pixon 0.8.0

SIMD-dispatched color conversion for the FFmpeg AVPixelFormat space: one-call decode of validated source frames to RGB / RGBA / Luma / HSV, fused resampling while converting, colorimetry-honest matrices and chroma siting, 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
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
1499
1500
1501
1502
1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
1563
1564
1565
1566
1567
1568
1569
1570
1571
1572
1573
1574
1575
1576
1577
1578
1579
1580
1581
1582
1583
1584
1585
1586
1587
1588
1589
1590
1591
1592
1593
1594
1595
1596
1597
1598
1599
1600
1601
1602
1603
1604
1605
1606
1607
1608
1609
1610
1611
1612
1613
1614
1615
1616
1617
1618
1619
1620
1621
1622
1623
1624
1625
1626
1627
1628
1629
1630
1631
1632
1633
1634
1635
1636
1637
1638
1639
1640
1641
1642
1643
1644
1645
1646
1647
1648
1649
1650
1651
1652
1653
1654
1655
1656
1657
1658
1659
1660
1661
1662
1663
1664
//! RFC #238 S3b/S4-C — chroma-siting-aware 4:2:0 **resample** for the
//! semi-planar `Nv12` (`U V U V …`) and `Nv21` (`V U V U …`).
//!
//! The 4:2:0 resample twin of the identity-decode `chroma_siting_nv` and the
//! semi-planar sibling of the planar `chroma_siting_420_resample` (`Yuv420p`).
//! S3b routes the HORIZONTAL centered siting (`Center` / `Top` / `Bottom`,
//! [`chroma_420_center_sited_h`](super::super::chroma_420_center_sited_h))
//! through the resample; S4-C additionally folds the VERTICAL `v = 1` triangle
//! for `Bottom`
//! ([`chroma_420_bottom_sited_v`](super::super::chroma_420_bottom_sited_v)) —
//! its even output row box-blends the previous chroma row — while `Center` /
//! `Top` keep the vertical pairing co-sited (`v_phase = 0`). The interleaved
//! chroma is de-interleaved to the SAME half-width U / V planes a `Yuv420p`
//! frame holds, so every `Nv12` / `Nv21` resample (centered AND bottom) is
//! **bit-identical** to the `Yuv420p` resample of those planes — the strongest
//! catch for a U/V swap in the de-interleave (`swap_uv`).
//!
//! Oracles (independent of the production kernel):
//!  - the **native** tier is pinned to the EXACT code-domain box-average of the
//!    UNROUNDED triangle-reconstructed chroma — a SINGLE rounding, the
//!    user-approved more-correct form — computed in the chroma CODE domain
//!    (never RGB, which would prove the wrong averaging domain), on EVEN source
//!    heights so the luma-domain vertical pairing equals the co-sited box over
//!    the `sh / 2` chroma rows;
//!  - the **encoded row-stage** tier is pinned to the RGB-domain
//!    reconstruct-to-u8-then-bin;
//!  - both tiers are ALSO cross-checked bit-identical against the `Yuv420p`
//!    resample of the de-interleaved planes.

use super::*;
use crate::{
  ChromaLocation, PixelSink,
  resample::{AreaResampler, FilteredResampler, ResampleError, Triangle},
};

const M: KernelMatrix = KernelMatrix::Bt601;
const FR: bool = true;

/// Round `a / d` half-up (ties toward `+∞`) — the production
/// `round_div_half_up`, replicated here so the oracle is independent.
fn rdhu(a: u64, d: u64) -> u64 {
  let q = a / d;
  let r = a % d;
  q + u64::from(r >= d - d / 2)
}

/// Exact box-overlap area weights for `src -> out`, mirroring
/// `resample::AxisSpans::area`. Returns per output `(first source cell, overlaps)`.
fn area_weights(src: usize, out: usize) -> Vec<(usize, Vec<u64>)> {
  let (src64, out64) = (src as u64, out as u64);
  (0..out)
    .map(|o| {
      let lo = o as u64 * src64;
      let hi = lo + src64;
      let start = (lo / out64) as usize;
      let mut w = Vec::new();
      let mut i = start as u64;
      loop {
        let clo = i * out64;
        if clo >= hi {
          break;
        }
        let chi = clo + out64;
        let ov = chi.min(hi) - clo.max(lo);
        if ov == 0 {
          break;
        }
        w.push(ov);
        if chi >= hi {
          break;
        }
        i += 1;
      }
      (start, w)
    })
    .collect()
}

/// Co-sited box-average of a full-resolution `sw x sh` u8 plane to `ow x oh`
/// (round-half-up) — the reference for a phase-free plane (luma).
fn bin_cosited(plane: &[u8], sw: usize, sh: usize, ow: usize, oh: usize) -> Vec<u8> {
  let hw = area_weights(sw, ow);
  let vw = area_weights(sh, oh);
  let denom = (sw * sh) as u64;
  let mut out = vec![0u8; ow * oh];
  for (oy, (vs, vwin)) in vw.iter().enumerate() {
    for (ox, (hs, hwin)) in hw.iter().enumerate() {
      let mut s = 0u64;
      for (dy, &vwt) in vwin.iter().enumerate() {
        let mut hsum = 0u64;
        for (dx, &hwt) in hwin.iter().enumerate() {
          hsum += hwt * u64::from(plane[(vs + dy) * sw + hs + dx]);
        }
        s += vwt * hsum;
      }
      out[oy * ow + ox] = rdhu(s, denom) as u8;
    }
  }
  out
}

/// The EXACT centered chroma oracle for the native tier: reconstruct the
/// `cw x ch` chroma to full width with the #302 `1/4`–`3/4` triangle kept
/// UNROUNDED (scaled ×4 to stay integral: `r ∈ {1, 3, 4}`), then box-average to
/// `ow x oh` — HORIZONTAL over `2·cw`, VERTICAL over the `ch` chroma rows
/// (co-sited: 4:2:0 vertical stays a box pairing) — with a SINGLE round-half-up
/// over `4·(2·cw)·ch`. The code-domain twin the folded
/// [`ResamplePlan::area_chroma_420`] weights realize (for EVEN `sh`).
fn bin_chroma_centered(c: &[u8], cw: usize, ch: usize, ow: usize, oh: usize) -> Vec<u8> {
  let full = 2 * cw;
  let mut r4 = vec![0u32; full * ch];
  for r in 0..ch {
    let row = &c[r * cw..r * cw + cw];
    for j in 0..cw {
      let l = u32::from(row[j.saturating_sub(1)]);
      let m = u32::from(row[j]);
      let rt = u32::from(row[if j + 1 < cw { j + 1 } else { j }]);
      r4[r * full + 2 * j] = l + 3 * m;
      r4[r * full + 2 * j + 1] = 3 * m + rt;
    }
  }
  let hw = area_weights(full, ow);
  let vw = area_weights(ch, oh);
  let denom = (4 * full * ch) as u64;
  let mut out = vec![0u8; ow * oh];
  for (oy, (vs, vwin)) in vw.iter().enumerate() {
    for (ox, (hs, hwin)) in hw.iter().enumerate() {
      let mut s = 0u64;
      for (dy, &vwt) in vwin.iter().enumerate() {
        let mut hsum = 0u64;
        for (dx, &hwt) in hwin.iter().enumerate() {
          hsum += hwt * u64::from(r4[(vs + dy) * full + hs + dx]);
        }
        s += vwt * hsum;
      }
      out[oy * ow + ox] = rdhu(s, denom) as u8;
    }
  }
  out
}

/// Independent #302 centered horizontal upsample (`1/4`–`3/4`, edge clamp,
/// round-half-up to u8) — the RGB-domain oracle's reconstruction step.
fn recon_full_row(c: &[u8], cw: usize) -> Vec<u8> {
  let mut out = vec![0u8; 2 * cw];
  for j in 0..cw {
    let l = u32::from(c[j.saturating_sub(1)]);
    let m = u32::from(c[j]);
    let r = u32::from(c[if j + 1 < cw { j + 1 } else { j }]);
    out[2 * j] = ((l + 3 * m + 2) >> 2) as u8;
    out[2 * j + 1] = ((3 * m + r + 2) >> 2) as u8;
  }
  out
}

/// A `Yuv420p`/`Nv12`/`Nv21` fixture (`cw = sw / 2`, `ch = sh / 2`) with a strong
/// HORIZONTAL chroma ramp (so the centered triangle, which pulls neighbours,
/// genuinely differs from the co-sited nearest decode) plus a per-row tilt (a
/// vertical mistake would show). `sw` / `sh` must be even.
fn ramp(sw: usize, sh: usize) -> (Vec<u8>, Vec<u8>, Vec<u8>) {
  let cw = sw / 2;
  let ch = sh / 2;
  let mut y = vec![0u8; sw * sh];
  let mut u = vec![0u8; cw * ch];
  let mut v = vec![0u8; cw * ch];
  for (i, p) in y.iter_mut().enumerate() {
    *p = 40 + ((i as u32 * 3) % 160) as u8;
  }
  for r in 0..ch {
    for c in 0..cw {
      u[r * cw + c] = (30 + c * 44 + r * 4).min(240) as u8;
      v[r * cw + c] = (230u32.saturating_sub((c * 44 + r * 4) as u32)).max(16) as u8;
    }
  }
  (y, u, v)
}

/// A flat-chroma fixture: the centered phase is a no-op on constant chroma, so
/// centered must equal co-sited. Luma still varies.
fn flat_chroma(sw: usize, sh: usize) -> (Vec<u8>, Vec<u8>, Vec<u8>) {
  let cw = sw / 2;
  let ch = sh / 2;
  let mut y = vec![0u8; sw * sh];
  for (i, p) in y.iter_mut().enumerate() {
    *p = 40 + ((i as u32 * 7) % 170) as u8;
  }
  (y, vec![110u8; cw * ch], vec![140u8; cw * ch])
}

/// Interleave the half-width planar U / V into a 4:2:0 semi-planar chroma plane
/// (`2·cw` bytes per row, `ch` rows): `swap = false` packs Nv12 (`U` at the even
/// byte), `true` packs Nv21 (`V` at the even byte).
fn interleave(u: &[u8], v: &[u8], cw: usize, ch: usize, swap: bool) -> Vec<u8> {
  let full = 2 * cw;
  let mut uv = vec![0u8; full * ch];
  for r in 0..ch {
    for c in 0..cw {
      let (even, odd) = if swap {
        (v[r * cw + c], u[r * cw + c])
      } else {
        (u[r * cw + c], v[r * cw + c])
      };
      uv[r * full + 2 * c] = even;
      uv[r * full + 2 * c + 1] = odd;
    }
  }
  uv
}

type Outs = (
  Vec<u8>,
  Vec<u8>,
  (Vec<u8>, Vec<u8>, Vec<u8>),
  Vec<u8>,
  Vec<u16>,
);

/// The `Yuv420p` twin cross-check: the SAME planar U / V driven through a planar
/// 4:2:0 resample. A centered `Nv12` / `Nv21` decode must equal this
/// byte-for-byte (the de-interleave reconstructs the SAME planes).
#[allow(clippy::too_many_arguments)]
fn run_yuv420p(
  y: &[u8],
  u: &[u8],
  v: &[u8],
  sw: usize,
  sh: usize,
  ow: usize,
  oh: usize,
  loc: ChromaLocation,
  native: bool,
  simd: bool,
) -> Outs {
  let cw = sw / 2;
  let mut rgb = vec![0u8; ow * oh * 3];
  let mut rgba = vec![0u8; ow * oh * 4];
  let (mut hh, mut ss, mut vv) = (vec![0u8; ow * oh], vec![0u8; ow * oh], vec![0u8; ow * oh]);
  let mut luma = vec![0u8; ow * oh];
  let mut luma_u16 = vec![0u16; ow * oh];
  {
    let mut sink =
      MixedSinker::<Yuv420p, AreaResampler>::with_resampler(sw, sh, AreaResampler::to(ow, oh))
        .unwrap()
        .with_native(native)
        .with_chroma_location(loc.clone())
        .with_simd(simd)
        .with_rgb(&mut rgb)
        .unwrap()
        .with_rgba(&mut rgba)
        .unwrap()
        .with_hsv(&mut hh, &mut ss, &mut vv)
        .unwrap()
        .with_luma(&mut luma)
        .unwrap()
        .with_luma_u16(&mut luma_u16)
        .unwrap();
    let f = Yuv420pFrame::new(
      y, u, v, sw as u32, sh as u32, sw as u32, cw as u32, cw as u32,
    );
    yuv420p_to(&f, FR, sink.set_kernel_matrix(M)).unwrap();
  }
  (rgb, rgba, (hh, ss, vv), luma, luma_u16)
}

/// The centered NATIVE oracle: bin Y co-sited and U / V through the exact
/// centered chroma oracle to `ow x oh`, then convert ONCE at output width via an
/// identity `Yuv444p` sink — the exact ground truth the native tier reproduces
/// (EVEN `sh` only).
#[allow(clippy::too_many_arguments)]
fn native_oracle(
  y: &[u8],
  u: &[u8],
  v: &[u8],
  sw: usize,
  sh: usize,
  ow: usize,
  oh: usize,
  simd: bool,
) -> Outs {
  let cw = sw / 2;
  let ch = sh / 2;
  let yb = bin_cosited(y, sw, sh, ow, oh);
  let ub = bin_chroma_centered(u, cw, ch, ow, oh);
  let vb = bin_chroma_centered(v, cw, ch, ow, oh);
  let mut rgb = vec![0u8; ow * oh * 3];
  let mut rgba = vec![0u8; ow * oh * 4];
  let (mut hh, mut ss, mut vv) = (vec![0u8; ow * oh], vec![0u8; ow * oh], vec![0u8; ow * oh]);
  let mut luma = vec![0u8; ow * oh];
  let mut luma_u16 = vec![0u16; ow * oh];
  {
    let mut sink = MixedSinker::<Yuv444p>::new(ow, oh)
      .with_simd(simd)
      .with_rgb(&mut rgb)
      .unwrap()
      .with_rgba(&mut rgba)
      .unwrap()
      .with_hsv(&mut hh, &mut ss, &mut vv)
      .unwrap()
      .with_luma(&mut luma)
      .unwrap()
      .with_luma_u16(&mut luma_u16)
      .unwrap();
    let f = Yuv444pFrame::new(
      &yb, &ub, &vb, ow as u32, oh as u32, ow as u32, ow as u32, ow as u32,
    );
    yuv444p_to(&f, FR, sink.set_kernel_matrix(M)).unwrap();
  }
  (rgb, rgba, (hh, ss, vv), luma, luma_u16)
}

/// The centered ENCODED row-stage oracle: reconstruct U / V to full width with
/// the #302 kernel (u8), replicate each chroma row across its two luma rows (the
/// co-sited vertical pairing), then run that full-resolution `Yuv444p` frame
/// through a `with_native(false)` RGB-domain resample — convert-each-row-then-bin.
#[allow(clippy::too_many_arguments)]
fn encoded_oracle_rgb(
  y: &[u8],
  u: &[u8],
  v: &[u8],
  sw: usize,
  sh: usize,
  ow: usize,
  oh: usize,
  simd: bool,
) -> Vec<u8> {
  let cw = sw / 2;
  let mut uf = vec![0u8; sw * sh];
  let mut vf = vec![0u8; sw * sh];
  for r in 0..sh {
    let cr = r / 2;
    uf[r * sw..r * sw + sw].copy_from_slice(&recon_full_row(&u[cr * cw..cr * cw + cw], cw));
    vf[r * sw..r * sw + sw].copy_from_slice(&recon_full_row(&v[cr * cw..cr * cw + cw], cw));
  }
  let mut rgb = vec![0u8; ow * oh * 3];
  {
    let mut sink =
      MixedSinker::<Yuv444p, AreaResampler>::with_resampler(sw, sh, AreaResampler::to(ow, oh))
        .unwrap()
        .with_native(false)
        .with_simd(simd)
        .with_rgb(&mut rgb)
        .unwrap();
    let f = Yuv444pFrame::new(
      y, &uf, &vf, sw as u32, sh as u32, sw as u32, sw as u32, sw as u32,
    );
    yuv444p_to(&f, FR, sink.set_kernel_matrix(M)).unwrap();
  }
  rgb
}

// ---- RFC #238 S4-C bottom-sited (v = 1) oracles ----------------------------
//
// The vertical `Bottom` fold reconstructs the same DE-INTERLEAVED planar U / V a
// `Yuv420p` frame holds, so every planar-derived Bottom oracle below is the
// byte-for-byte ground truth an `Nv12` / `Nv21` Bottom resample must match.

/// The EXACT bottom-sited (`v = 1`) chroma oracle for the native tier: the
/// centered horizontal `1/4`–`3/4` triangle (×4) composed with the vertical
/// `v = 1` triangle (×2 — even luma row `2i` box-blends chroma rows `{i - 1, i}`
/// with weights `{1, 1}`, odd row `2i + 1` takes chroma row `i`, top-edge clamp),
/// the combined ×8 UNROUNDED reconstruction box-averaged to `ow x oh` with a
/// SINGLE round-half-up — the code-domain twin the folded
/// [`ResamplePlan::area_chroma_420`] realizes for `v = 1` (EVEN `sh` only).
fn bin_chroma_bottom(c: &[u8], cw: usize, ch: usize, ow: usize, oh: usize) -> Vec<u8> {
  let full = 2 * cw;
  let sh = 2 * ch;
  let mut r8 = vec![0u64; full * sh];
  for r in 0..sh {
    let cr = r / 2;
    let prev = cr.saturating_sub(1);
    let vrow: Vec<u32> = (0..cw)
      .map(|j| {
        if r & 1 == 0 {
          u32::from(c[prev * cw + j]) + u32::from(c[cr * cw + j]) // even: {1, 1}
        } else {
          2 * u32::from(c[cr * cw + j]) // odd: {2}
        }
      })
      .collect();
    for j in 0..cw {
      let l = vrow[j.saturating_sub(1)];
      let m = vrow[j];
      let rt = vrow[if j + 1 < cw { j + 1 } else { j }];
      r8[r * full + 2 * j] = u64::from(l + 3 * m);
      r8[r * full + 2 * j + 1] = u64::from(3 * m + rt);
    }
  }
  let hw = area_weights(full, ow);
  let vw = area_weights(sh, oh);
  let denom = (8 * full * sh) as u64;
  let mut out = vec![0u8; ow * oh];
  for (oy, (vs, vwin)) in vw.iter().enumerate() {
    for (ox, (hs, hwin)) in hw.iter().enumerate() {
      let mut s = 0u64;
      for (dy, &vwt) in vwin.iter().enumerate() {
        let mut hsum = 0u64;
        for (dx, &hwt) in hwin.iter().enumerate() {
          hsum += hwt * r8[(vs + dy) * full + hs + dx];
        }
        s += vwt * hsum;
      }
      out[oy * ow + ox] = rdhu(s, denom) as u8;
    }
  }
  out
}

/// The bottom-sited NATIVE oracle: bin Y co-sited and U / V through the exact
/// bottom V-fold chroma oracle to `ow x oh`, then convert ONCE at output width
/// via an identity `Yuv444p` sink (EVEN `sh` only).
#[allow(clippy::too_many_arguments)]
fn bottom_native_oracle(
  y: &[u8],
  u: &[u8],
  v: &[u8],
  sw: usize,
  sh: usize,
  ow: usize,
  oh: usize,
  simd: bool,
) -> Outs {
  let cw = sw / 2;
  let ch = sh / 2;
  let yb = bin_cosited(y, sw, sh, ow, oh);
  let ub = bin_chroma_bottom(u, cw, ch, ow, oh);
  let vb = bin_chroma_bottom(v, cw, ch, ow, oh);
  let mut rgb = vec![0u8; ow * oh * 3];
  let mut rgba = vec![0u8; ow * oh * 4];
  let (mut hh, mut ss, mut vv) = (vec![0u8; ow * oh], vec![0u8; ow * oh], vec![0u8; ow * oh]);
  let mut luma = vec![0u8; ow * oh];
  let mut luma_u16 = vec![0u16; ow * oh];
  {
    let mut sink = MixedSinker::<Yuv444p>::new(ow, oh)
      .with_simd(simd)
      .with_rgb(&mut rgb)
      .unwrap()
      .with_rgba(&mut rgba)
      .unwrap()
      .with_hsv(&mut hh, &mut ss, &mut vv)
      .unwrap()
      .with_luma(&mut luma)
      .unwrap()
      .with_luma_u16(&mut luma_u16)
      .unwrap();
    let f = Yuv444pFrame::new(
      &yb, &ub, &vb, ow as u32, oh as u32, ow as u32, ow as u32, ow as u32,
    );
    yuv444p_to(&f, FR, sink.set_kernel_matrix(M)).unwrap();
  }
  (rgb, rgba, (hh, ss, vv), luma, luma_u16)
}

/// Reconstruct planar U / V to full resolution (`sw x sh`) for the bottom-sited
/// (`v = 1`) decode: per luma row the even rows vertically box-blend chroma rows
/// `i - 1` (clamped) and `i` (round-half-up), the odd rows take chroma row `i`,
/// each then horizontally upsampled with the #302 centered `1/4`–`3/4` kernel.
/// The shared reconstruction step for every reconstruct-then-bin bottom oracle;
/// NV de-interleaves to these SAME planes, so the oracle is format-agnostic.
fn recon_full_bottom(u: &[u8], v: &[u8], sw: usize, sh: usize) -> (Vec<u8>, Vec<u8>) {
  let cw = sw / 2;
  let vblend = |plane: &[u8], cr: usize, prev: usize| -> Vec<u8> {
    (0..cw)
      .map(|c| {
        let a = u32::from(plane[prev * cw + c]);
        let b = u32::from(plane[cr * cw + c]);
        ((a + b + 1) >> 1) as u8
      })
      .collect::<Vec<u8>>()
  };
  let mut uf = vec![0u8; sw * sh];
  let mut vf = vec![0u8; sw * sh];
  for r in 0..sh {
    let cr = r / 2;
    let (uh, vh) = if r & 1 == 0 {
      let prev = cr.saturating_sub(1);
      (vblend(u, cr, prev), vblend(v, cr, prev))
    } else {
      (
        u[cr * cw..cr * cw + cw].to_vec(),
        v[cr * cw..cr * cw + cw].to_vec(),
      )
    };
    uf[r * sw..r * sw + sw].copy_from_slice(&recon_full_row(&uh, cw));
    vf[r * sw..r * sw + sw].copy_from_slice(&recon_full_row(&vh, cw));
  }
  (uf, vf)
}

/// The bottom-sited ENCODED row-stage oracle: reconstruct U / V to full width
/// with the vertical bottom blend ([`recon_full_bottom`]) then run that
/// full-resolution `Yuv444p` frame through a `with_native(false)` RGB-domain
/// resample — convert-each-row-then-bin-RGB, exactly what the NV encoded arm does
/// for `Bottom`.
#[allow(clippy::too_many_arguments)]
fn encoded_oracle_rgb_bottom(
  y: &[u8],
  u: &[u8],
  v: &[u8],
  sw: usize,
  sh: usize,
  ow: usize,
  oh: usize,
  simd: bool,
) -> Vec<u8> {
  let (uf, vf) = recon_full_bottom(u, v, sw, sh);
  let mut rgb = vec![0u8; ow * oh * 3];
  {
    let mut sink =
      MixedSinker::<Yuv444p, AreaResampler>::with_resampler(sw, sh, AreaResampler::to(ow, oh))
        .unwrap()
        .with_native(false)
        .with_simd(simd)
        .with_rgb(&mut rgb)
        .unwrap();
    let f = Yuv444pFrame::new(
      y, &uf, &vf, sw as u32, sh as u32, sw as u32, sw as u32, sw as u32,
    );
    yuv444p_to(&f, FR, sink.set_kernel_matrix(M)).unwrap();
  }
  rgb
}

/// A fixture with flat luma and a strong per-ROW chroma step (flat across
/// columns), so the vertical bottom fold is observable in isolation: a
/// horizontal-only siting leaves it untouched, the `v = 1` blend visibly moves
/// it. `sw` / `sh` must be even.
fn vramp(sw: usize, sh: usize) -> (Vec<u8>, Vec<u8>, Vec<u8>) {
  let cw = sw / 2;
  let ch = sh / 2;
  let y = vec![128u8; sw * sh];
  let mut u = vec![0u8; cw * ch];
  let mut v = vec![0u8; cw * ch];
  for r in 0..ch {
    for c in 0..cw {
      u[r * cw + c] = (20 + r * 40).min(240) as u8;
      v[r * cw + c] = (220u32.saturating_sub((r * 40) as u32)).max(16) as u8;
    }
  }
  (y, u, v)
}

// The full test set is identical bar the format marker, frame type, walker, row
// type, and chroma interleave order, so generate it once per format. Each test
// lands in its own `mod` so the names don't collide.
macro_rules! nv_resample_siting_tests {
  ($mod:ident, $Marker:ident, $Frame:ident, $walker:ident, $Row:ident, $swap:expr) => {
    mod $mod {
      use super::*;

      /// Drive an `Nv12` / `Nv21` area resample (`sw x sh -> ow x oh`) for the
      /// full output set, at `loc` siting and `native` tier — the interleaved
      /// chroma is built from the planar U / V fixture.
      #[allow(clippy::too_many_arguments)]
      fn run(
        y: &[u8],
        u: &[u8],
        v: &[u8],
        sw: usize,
        sh: usize,
        ow: usize,
        oh: usize,
        loc: ChromaLocation,
        native: bool,
        simd: bool,
      ) -> Outs {
        let (cw, ch) = (sw / 2, sh / 2);
        let uv = interleave(u, v, cw, ch, $swap);
        let mut rgb = vec![0u8; ow * oh * 3];
        let mut rgba = vec![0u8; ow * oh * 4];
        let (mut hh, mut ss, mut vv) = (vec![0u8; ow * oh], vec![0u8; ow * oh], vec![0u8; ow * oh]);
        let mut luma = vec![0u8; ow * oh];
        let mut luma_u16 = vec![0u16; ow * oh];
        {
          let mut sink = MixedSinker::<$Marker, AreaResampler>::with_resampler(
            sw,
            sh,
            AreaResampler::to(ow, oh),
          )
          .unwrap()
          .with_native(native)
          .with_chroma_location(loc.clone())
          .with_simd(simd)
          .with_rgb(&mut rgb)
          .unwrap()
          .with_rgba(&mut rgba)
          .unwrap()
          .with_hsv(&mut hh, &mut ss, &mut vv)
          .unwrap()
          .with_luma(&mut luma)
          .unwrap()
          .with_luma_u16(&mut luma_u16)
          .unwrap();
          let f = $Frame::new(y, &uv, sw as u32, sh as u32, sw as u32, sw as u32);
          $walker(&f, FR, sink.set_kernel_matrix(M)).unwrap();
        }
        (rgb, rgba, (hh, ss, vv), luma, luma_u16)
      }

      /// A single-kernel Triangle FILTER resample of the fixture, as this NV
      /// format — the RGB output only.
      #[allow(clippy::too_many_arguments)]
      fn filter_rgb(
        y: &[u8],
        u: &[u8],
        v: &[u8],
        sw: usize,
        sh: usize,
        ow: usize,
        oh: usize,
        loc: ChromaLocation,
      ) -> Vec<u8> {
        let (cw, ch) = (sw / 2, sh / 2);
        let uv = interleave(u, v, cw, ch, $swap);
        let mut rgb = vec![0u8; ow * oh * 3];
        {
          let mut sink = MixedSinker::<$Marker, FilteredResampler<Triangle>>::with_resampler(
            sw,
            sh,
            FilteredResampler::new(ow, oh, Triangle),
          )
          .unwrap()
          .with_chroma_location(loc.clone())
          .with_rgb(&mut rgb)
          .unwrap();
          let f = $Frame::new(y, &uv, sw as u32, sh as u32, sw as u32, sw as u32);
          $walker(&f, FR, sink.set_kernel_matrix(M)).unwrap();
        }
        rgb
      }

      /// Reuse ONE native-tier sink across two frames of the SAME content, siting
      /// `loc1` then `loc2` (via `set_chroma_location` between walks).
      #[allow(clippy::too_many_arguments)]
      fn run_reuse_native(
        y: &[u8],
        u: &[u8],
        v: &[u8],
        sw: usize,
        sh: usize,
        ow: usize,
        oh: usize,
        loc1: ChromaLocation,
        loc2: ChromaLocation,
      ) -> Outs {
        let (cw, ch) = (sw / 2, sh / 2);
        let uv = interleave(u, v, cw, ch, $swap);
        let mut rgb = vec![0u8; ow * oh * 3];
        let mut rgba = vec![0u8; ow * oh * 4];
        let (mut hh, mut ss, mut vv) = (vec![0u8; ow * oh], vec![0u8; ow * oh], vec![0u8; ow * oh]);
        let mut luma = vec![0u8; ow * oh];
        let mut luma_u16 = vec![0u16; ow * oh];
        {
          let mut sink = MixedSinker::<$Marker, AreaResampler>::with_resampler(
            sw,
            sh,
            AreaResampler::to(ow, oh),
          )
          .unwrap()
          .with_native(true)
          .with_rgb(&mut rgb)
          .unwrap()
          .with_rgba(&mut rgba)
          .unwrap()
          .with_hsv(&mut hh, &mut ss, &mut vv)
          .unwrap()
          .with_luma(&mut luma)
          .unwrap()
          .with_luma_u16(&mut luma_u16)
          .unwrap();
          let f = $Frame::new(y, &uv, sw as u32, sh as u32, sw as u32, sw as u32);
          sink.set_chroma_location(loc1.clone());
          $walker(&f, FR, sink.set_kernel_matrix(M)).unwrap();
          sink.set_chroma_location(loc2.clone());
          $walker(&f, FR, sink.set_kernel_matrix(M)).unwrap();
        }
        (rgb, rgba, (hh, ss, vv), luma, luma_u16)
      }

      // ---- co-sited byte-identity (the regression contract) ----------------

      #[test]
      #[cfg_attr(
        miri,
        ignore = "SIMD-dispatched row kernels use intrinsics unsupported by Miri"
      )]
      fn cosited_group_is_byte_identical_across_tiers() {
        // Every co-sited / unspecified siting must produce the byte-identical
        // pre-siting resample (phase 0 → the folded plan is never built), on BOTH
        // tiers. `Unspecified` is the baseline. `TopLeft` is EXCLUDED (RFC #238):
        // co-sited horizontally but top-sited vertically (`v = 0`), so it folds the
        // FORWARD odd-row blend and leaves the co-sited byte-identity group.
        let (y, u, v) = ramp(8, 8);
        for native in [true, false] {
          let base = run(
            &y,
            &u,
            &v,
            8,
            8,
            4,
            4,
            ChromaLocation::Unspecified,
            native,
            true,
          );
          for loc in [
            ChromaLocation::Left,
            ChromaLocation::other("unassigned-7"),
          ] {
            let got = run(&y, &u, &v, 8, 8, 4, 4, loc.clone(), native, true);
            assert_eq!(got.0, base.0, "rgb {loc:?} native={native}");
            assert_eq!(got.1, base.1, "rgba {loc:?} native={native}");
            assert_eq!(got.2, base.2, "hsv {loc:?} native={native}");
            assert_eq!(got.3, base.3, "luma {loc:?} native={native}");
            assert_eq!(got.4, base.4, "luma_u16 {loc:?} native={native}");
          }
        }
      }

      // ---- centered native == the exact code-domain oracle -----------------

      #[test]
      #[cfg_attr(
        miri,
        ignore = "SIMD-dispatched row kernels use intrinsics unsupported by Miri"
      )]
      fn centered_native_equals_code_domain_oracle() {
        // Clean 2:1 and fractional ratios (EVEN source height so the vertical
        // luma-domain pairing equals the co-sited chroma box), for the
        // vertically-co-sited centered siting (Center). RFC #238 S4-C Bottom (and
        // the Top FORWARD fold) fold the vertical phase and are pinned to their own
        // oracles / the Yuv420p cross-format cross-check below.
        for (sw, sh, ow, oh) in [(8, 8, 4, 4), (8, 8, 5, 3), (12, 8, 4, 4), (16, 8, 6, 5)] {
          let (y, u, v) = ramp(sw, sh);
          let o = native_oracle(&y, &u, &v, sw, sh, ow, oh, true);
          for loc in [ChromaLocation::Center] {
            let n = run(&y, &u, &v, sw, sh, ow, oh, loc.clone(), true, true);
            assert_eq!(n.0, o.0, "rgb {loc:?} {sw}x{sh}->{ow}x{oh}");
            assert_eq!(n.1, o.1, "rgba {loc:?} {sw}x{sh}->{ow}x{oh}");
            assert_eq!(n.2, o.2, "hsv {loc:?} {sw}x{sh}->{ow}x{oh}");
            assert_eq!(n.3, o.3, "luma {loc:?} {sw}x{sh}->{ow}x{oh}");
            assert_eq!(n.4, o.4, "luma_u16 {loc:?} {sw}x{sh}->{ow}x{oh}");
          }
        }
      }

      #[test]
      #[cfg_attr(
        miri,
        ignore = "SIMD-dispatched row kernels use intrinsics unsupported by Miri"
      )]
      fn centered_native_simd_matches_scalar() {
        // Weights are precomputed integers, so the SIMD H/V passes must be 0-ULP.
        let (y, u, v) = ramp(8, 8);
        let s = run(&y, &u, &v, 8, 8, 4, 4, ChromaLocation::Center, true, false);
        let d = run(&y, &u, &v, 8, 8, 4, 4, ChromaLocation::Center, true, true);
        assert_eq!(s.0, d.0, "rgb scalar vs simd");
        assert_eq!(s.2, d.2, "hsv scalar vs simd");
        assert_eq!(s.3, d.3, "luma scalar vs simd");
      }

      // ---- centered encoded row-stage == RGB-domain reconstruct-then-bin ----

      #[test]
      #[cfg_attr(
        miri,
        ignore = "SIMD-dispatched row kernels use intrinsics unsupported by Miri"
      )]
      fn centered_encoded_output_equals_rgb_reconstruct_then_bin() {
        // Center only — RFC #238 S4-C Bottom (and the Top FORWARD fold) fold the
        // vertical phase and are pinned to their own reconstruct-then-bin oracles /
        // the Yuv420p cross-format cross-check below.
        for (sw, sh, ow, oh) in [(8, 8, 4, 4), (8, 8, 5, 3), (12, 8, 6, 4)] {
          let (y, u, v) = ramp(sw, sh);
          let oracle = encoded_oracle_rgb(&y, &u, &v, sw, sh, ow, oh, true);
          for loc in [ChromaLocation::Center] {
            let got = run(&y, &u, &v, sw, sh, ow, oh, loc.clone(), false, true);
            assert_eq!(got.0, oracle, "rgb {loc:?} {sw}x{sh}->{ow}x{oh}");
          }
        }
      }

      // ---- centered NV is bit-identical to centered Yuv420p (U/V-swap catch) --

      #[test]
      #[cfg_attr(
        miri,
        ignore = "SIMD-dispatched row kernels use intrinsics unsupported by Miri"
      )]
      fn centered_equals_centered_yuv420p_area_tiers() {
        // De-interleaving the packed chroma back to planar U / V reconstructs the
        // SAME planes the planar twin holds, so a centered NV area resample is
        // byte-identical to the centered Yuv420p resample of those planes — on
        // BOTH the native fast tier and the encoded row-stage tier. The strongest
        // catch for a U/V swap in the de-interleave (`swap_uv`). Bt601 keeps this
        // on the shared matrix-tag path (ChromaDerivedNcl is the lone divergence).
        // Now that RFC #238 S4-C folds the NV vertical `Bottom` v=1 phase through
        // the resample exactly as the Yuv420p twin (S4-B), Bottom rejoins the
        // loop: NV-Bottom == Yuv420p-Bottom is a strong cross-format cross-check.
        // `BottomLeft` (co-sited h + v=1) rides the SAME shared reconstruction, so
        // it joins too — pinning NV BottomLeft against the validated Yuv420p one.
        // `Top` (centered h + v=0) / `TopLeft` (co-sited h + v=0) ride the SAME
        // shared FORWARD one-row delay now activated for NV (RFC #238), so they too
        // must match the validated Yuv420p Top / TopLeft on BOTH area tiers — the
        // strongest cross-tier (native == row-stage) + cross-format Top proof.
        for (sw, sh, ow, oh) in [(8, 8, 4, 4), (8, 8, 5, 3), (12, 8, 4, 4), (16, 8, 6, 5)] {
          let (y, u, v) = ramp(sw, sh);
          for loc in [
            ChromaLocation::Center,
            ChromaLocation::Bottom,
            ChromaLocation::BottomLeft,
            ChromaLocation::Top,
            ChromaLocation::TopLeft,
          ] {
            for native in [true, false] {
              let nv = run(&y, &u, &v, sw, sh, ow, oh, loc.clone(), native, true);
              let yuv420p = run_yuv420p(&y, &u, &v, sw, sh, ow, oh, loc.clone(), native, true);
              assert_eq!(
                nv, yuv420p,
                "NV vs Yuv420p {loc:?} native={native} {sw}x{sh}->{ow}x{oh}"
              );
            }
          }
        }
      }

      #[test]
      #[cfg_attr(
        miri,
        ignore = "SIMD-dispatched row kernels use intrinsics unsupported by Miri"
      )]
      fn centered_equals_centered_yuv420p_filter_tier() {
        // The filter tier's centered reconstruction is the SAME as the planar
        // twin's, so the Triangle-filtered centered NV RGB equals the Yuv420p one;
        // the centered result also genuinely differs from co-sited on the ramp.
        for (sw, sh, ow, oh) in [(8, 8, 4, 4), (8, 8, 5, 3)] {
          let (y, u, v) = ramp(sw, sh);
          let cw = sw / 2;
          let mut rgb420 = vec![0u8; ow * oh * 3];
          {
            let mut sink = MixedSinker::<Yuv420p, FilteredResampler<Triangle>>::with_resampler(
              sw,
              sh,
              FilteredResampler::new(ow, oh, Triangle),
            )
            .unwrap()
            .with_chroma_location(ChromaLocation::Center)
            .with_rgb(&mut rgb420)
            .unwrap();
            let f = Yuv420pFrame::new(
              &y, &u, &v, sw as u32, sh as u32, sw as u32, cw as u32, cw as u32,
            );
            yuv420p_to(&f, FR, sink.set_kernel_matrix(M)).unwrap();
          }
          let nv = filter_rgb(&y, &u, &v, sw, sh, ow, oh, ChromaLocation::Center);
          assert_eq!(nv, rgb420, "filter centered {sw}x{sh}->{ow}x{oh}");
          let cosited = filter_rgb(&y, &u, &v, sw, sh, ow, oh, ChromaLocation::Left);
          assert_ne!(nv, cosited, "filter centered must differ from co-sited");
        }
      }

      #[test]
      #[cfg_attr(
        miri,
        ignore = "SIMD-dispatched row kernels use intrinsics unsupported by Miri"
      )]
      fn bottomleft_filter_equals_bottomleft_yuv420p() {
        // The filter tier's BottomLeft reconstruction (co-sited h + v=1 vertical
        // blend) is the SAME as the planar twin's, so the Triangle-filtered NV
        // BottomLeft RGB equals the Yuv420p one, and genuinely differs from Bottom
        // (the horizontal phase) on a ramp that varies BOTH axes.
        for (sw, sh, ow, oh) in [(8, 8, 4, 4), (8, 8, 5, 3)] {
          let (y, u, v) = ramp(sw, sh);
          let cw = sw / 2;
          let mut rgb420 = vec![0u8; ow * oh * 3];
          {
            let mut sink = MixedSinker::<Yuv420p, FilteredResampler<Triangle>>::with_resampler(
              sw,
              sh,
              FilteredResampler::new(ow, oh, Triangle),
            )
            .unwrap()
            .with_chroma_location(ChromaLocation::BottomLeft)
            .with_rgb(&mut rgb420)
            .unwrap();
            let f = Yuv420pFrame::new(
              &y, &u, &v, sw as u32, sh as u32, sw as u32, cw as u32, cw as u32,
            );
            yuv420p_to(&f, FR, sink.set_kernel_matrix(M)).unwrap();
          }
          let nv = filter_rgb(&y, &u, &v, sw, sh, ow, oh, ChromaLocation::BottomLeft);
          assert_eq!(nv, rgb420, "filter bottom-left {sw}x{sh}->{ow}x{oh}");
          let bottom = filter_rgb(&y, &u, &v, sw, sh, ow, oh, ChromaLocation::Bottom);
          assert_ne!(nv, bottom, "filter bottom-left must differ from bottom (h phase)");
        }
      }

      #[test]
      #[cfg_attr(
        miri,
        ignore = "SIMD-dispatched row kernels use intrinsics unsupported by Miri"
      )]
      fn top_filter_equals_top_yuv420p() {
        // The filter tier's `Top` / `TopLeft` FORWARD one-row-delay reconstruction
        // is the SAME as the planar twin's, so the Triangle-filtered NV RGB equals
        // the validated Yuv420p one (RFC #238) and genuinely differs from Center
        // (horizontal-only) and Bottom (the backward fold) on a two-axis ramp.
        for (sw, sh, ow, oh) in [(8, 8, 4, 4), (8, 8, 5, 3)] {
          let (y, u, v) = ramp(sw, sh);
          let cw = sw / 2;
          for loc in [ChromaLocation::Top, ChromaLocation::TopLeft] {
            let mut rgb420 = vec![0u8; ow * oh * 3];
            {
              let mut sink = MixedSinker::<Yuv420p, FilteredResampler<Triangle>>::with_resampler(
                sw,
                sh,
                FilteredResampler::new(ow, oh, Triangle),
              )
              .unwrap()
              .with_chroma_location(loc.clone())
              .with_rgb(&mut rgb420)
              .unwrap();
              let f = Yuv420pFrame::new(
                &y, &u, &v, sw as u32, sh as u32, sw as u32, cw as u32, cw as u32,
              );
              yuv420p_to(&f, FR, sink.set_kernel_matrix(M)).unwrap();
            }
            let nv = filter_rgb(&y, &u, &v, sw, sh, ow, oh, loc.clone());
            assert_eq!(nv, rgb420, "filter {loc:?} {sw}x{sh}->{ow}x{oh}");
          }
          let top = filter_rgb(&y, &u, &v, sw, sh, ow, oh, ChromaLocation::Top);
          let center = filter_rgb(&y, &u, &v, sw, sh, ow, oh, ChromaLocation::Center);
          let bottom = filter_rgb(&y, &u, &v, sw, sh, ow, oh, ChromaLocation::Bottom);
          assert_ne!(top, center, "filter Top (forward v) must differ from Center");
          assert_ne!(top, bottom, "filter Top (forward) must differ from Bottom (backward)");
        }
      }

      // ---- non-vacuous + flat-chroma sanity --------------------------------

      #[test]
      #[cfg_attr(
        miri,
        ignore = "SIMD-dispatched row kernels use intrinsics unsupported by Miri"
      )]
      fn centered_differs_from_cosited_on_a_chroma_ramp() {
        // The phase must actually DO something: on a horizontal chroma ramp the
        // centered decode diverges from co-sited on both tiers.
        let (y, u, v) = ramp(8, 8);
        for native in [true, false] {
          let cos = run(
            &y,
            &u,
            &v,
            8,
            8,
            4,
            4,
            ChromaLocation::Unspecified,
            native,
            true,
          );
          let cen = run(&y, &u, &v, 8, 8, 4, 4, ChromaLocation::Center, native, true);
          assert_ne!(
            cen.0, cos.0,
            "centered rgb must differ from co-sited (native={native})"
          );
        }
      }

      #[test]
      #[cfg_attr(
        miri,
        ignore = "SIMD-dispatched row kernels use intrinsics unsupported by Miri"
      )]
      fn centered_equals_cosited_on_flat_chroma() {
        // Sanity: on constant chroma the centered triangle is a no-op, so centered
        // and co-sited agree byte-for-byte (the phase machinery corrupts nothing).
        let (y, u, v) = flat_chroma(8, 8);
        for native in [true, false] {
          let cos = run(&y, &u, &v, 8, 8, 4, 4, ChromaLocation::Left, native, true);
          let cen = run(&y, &u, &v, 8, 8, 4, 4, ChromaLocation::Center, native, true);
          assert_eq!(cen.0, cos.0, "flat-chroma rgb (native={native})");
          assert_eq!(cen.2, cos.2, "flat-chroma hsv (native={native})");
        }
      }

      // ---- RFC #238 S4-C bottom-sited (v = 1) vertical fold -----------------

      /// The DIRECT (non-resample) NV decode to RGB at `loc` — the already
      /// validated identity delay-line path the resample must match at identity
      /// dimensions.
      fn direct_rgb(
        y: &[u8],
        u: &[u8],
        v: &[u8],
        sw: usize,
        sh: usize,
        loc: ChromaLocation,
        simd: bool,
      ) -> Vec<u8> {
        let (cw, ch) = (sw / 2, sh / 2);
        let uv = interleave(u, v, cw, ch, $swap);
        let mut rgb = vec![0u8; sw * sh * 3];
        {
          let mut sink = MixedSinker::<$Marker>::new(sw, sh)
            .with_chroma_location(loc.clone())
            .with_simd(simd)
            .with_rgb(&mut rgb)
            .unwrap();
          let f = $Frame::new(y, &uv, sw as u32, sh as u32, sw as u32, sw as u32);
          $walker(&f, FR, sink.set_kernel_matrix(M)).unwrap();
        }
        rgb
      }

      #[test]
      #[cfg_attr(
        miri,
        ignore = "SIMD-dispatched row kernels use intrinsics unsupported by Miri"
      )]
      fn bottom_native_equals_code_domain_oracle() {
        // The native tier folds the vertical v=1 triangle into the chroma area
        // weights; its output is the EXACT code-domain box-average of the
        // UNROUNDED H⊗V reconstruction (single rounding), EVEN source heights.
        for (sw, sh, ow, oh) in [(8, 8, 4, 4), (8, 8, 5, 3), (12, 8, 4, 4), (16, 8, 6, 5)] {
          let (y, u, v) = ramp(sw, sh);
          let o = bottom_native_oracle(&y, &u, &v, sw, sh, ow, oh, true);
          let n = run(
            &y,
            &u,
            &v,
            sw,
            sh,
            ow,
            oh,
            ChromaLocation::Bottom,
            true,
            true,
          );
          assert_eq!(n.0, o.0, "rgb {sw}x{sh}->{ow}x{oh}");
          assert_eq!(n.1, o.1, "rgba {sw}x{sh}->{ow}x{oh}");
          assert_eq!(n.2, o.2, "hsv {sw}x{sh}->{ow}x{oh}");
          assert_eq!(n.3, o.3, "luma {sw}x{sh}->{ow}x{oh}");
          assert_eq!(n.4, o.4, "luma_u16 {sw}x{sh}->{ow}x{oh}");
        }
      }

      #[test]
      #[cfg_attr(
        miri,
        ignore = "SIMD-dispatched row kernels use intrinsics unsupported by Miri"
      )]
      fn bottom_native_simd_matches_scalar() {
        // Precomputed integer weights → the SIMD H/V passes are 0-ULP vs scalar.
        let (y, u, v) = ramp(8, 8);
        let s = run(&y, &u, &v, 8, 8, 4, 4, ChromaLocation::Bottom, true, false);
        let d = run(&y, &u, &v, 8, 8, 4, 4, ChromaLocation::Bottom, true, true);
        assert_eq!(s.0, d.0, "rgb scalar vs simd");
        assert_eq!(s.2, d.2, "hsv scalar vs simd");
        assert_eq!(s.3, d.3, "luma scalar vs simd");
      }

      #[test]
      #[cfg_attr(
        miri,
        ignore = "SIMD-dispatched row kernels use intrinsics unsupported by Miri"
      )]
      fn bottom_encoded_output_equals_rgb_reconstruct_then_bin() {
        // The encoded row-stage tier reconstructs full-width chroma with the
        // vertical bottom blend (de-interleave + v=1 fold) then bins in RGB.
        for (sw, sh, ow, oh) in [(8, 8, 4, 4), (8, 8, 5, 3), (12, 8, 6, 4)] {
          let (y, u, v) = ramp(sw, sh);
          let oracle = encoded_oracle_rgb_bottom(&y, &u, &v, sw, sh, ow, oh, true);
          let got = run(
            &y,
            &u,
            &v,
            sw,
            sh,
            ow,
            oh,
            ChromaLocation::Bottom,
            false,
            true,
          );
          assert_eq!(got.0, oracle, "rgb {sw}x{sh}->{ow}x{oh}");
        }
      }

      #[test]
      #[cfg_attr(
        miri,
        ignore = "SIMD-dispatched row kernels use intrinsics unsupported by Miri"
      )]
      fn bottom_differs_from_center_on_a_vertical_chroma_ramp() {
        // The v=1 fold must actually MOVE the chroma: on a purely-vertical chroma
        // ramp Bottom diverges from the vertically-co-sited Center on BOTH tiers.
        let (y, u, v) = vramp(8, 8);
        for native in [true, false] {
          let cen = run(&y, &u, &v, 8, 8, 4, 4, ChromaLocation::Center, native, true);
          let bot = run(&y, &u, &v, 8, 8, 4, 4, ChromaLocation::Bottom, native, true);
          assert_ne!(
            bot.0, cen.0,
            "bottom rgb must differ from center (native={native})"
          );
        }
      }

      #[test]
      #[cfg_attr(
        miri,
        ignore = "SIMD-dispatched row kernels use intrinsics unsupported by Miri"
      )]
      fn bottom_equals_cosited_on_flat_chroma() {
        // On constant chroma the vertical blend (and horizontal triangle) are
        // no-ops, so Bottom collapses to the co-sited decode byte-for-byte.
        let (y, u, v) = flat_chroma(8, 8);
        for native in [true, false] {
          let cos = run(&y, &u, &v, 8, 8, 4, 4, ChromaLocation::Left, native, true);
          let bot = run(&y, &u, &v, 8, 8, 4, 4, ChromaLocation::Bottom, native, true);
          assert_eq!(bot.0, cos.0, "flat-chroma bottom rgb (native={native})");
          assert_eq!(bot.2, cos.2, "flat-chroma bottom hsv (native={native})");
        }
      }

      #[test]
      #[cfg_attr(
        miri,
        ignore = "SIMD-dispatched row kernels use intrinsics unsupported by Miri"
      )]
      fn bottom_encoded_identity_matches_direct_decode() {
        // At identity dimensions (out == src) the encoded resample tier
        // reconstructs chroma with the SAME bottom kernel as the direct NV decode
        // and bins with pass-through area weights, so routing Bottom through the
        // resample preserves the decode byte-for-byte.
        let (y, u, v) = vramp(8, 8);
        let res = run(&y, &u, &v, 8, 8, 8, 8, ChromaLocation::Bottom, false, true);
        let direct = direct_rgb(&y, &u, &v, 8, 8, ChromaLocation::Bottom, true);
        assert_eq!(
          res.0, direct,
          "identity encoded resample bottom == direct decode"
        );
      }

      #[test]
      #[cfg_attr(
        miri,
        ignore = "SIMD-dispatched row kernels use intrinsics unsupported by Miri"
      )]
      fn bottom_filter_equals_reconstruct_then_filter() {
        // The single-kernel filter tier reconstructs full-width chroma with the
        // vertical bottom blend then Triangle-filters it — equal to feeding the
        // same reconstruction through a Yuv444p Triangle filter, and non-vacuously
        // different from the vertically-co-sited Center on a vertical ramp.
        use crate::resample::{FilteredResampler, Triangle};
        let (sw, sh, ow, oh) = (8usize, 8usize, 4usize, 4usize);
        let (y, u, v) = vramp(sw, sh);
        let got = filter_rgb(&y, &u, &v, sw, sh, ow, oh, ChromaLocation::Bottom);
        let (uf, vf) = recon_full_bottom(&u, &v, sw, sh);
        let mut oracle = vec![0u8; ow * oh * 3];
        {
          let mut sink = MixedSinker::<Yuv444p, FilteredResampler<Triangle>>::with_resampler(
            sw,
            sh,
            FilteredResampler::new(ow, oh, Triangle),
          )
          .unwrap()
          .with_simd(true)
          .with_rgb(&mut oracle)
          .unwrap();
          let f = Yuv444pFrame::new(
            &y, &uf, &vf, sw as u32, sh as u32, sw as u32, sw as u32, sw as u32,
          );
          yuv444p_to(&f, FR, sink.set_kernel_matrix(M)).unwrap();
        }
        assert_eq!(got, oracle, "filter-tier bottom == reconstruct-then-filter");
        assert_ne!(
          got,
          filter_rgb(&y, &u, &v, sw, sh, ow, oh, ChromaLocation::Center),
          "filter-tier bottom must differ from center on a vertical ramp"
        );
      }

      #[test]
      #[cfg_attr(
        miri,
        ignore = "SIMD-dispatched row kernels use intrinsics unsupported by Miri"
      )]
      fn bottom_final_even_row_reserve_failure_retries_with_blend() {
        // ODD source height => the FINAL luma row is EVEN, so its Bottom vertical
        // blend reads the PREDECESSOR chroma row via the lookback. The `chroma_prev`
        // reserve is ordered AFTER the resample preflight but BEFORE the commit
        // (#180): a simulated allocator refusal on that final row must surface
        // AllocationFailed with NO output written and the lookback left intact, so
        // the retry of the SAME row still box-blends the predecessor (not clamps).
        // (The streaming row-stage commit is infallible past the reserve, so the
        // reserve is the injectable failpoint; the deferred staging — advancing the
        // lookback only after the commit accepts — mirrors the planar twin and is
        // covered for correctness by the multi-row encoded / identity oracles.)
        use super::super::MixedSinkerError;
        let (sw, sh, ow, oh) = (8usize, 7usize, 4usize, 3usize);
        let cw = sw / 2;
        let ch = sh.div_ceil(2);
        let y = vec![128u8; sw * sh];
        let mut u = vec![0u8; cw * ch];
        let mut v = vec![0u8; cw * ch];
        for r in 0..ch {
          for c in 0..cw {
            u[r * cw + c] = (20 + r * 40).min(240) as u8;
            v[r * cw + c] = (220u32.saturating_sub((r * 40) as u32)).max(16) as u8;
          }
        }
        let full = 2 * cw;
        let uv = interleave(&u, &v, cw, ch, $swap);
        let run_frame = |arm_fail: bool, loc: ChromaLocation| -> Vec<u8> {
          let mut rgb = vec![0u8; ow * oh * 3];
          {
            let mut sink = MixedSinker::<$Marker, AreaResampler>::with_resampler(
              sw,
              sh,
              AreaResampler::to(ow, oh),
            )
            .unwrap()
            .with_native(false)
            .with_chroma_location(loc.clone())
            .with_rgb(&mut rgb)
            .unwrap();
            sink.begin_frame(sw as u32, sh as u32).unwrap();
            let feed = |sink: &mut MixedSinker<'_, $Marker, AreaResampler>, r: usize| {
              let yr = &y[r * sw..(r + 1) * sw];
              let cr = r / 2;
              let uvr = &uv[cr * full..(cr + 1) * full];
              sink.process($Row::for_tests(yr, uvr, r, M, FR))
            };
            for r in 0..sh - 1 {
              feed(&mut sink, r).unwrap();
            }
            if arm_fail {
              crate::sinker::mixed::arm_chroma_prev_alloc_failure();
              let err = feed(&mut sink, sh - 1).unwrap_err();
              assert!(
                matches!(
                  err,
                  MixedSinkerError::Resample(ResampleError::AllocationFailed(_))
                ),
                "armed final-row chroma_prev reserve must surface AllocationFailed, got {err:?}"
              );
            }
            feed(&mut sink, sh - 1).unwrap();
          }
          rgb
        };
        let reference = run_frame(false, ChromaLocation::Bottom);
        let retried = run_frame(true, ChromaLocation::Bottom);
        assert_eq!(
          retried, reference,
          "the post-failure retry of the final EVEN row must blend with the predecessor (not clamp)"
        );
        assert_ne!(
          reference,
          run_frame(false, ChromaLocation::Center),
          "the bottom vertical blend must move the final even row vs co-sited Center"
        );
      }

      // ---- no-output rows are invisible to the lookback (no-output invariant) --

      #[test]
      #[cfg_attr(
        miri,
        ignore = "constructs an absurd geometry; the no-op contract is the point, not Miri"
      )]
      fn bottom_no_output_large_geometry_does_not_overflow() {
        // RFC #238 S4-C parity with the Yuv420p direct twin: the no-output guard
        // must run BEFORE the `idx * w` single-plane offset arithmetic. A
        // no-output `process` never ran an attach-time `w x h x 1` validation, so
        // on a 32-bit target `idx * w` could overflow `usize` and panic under
        // overflow checks. With no outputs attached, `process` must return Ok
        // having done NO row math and NO allocation (the vertical lookback
        // included). On 64-bit this documents the contract; on the i686 /
        // overflow-check CI it catches a regression if the guard moves back below
        // the arithmetic. w = 4, idx = 2^30 → idx * w = 2^32 (overflows u32).
        let w: usize = 4;
        let idx: usize = 1 << 30;
        let h: usize = idx + 1;
        assert!(
          (idx as u64) * (w as u64) > u32::MAX as u64,
          "test geometry must exceed u32::MAX to exercise the 32-bit offset overflow"
        );
        let y = vec![128u8; w];
        let uv = vec![128u8; w];
        let mut sink =
          MixedSinker::<$Marker>::new(w, h).with_chroma_location(ChromaLocation::Bottom);
        crate::PixelSink::process(&mut sink, $Row::for_tests(&y, &uv, idx, M, FR)).unwrap();
        let prev_len = sink.chroma_prev.len();
        drop(sink);
        assert_eq!(
          prev_len, 0,
          "a no-output large-geometry bottom row must allocate nothing (lookback included)"
        );
      }

      #[test]
      #[cfg_attr(
        miri,
        ignore = "SIMD-dispatched row kernels use intrinsics unsupported by Miri"
      )]
      fn bottom_no_output_rows_do_not_enable_late_color_blend() {
        // Rows 0,1 have NO output attached — invisible, so they must NOT prime the
        // vertical lookback. After a late RGB attach, row 2 (Bottom even, pair 1)
        // must CLAMP to the centered upsample of the CURRENT chroma row (no
        // legitimately-output predecessor was staged), NOT box-blend through the
        // invisible rows — so it equals the co-sited-vertical Center decode's row
        // 2, and differs from the all-output Bottom box-blend.
        let (y, u, v) = vramp(8, 8);
        let (w, h, cw, ch) = (8usize, 8usize, 4usize, 4usize);
        let full = 2 * cw;
        let uv = interleave(&u, &v, cw, ch, $swap);
        let mut rgb = vec![0u8; w * h * 3];
        {
          let mut sink =
            MixedSinker::<$Marker>::new(w, h).with_chroma_location(ChromaLocation::Bottom);
          crate::PixelSink::begin_frame(&mut sink, w as u32, h as u32).unwrap();
          let feed = |sink: &mut MixedSinker<'_, $Marker>, r: usize| {
            let uvr = &uv[(r / 2) * full..(r / 2) * full + full];
            crate::PixelSink::process(sink, $Row::for_tests(&y[r * w..r * w + w], uvr, r, M, FR)).unwrap();
          };
          feed(&mut sink, 0);
          feed(&mut sink, 1);
          sink.set_rgb(&mut rgb).unwrap();
          feed(&mut sink, 2);
          feed(&mut sink, 3);
        }
        let center = direct_rgb(&y, &u, &v, w, h, ChromaLocation::Center, true);
        let bottom = direct_rgb(&y, &u, &v, w, h, ChromaLocation::Bottom, true);
        assert_eq!(
          &rgb[2 * w * 3..3 * w * 3],
          &center[2 * w * 3..3 * w * 3],
          "no-output predecessor rows must not enable a later colour even row to box-blend"
        );
        assert_ne!(
          &center[2 * w * 3..3 * w * 3],
          &bottom[2 * w * 3..3 * w * 3],
          "guard: the clamp (Center) must differ from the Bottom box-blend on row 2"
        );
      }

      #[test]
      #[cfg_attr(
        miri,
        ignore = "SIMD-dispatched row kernels use intrinsics unsupported by Miri"
      )]
      fn bottom_luma_only_then_late_color_box_blends() {
        // The correctness twin: rows 0,1 are LUMA-ONLY — they DO prime the
        // lookback via the luma-only de-interleave+stage path. After a late RGB
        // attach, row 2 (Bottom even, pair 1) box-blends chroma rows 0,1, matching
        // the all-output Bottom decode's row 2 (and differing from the Center
        // clamp) — proving the luma-only staging works end-to-end.
        let (y, u, v) = vramp(8, 8);
        let (w, h, cw, ch) = (8usize, 8usize, 4usize, 4usize);
        let full = 2 * cw;
        let uv = interleave(&u, &v, cw, ch, $swap);
        let mut luma = vec![0u8; w * h];
        let mut rgb = vec![0u8; w * h * 3];
        {
          let mut sink = MixedSinker::<$Marker>::new(w, h)
            .with_chroma_location(ChromaLocation::Bottom)
            .with_luma(&mut luma)
            .unwrap();
          crate::PixelSink::begin_frame(&mut sink, w as u32, h as u32).unwrap();
          let feed = |sink: &mut MixedSinker<'_, $Marker>, r: usize| {
            let uvr = &uv[(r / 2) * full..(r / 2) * full + full];
            crate::PixelSink::process(sink, $Row::for_tests(&y[r * w..r * w + w], uvr, r, M, FR)).unwrap();
          };
          feed(&mut sink, 0);
          feed(&mut sink, 1);
          sink.set_rgb(&mut rgb).unwrap();
          feed(&mut sink, 2);
          feed(&mut sink, 3);
        }
        let bottom = direct_rgb(&y, &u, &v, w, h, ChromaLocation::Bottom, true);
        assert_eq!(
          &rgb[2 * w * 3..3 * w * 3],
          &bottom[2 * w * 3..3 * w * 3],
          "luma-only predecessor rows must prime the lookback so the late-colour even row box-blends"
        );
        let center = direct_rgb(&y, &u, &v, w, h, ChromaLocation::Center, true);
        assert_ne!(
          &bottom[2 * w * 3..3 * w * 3],
          &center[2 * w * 3..3 * w * 3],
          "guard: the box-blend must differ from the Center clamp on row 2"
        );
      }

      // ---- cross-frame sink reuse rebuilds the phased native join ----------

      #[test]
      #[cfg_attr(
        miri,
        ignore = "SIMD-dispatched row kernels use intrinsics unsupported by Miri"
      )]
      fn native_join_rebuilds_on_siting_change_across_frames() {
        // Reuse one native-tier sink flipping Left ⇄ Center (both directions):
        // frame 2 must match a FRESH sink for frame 2's siting — no stale-phase
        // carryover from the cached folded/co-sited chroma plan.
        let (y, u, v) = ramp(8, 8);
        for (a, b) in [
          (ChromaLocation::Left, ChromaLocation::Center),
          (ChromaLocation::Center, ChromaLocation::Left),
        ] {
          let reused = run_reuse_native(&y, &u, &v, 8, 8, 4, 4, a.clone(), b.clone());
          let fresh = run(&y, &u, &v, 8, 8, 4, 4, b.clone(), true, true);
          assert_eq!(
            reused.0, fresh.0,
            "native rgb {a:?}->{b:?} stale-phase carryover"
          );
          assert_eq!(reused.2, fresh.2, "native hsv {a:?}->{b:?}");
          assert_eq!(reused.3, fresh.3, "native luma {a:?}->{b:?}");
          let stale = run(&y, &u, &v, 8, 8, 4, 4, a.clone(), true, true);
          assert_ne!(
            fresh.0, stale.0,
            "sitings {a:?} vs {b:?} must differ (non-vacuous)"
          );
        }
      }

      #[test]
      #[cfg_attr(
        miri,
        ignore = "SIMD-dispatched row kernels use intrinsics unsupported by Miri"
      )]
      fn native_out_of_sequence_first_row_does_not_drop_the_cached_join() {
        // The point-of-use phase drop is state-atomic: a reused sink, new frame,
        // changed siting, then an OUT-OF-SEQUENCE first process call must be
        // rejected (OutOfSequenceRow) with the cached join INTACT, so a corrected
        // row-0 retry rebuilds cleanly.
        let (y, u, v) = ramp(8, 8);
        let uv = interleave(&u, &v, 4, 4, $swap);
        let mut rgb = vec![0u8; 4 * 4 * 3];
        let mut sink =
          MixedSinker::<$Marker, AreaResampler>::with_resampler(8, 8, AreaResampler::to(4, 4))
            .unwrap()
            .with_native(true)
            .with_chroma_location(ChromaLocation::Left)
            .with_rgb(&mut rgb)
            .unwrap();
        // Frame 1 at Left builds the native join (chroma row `r / 2` per luma row).
        PixelSink::begin_frame(&mut sink, 8, 8).unwrap();
        for r in 0..8 {
          let row = $Row::for_tests(
            &y[r * 8..r * 8 + 8],
            &uv[(r / 2) * 8..(r / 2) * 8 + 8],
            r,
            M,
            FR,
          );
          PixelSink::process(&mut sink, row).unwrap();
        }
        assert!(sink.native_420.is_some(), "frame 1 builds the native join");
        // Frame 2: change siting to Center, then feed an OUT-OF-SEQUENCE first row.
        PixelSink::begin_frame(&mut sink, 8, 8).unwrap();
        sink.set_chroma_location(ChromaLocation::Center);
        let bad = $Row::for_tests(&y[4 * 8..5 * 8], &uv[2 * 8..3 * 8], 4, M, FR);
        let err = PixelSink::process(&mut sink, bad).unwrap_err();
        assert!(
          matches!(
            err,
            MixedSinkerError::Resample(ResampleError::OutOfSequenceRow(_))
          ),
          "out-of-sequence first row must be OutOfSequenceRow, got {err:?}"
        );
        assert!(
          sink.native_420.is_some(),
          "an out-of-sequence first row must NOT drop the cached native join"
        );
        // The corrected retry (row 0, now rebuilding for Center) succeeds.
        for r in 0..8 {
          let row = $Row::for_tests(
            &y[r * 8..r * 8 + 8],
            &uv[(r / 2) * 8..(r / 2) * 8 + 8],
            r,
            M,
            FR,
          );
          PixelSink::process(&mut sink, row).unwrap();
        }
      }

      // ---- IN-SEQUENCE mid-frame phase change is rejected (not silently mixed) --

      /// Drive one frame: `begin_frame`, accept row 0 at `loc1` (freezes the
      /// phase), flip to `loc2`, then feed the IN-SEQUENCE row 1 and return its
      /// `process` result. In 4:2:0 rows 0 and 1 SHARE chroma row 0 (`uv[0..8]`).
      fn in_sequence_flip_row1<R>(
        mut sink: MixedSinker<'_, $Marker, R>,
        y: &[u8],
        uv: &[u8],
        loc1: ChromaLocation,
        loc2: ChromaLocation,
      ) -> Result<(), MixedSinkerError> {
        sink.set_chroma_location(loc1.clone());
        PixelSink::begin_frame(&mut sink, 8, 8).unwrap();
        PixelSink::process(&mut sink, $Row::for_tests(&y[0..8], &uv[0..8], 0, M, FR)).unwrap();
        sink.set_chroma_location(loc2.clone());
        PixelSink::process(&mut sink, $Row::for_tests(&y[8..16], &uv[0..8], 1, M, FR))
      }

      #[test]
      #[cfg_attr(
        miri,
        ignore = "SIMD-dispatched row kernels use intrinsics unsupported by Miri"
      )]
      fn in_sequence_mid_frame_phase_change_rejected_across_tiers() {
        // The effective siting is frozen on the first output-bearing row; a later
        // IN-SEQUENCE row observing a different phase (after a mid-frame
        // `set_chroma_location`) passes the sequence preflight but must be rejected
        // with ChromaSitingChanged, uniformly across tiers — else the frame bins a
        // mixture of co-sited and centered chroma.
        let (y, u, v) = ramp(8, 8);
        let uv = interleave(&u, &v, 4, 4, $swap);
        // Both HORIZONTAL flips (Center⇄Left) plus the RFC #238 S4-C VERTICAL flip
        // Center⇄Bottom (same horizontal center, differing v=1 fold — caught only
        // by the vertical-phase freeze). BottomLeft adds BottomLeft⇄Left (vertical
        // fold flip on a co-sited h) and BottomLeft⇄Bottom (horizontal flip on a
        // shared v=1). The RFC #238 Top fold adds Center⇄Top (shared h center,
        // differing v=0 FORWARD fold — caught only by the Top-phase freeze),
        // TopLeft⇄Left (v=0 fold flip on a co-sited h), and Top⇄Bottom (opposite
        // vertical reaches on a shared h center). Each must reject the in-sequence
        // row 1.
        for (loc1, loc2) in [
          (ChromaLocation::Center, ChromaLocation::Left),
          (ChromaLocation::Left, ChromaLocation::Center),
          (ChromaLocation::Center, ChromaLocation::Bottom),
          (ChromaLocation::Bottom, ChromaLocation::Center),
          (ChromaLocation::Left, ChromaLocation::BottomLeft),
          (ChromaLocation::BottomLeft, ChromaLocation::Left),
          (ChromaLocation::BottomLeft, ChromaLocation::Bottom),
          (ChromaLocation::Bottom, ChromaLocation::BottomLeft),
          (ChromaLocation::Center, ChromaLocation::Top),
          (ChromaLocation::Top, ChromaLocation::Center),
          (ChromaLocation::TopLeft, ChromaLocation::Left),
          (ChromaLocation::Left, ChromaLocation::TopLeft),
          (ChromaLocation::Top, ChromaLocation::Bottom),
          (ChromaLocation::Bottom, ChromaLocation::Top),
        ] {
          // Native fast tier.
          let mut rgb = vec![0u8; 4 * 4 * 3];
          let sink =
            MixedSinker::<$Marker, AreaResampler>::with_resampler(8, 8, AreaResampler::to(4, 4))
              .unwrap()
              .with_native(true)
              .with_rgb(&mut rgb)
              .unwrap();
          let err = in_sequence_flip_row1(sink, &y, &uv, loc1.clone(), loc2.clone()).unwrap_err();
          assert!(
            matches!(err, MixedSinkerError::ChromaSitingChanged(_)),
            "native {loc1:?}->{loc2:?}: want ChromaSitingChanged, got {err:?}"
          );

          // Encoded row-stage RGB tier (`with_native(false)`).
          let mut rgb = vec![0u8; 4 * 4 * 3];
          let sink =
            MixedSinker::<$Marker, AreaResampler>::with_resampler(8, 8, AreaResampler::to(4, 4))
              .unwrap()
              .with_native(false)
              .with_rgb(&mut rgb)
              .unwrap();
          let err = in_sequence_flip_row1(sink, &y, &uv, loc1.clone(), loc2.clone()).unwrap_err();
          assert!(
            matches!(err, MixedSinkerError::ChromaSitingChanged(_)),
            "encoded-rgb {loc1:?}->{loc2:?}: want ChromaSitingChanged, got {err:?}"
          );

          // HSV-only row-stage tier (no separate hsv-direct join for NV 4:2:0).
          let (mut hh, mut ss, mut vv) = (vec![0u8; 4 * 4], vec![0u8; 4 * 4], vec![0u8; 4 * 4]);
          let sink =
            MixedSinker::<$Marker, AreaResampler>::with_resampler(8, 8, AreaResampler::to(4, 4))
              .unwrap()
              .with_native(false)
              .with_hsv(&mut hh, &mut ss, &mut vv)
              .unwrap();
          let err = in_sequence_flip_row1(sink, &y, &uv, loc1.clone(), loc2.clone()).unwrap_err();
          assert!(
            matches!(err, MixedSinkerError::ChromaSitingChanged(_)),
            "hsv {loc1:?}->{loc2:?}: want ChromaSitingChanged, got {err:?}"
          );

          // Filter tier (single-kernel Triangle FilteredResampler).
          let mut rgb = vec![0u8; 4 * 4 * 3];
          let sink = MixedSinker::<$Marker, FilteredResampler<Triangle>>::with_resampler(
            8,
            8,
            FilteredResampler::new(4, 4, Triangle),
          )
          .unwrap()
          .with_rgb(&mut rgb)
          .unwrap();
          let err = in_sequence_flip_row1(sink, &y, &uv, loc1.clone(), loc2.clone()).unwrap_err();
          assert!(
            matches!(err, MixedSinkerError::ChromaSitingChanged(_)),
            "filter {loc1:?}->{loc2:?}: want ChromaSitingChanged, got {err:?}"
          );
        }
      }

      #[test]
      #[cfg_attr(
        miri,
        ignore = "SIMD-dispatched row kernels use intrinsics unsupported by Miri"
      )]
      fn native_mid_frame_phase_change_rejection_keeps_the_stream_retryable() {
        // Advance rows 0,1 (Center), then flip siting mid-frame (Left): the
        // frozen-phase CHECK rejects it with ChromaSitingChanged at the choke
        // point; a mixed-phase frame is never emitted and the frame restarts
        // cleanly (a rejected row mutates no state).
        let (y, u, v) = ramp(8, 8);
        let uv = interleave(&u, &v, 4, 4, $swap);
        let mut rgb = vec![0u8; 4 * 4 * 3];
        let mut sink =
          MixedSinker::<$Marker, AreaResampler>::with_resampler(8, 8, AreaResampler::to(4, 4))
            .unwrap()
            .with_native(true)
            .with_chroma_location(ChromaLocation::Center)
            .with_rgb(&mut rgb)
            .unwrap();
        PixelSink::begin_frame(&mut sink, 8, 8).unwrap();
        for r in 0..2 {
          let row = $Row::for_tests(
            &y[r * 8..r * 8 + 8],
            &uv[(r / 2) * 8..(r / 2) * 8 + 8],
            r,
            M,
            FR,
          );
          PixelSink::process(&mut sink, row).unwrap();
        }
        sink.set_chroma_location(ChromaLocation::Left);
        let bad = $Row::for_tests(&y[2 * 8..3 * 8], &uv[8..16], 2, M, FR);
        let err = PixelSink::process(&mut sink, bad).unwrap_err();
        assert!(
          matches!(err, MixedSinkerError::ChromaSitingChanged(_)),
          "mid-frame siting change must be ChromaSitingChanged, got {err:?}"
        );
        // The rejected row mutated no stream state: begin_frame restarts cleanly
        // and a fresh frame at the new siting processes without error.
        PixelSink::begin_frame(&mut sink, 8, 8).unwrap();
        for r in 0..8 {
          let row = $Row::for_tests(
            &y[r * 8..r * 8 + 8],
            &uv[(r / 2) * 8..(r / 2) * 8 + 8],
            r,
            M,
            FR,
          );
          PixelSink::process(&mut sink, row).unwrap();
        }
      }
    }
  };
}

nv_resample_siting_tests!(nv12, Nv12, Nv12Frame, nv12_to, Nv12Row, false);
nv_resample_siting_tests!(nv21, Nv21, Nv21Frame, nv21_to, Nv21Row, true);