mag_a;
            super::newton_pow10_mag_u128(&mut mag_a[..mag_limbs], true, mode, &table);
            super::newton_pow10_mag_u128_packed(&mut mag_b[..mag_limbs], true, mode, &table);
            assert_eq!(
                mag_a, mag_b,
                "u64 != u128 Newton (neg) at scale={scale} width={width_limbs} mode={mode:?}"
            );
        }
    }

    #[cfg(feature = "d307")]
    #[test]
    fn newton_u64_eq_u128_d307_s150() {
        assert_u64_u128_match(150, 16, 8, 6, 1u128 << 32, (1, 0xdeadbeef_cafef00d_u128));
    }

    #[cfg(feature = "d307")]
    #[test]
    fn newton_u64_eq_u128_d307_s307() {
        assert_u64_u128_match(307, 16, 8, 7, 0x1234_5678_9abc_def0u128, (0, 1));
    }

    // Int<24> = 1536-bit. The PRODUCTION row: D462 storage, D230 Work.
    #[cfg(feature = "d462")]
    #[test]
    fn newton_u64_eq_u128_b1536_s200() {
        assert_u64_u128_match(200, 24, 12, 10, 1u128 << 24, (2, 0xfeedfacecafef00d_u128));
    }

    #[cfg(feature = "d462")]
    #[test]
    fn newton_u64_eq_u128_b1536_s202() {
        assert_u64_u128_match(202, 24, 12, 10, 1u128 << 24, (2, 0xfeedfacecafef00d_u128));
    }

    #[cfg(feature = "d462")]
    #[test]
    fn newton_u64_eq_u128_b1536_s259() {
        assert_u64_u128_match(259, 24, 12, 10, 1u128 << 8, (1, 0xdeadbeef_cafef00d_u128));
    }

    #[cfg(feature = "d462")]
    #[test]
    fn newton_u64_eq_u128_b1536_s461() {
        assert_u64_u128_match(461, 24, 12, 11, 0x1u128, (3, 0xfacefacef00d_u128));
    }

    #[cfg(feature = "d616")]
    #[test]
    fn newton_u64_eq_u128_d616_s308() {
        assert_u64_u128_match(308, 32, 16, 14, 1u128 << 16, (3, 0xdeadbeef));
    }

    #[cfg(feature = "d616")]
    #[test]
    fn newton_u64_eq_u128_d616_s616() {
        assert_u64_u128_match(616, 32, 16, 15, 1u128 << 8, (4, 0xfeedface));
    }

    #[cfg(feature = "d924")]
    #[test]
    fn newton_u64_eq_u128_d924_s460() {
        assert_u64_u128_match(460, 48, 24, 22, 1u128 << 8, (5, 0xcafef00d));
    }

    #[cfg(feature = "d924")]
    #[test]
    fn newton_u64_eq_u128_d924_s924() {
        assert_u64_u128_match(924, 48, 24, 23, 1u128, (6, 0xfeedfacef00d));
    }

    #[cfg(feature = "d1232")]
    #[test]
    fn newton_u64_eq_u128_d1232_s615() {
        assert_u64_u128_match(615, 64, 32, 30, 1u128 << 8, (5, 0xcafef00d));
    }

    #[cfg(feature = "d1232")]
    #[test]
    fn newton_u64_eq_u128_d1232_s1231() {
        assert_u64_u128_match(1231, 64, 32, 31, 1u128, (7, 0xdeadbeef_feedface_u128));
    }

    // ── Wider-width validity wall ──────────────────
    //
    // Int<96> / Int<128> / Int<192> exercised at the benchmarked anchor scales
    // and the maxima. Bit-identical agreement with the Knuth-routed
    // `div_rem_mag_slice` path is the validity wall for the wide Newton
    // kept-alt cells (6144 / 8192 / 12288; the `rescale` `Newton` arm).

    #[cfg(any(feature = "d924", feature = "xx-wide"))]
    #[test]
    fn newton_matches_mg_chain_b6144_s200() {
        let scale = 200u32;
        let width_limbs = 96; // Int<96> = 48 u128 = 96 u64
        let table = NewtonReciprocal::precompute(scale, width_limbs);
        let mut limbs = [0u128; 64];
        limbs[40] = 1u128 << 24;
        limbs[3] = 0xfeedfacecafef00d_u128;
        let n = <Int<96> as crate::int::types::traits::BigInt>::from_mag_sign_u128(&limbs, false);
        let got = div_wide_pow10_newton_with(n, scale, RoundingMode::HalfToEven, &table);
        let want = div_wide_pow10_chain::<Int<96>>(n, scale, RoundingMode::HalfToEven);
        assert_eq!(got, want, "Newton differs from MG chain at Int<96> s=200");
    }

    #[cfg(any(feature = "d924", feature = "xx-wide"))]
    #[test]
    fn newton_matches_mg_chain_b6144_s953() {
        let scale = 953u32;
        let width_limbs = 96;
        let table = NewtonReciprocal::precompute(scale, width_limbs);
        let mut limbs = [0u128; 64];
        limbs[46] = 1u128 << 8;
        limbs[1] = 0xdeadbeef_cafef00d_u128;
        let n = <Int<96> as crate::int::types::traits::BigInt>::from_mag_sign_u128(&limbs, false);
        let got = div_wide_pow10_newton_with(n, scale, RoundingMode::HalfToEven, &table);
        let want = div_wide_pow10_chain::<Int<96>>(n, scale, RoundingMode::HalfToEven);
        assert_eq!(got, want, "Newton differs from MG chain at Int<96> s=953");
    }

    // u64 vs u128 bit-identity at the 6144 width covered by
    // `newton_u128_wins` — production-shape cells across the D924
    // SCALE band AND the AGM-widened scales the strict_agm transcendentals
    // exercise.

    #[cfg(any(feature = "d924", feature = "xx-wide"))]
    #[test]
    fn newton_u64_eq_u128_b6144_s200() {
        assert_u64_u128_match(200, 96, 48, 40, 1u128 << 24, (3, 0xfeedfacecafef00d_u128));
    }

    #[cfg(any(feature = "d924", feature = "xx-wide"))]
    #[test]
    fn newton_u64_eq_u128_b6144_s953() {
        assert_u64_u128_match(953, 96, 48, 46, 1u128 << 8, (1, 0xdeadbeef_cafef00d_u128));
    }

    // AGM-band cells — D924 strict_agm runs at `w_prime ≤ 1850`.
    #[cfg(any(feature = "d924", feature = "xx-wide"))]
    #[test]
    fn newton_u64_eq_u128_b6144_s1234() {
        assert_u64_u128_match(1234, 96, 48, 46, 1u128 << 16, (2, 0xcafef00dbeef_u128));
    }

    #[cfg(any(feature = "d924", feature = "xx-wide"))]
    #[test]
    fn newton_u64_eq_u128_b6144_s1850() {
        assert_u64_u128_match(1850, 96, 48, 47, 1u128, (5, 0xfacefacef00d_u128));
    }
}