heddle-semantic 0.4.0

An AI-native version control system
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
// SPDX-License-Identifier: Apache-2.0
//! Per-item 3-way merge + file reconstruction.
//!
//! Given three [`FileSegments`] (base / ours / theirs), produce a merged file
//! by:
//!
//! 1. Resolving each *item* (matched by [`ItemKey`]) via 3-way merge of its
//!    bytes, falling through to `heddle-merge::text_hunk_merge` when both
//!    sides modify the same item.
//! 2. Stitching the resolved items back together in *base order*, with
//!    one-side-only added items spliced in at their natural position relative
//!    to neighbours.
//! 3. Weaving the inter-item segments back between the items at their
//!    original positions. Each output gap is the per-segment 3-way merge of
//!    the corresponding source segments on the sides that have the trailing
//!    item; added items bring their adding-side segment with them.

use std::collections::{BTreeMap, BTreeSet};

use merge::{ConflictMarkers, MergeOutcome, text_hunk_merge_with_markers};

use super::items::{FileSegments, Item, ItemKey, ItemKind, inter_ranges};

/// Three sides of the merge: `[base, ours, theirs]`. Each per-iteration
/// segment contribution is indexed by [`Side`] so emission tracking can
/// say "this side has already contributed range N — don't re-emit".
const N_SIDES: usize = 3;

/// Identity used for cross-side item matching. The bare [`ItemKey`]
/// would collapse repeated declarations that share the same key — e.g.
/// two top-level JavaScript `function foo() {}` statements with
/// identical signatures, or two Python module-level `def foo()` blocks
/// — and `BTreeMap<ItemKey, _>` would silently keep only the LAST
/// occurrence per side. `MatchKey` pairs the key with the item's
/// per-key occurrence index within its side (0 for the first, 1 for
/// the second, …) so each duplicate gets a distinct slot. Matching
/// across sides pairs same-key items positionally — base's first `foo`
/// pairs with ours's first `foo` pairs with theirs's first `foo`.
///
/// Positional occurrence governs only NON-`use` items. `use` items are
/// resolved as whole leaf-components by [`resolve_use_component`] (a set
/// comparison over every declaration each side contributes), so no `use`
/// item's content is ever decided by its occurrence index — the occurrence
/// slot survives for them only to anchor inter-item whitespace weaving
/// (heddle#468 r5: the duplicate-import class the positional path produced).
type MatchKey = (ItemKey, usize);

/// Stitch three sides together via recursive per-region tree merge.
///
/// The whole file is the outermost region; each matched container body is a
/// nested region merged the same way ([`merge_region`]). The trailing-newline
/// reconcile + outcome wrapping happen once, here, around the top-level merge.
pub(crate) fn reconstruct_merged_file(
    base: &str,
    ours: &str,
    theirs: &str,
    base_segments: &FileSegments,
    ours_segments: &FileSegments,
    theirs_segments: &FileSegments,
    markers: ConflictMarkers<'_>,
) -> MergeOutcome {
    // Whole-file source bundle: lets `resolve_item` slice per-item bytes AND
    // carries a whole-file `EolPolicy` used by the trailing newline path
    // (`reconcile_trailing_newline`) and as a fallback by the marker path
    // (`emit_addadd_conflict`) when the conflicting item bodies carry zero
    // EOL observations (Codex r8, cid 3256283857).
    let sides = SideSources::new(base, ours, theirs);

    let (mut output, total_conflicts) = merge_region(
        sides,
        &base_segments.items,
        &ours_segments.items,
        &theirs_segments.items,
        (0, base_segments.source_len),
        (0, ours_segments.source_len),
        (0, theirs_segments.source_len),
        markers,
    );

    reconcile_trailing_newline(&mut output, sides);

    if total_conflicts == 0 {
        MergeOutcome::Clean(output)
    } else {
        MergeOutcome::Conflicts {
            merged_bytes_with_markers: output,
            conflict_count: total_conflicts,
        }
    }
}

/// Merge one *region* — a list of sibling items occupying `[start, end)` on
/// each side — into a byte string + conflict count. Called on the whole file
/// at top level and, recursively (via [`resolve_container`]), on each matched
/// container body. Recursion depth is bounded by the container-nesting cap in
/// [`super::items`], so it cannot overflow the stack.
///
/// The algorithm is the heddle#68/#468 weave, generalized to a region:
/// resolve each item by `MatchKey`, compute an emit order, and weave the
/// per-side inter-item segments (including a container body's own braces,
/// which live in its region's preamble/postamble) back between the items.
#[allow(clippy::too_many_arguments)]
fn merge_region(
    sides: SideSources<'_>,
    base_items: &[Item],
    ours_items: &[Item],
    theirs_items: &[Item],
    base_bounds: (usize, usize),
    ours_bounds: (usize, usize),
    theirs_bounds: (usize, usize),
    markers: ConflictMarkers<'_>,
) -> (Vec<u8>, usize) {
    // Per-side match keys in source order — (ItemKey, discriminator) tuples.
    // Leaves use a positional occurrence index; container instances are
    // aligned to base by child-key overlap so a prepended/appended/reordered
    // same-name container keeps an identity distinct from the matched base
    // block (the heddle#484 r3 cross-side class). See [`build_aligned_match_keys`].
    let (base_mks, ours_mks, theirs_mks) =
        build_aligned_match_keys(base_items, ours_items, theirs_items, sides);

    let base_map: BTreeMap<MatchKey, &Item> = base_mks
        .iter()
        .zip(base_items.iter())
        .map(|(mk, i)| (mk.clone(), i))
        .collect();
    let ours_map: BTreeMap<MatchKey, &Item> = ours_mks
        .iter()
        .zip(ours_items.iter())
        .map(|(mk, i)| (mk.clone(), i))
        .collect();
    let theirs_map: BTreeMap<MatchKey, &Item> = theirs_mks
        .iter()
        .zip(theirs_items.iter())
        .map(|(mk, i)| (mk.clone(), i))
        .collect();

    let all_keys: BTreeSet<&MatchKey> = base_map
        .keys()
        .chain(ours_map.keys())
        .chain(theirs_map.keys())
        .collect();

    let mut resolved: BTreeMap<MatchKey, (Option<Vec<u8>>, usize)> = BTreeMap::new();
    let mut total_conflicts = 0usize;

    // Non-`use` items: per-item positional resolution, matched by
    // (key, occurrence). A matched *container* recurses into its body via
    // `resolve_node` → `resolve_container`; a leaf is a byte 3-way merge.
    // `use` items are skipped here — their content is NEVER decided by
    // positional occurrence index (the heddle#468 r5 bug class). They are
    // resolved below as whole leaf-components.
    for key in &all_keys {
        if key.0.kind == ItemKind::Use {
            continue;
        }
        let resolution = resolve_node(
            sides,
            base_map.get(*key).copied(),
            ours_map.get(*key).copied(),
            theirs_map.get(*key).copied(),
            markers,
        );
        total_conflicts += resolution.1;
        resolved.insert((*key).clone(), resolution);
    }

    // `use` items at THIS level: resolve each canonical leaf-component as ONE
    // set-valued unit (the heddle#468 r5 fix). After `canonicalize_use_keys`,
    // every declaration in a component shares one `ItemKey`; comparing full
    // component leaf-SETS rather than occurrence positions makes the
    // duplicate-import class impossible. Components are scoped to the region
    // because the `ItemKey` carries the enclosing scope, so uses in different
    // containers never group together.
    let mut use_components: BTreeMap<ItemKey, [Vec<&Item>; 3]> = BTreeMap::new();
    for (side, items) in [base_items, ours_items, theirs_items].iter().enumerate() {
        for item in *items {
            if item.key.kind == ItemKind::Use {
                use_components
                    .entry(item.key.clone())
                    .or_insert_with(|| [Vec::new(), Vec::new(), Vec::new()])[side]
                    .push(item);
            }
        }
    }
    for (key, [base_uses, ours_uses, theirs_uses]) in &use_components {
        let (bytes, conflicts) =
            resolve_use_component(sides, base_uses, ours_uses, theirs_uses, markers);
        total_conflicts += conflicts;
        resolved.insert((key.clone(), 0), (bytes, conflicts));
        // Higher-occurrence slots of this component exist only so the
        // inter-item segment weaver can place the surrounding whitespace;
        // they carry no item bytes (the verdict above is the whole unit).
        let slots = base_uses.len().max(ours_uses.len()).max(theirs_uses.len());
        for occ in 1..slots {
            resolved.insert((key.clone(), occ), (None, 0));
        }
    }

    let item_emit_order = compute_item_emit_order(&base_mks, &ours_mks, &theirs_mks, &all_keys);

    let side_idx_maps = [
        match_key_index(&base_mks),
        match_key_index(&ours_mks),
        match_key_index(&theirs_mks),
    ];
    let side_ranges = [
        inter_ranges(base_items, base_bounds.0, base_bounds.1),
        inter_ranges(ours_items, ours_bounds.0, ours_bounds.1),
        inter_ranges(theirs_items, theirs_bounds.0, theirs_bounds.1),
    ];
    let side_sources = [sides.base, sides.ours, sides.theirs];

    // Per-side set of inter-item range indices already emitted. A side's
    // range is contributed to at most one slot (see the Codex r2 P2 #2 /
    // P1 #2 duplication shapes the tracking prevents).
    let mut emitted: [BTreeSet<usize>; N_SIDES] =
        [BTreeSet::new(), BTreeSet::new(), BTreeSet::new()];

    let mut output: Vec<u8> = Vec::new();

    for key in &item_emit_order {
        let mut segs: [Option<&str>; N_SIDES] = [None, None, None];
        for s in 0..N_SIDES {
            // A side contributes the gap PRECEDING `key` only if it actually
            // has `key`. A side that lacks `key` (an item added on another
            // side, or one this side deleted) contributes nothing for this
            // slot — its surrounding content flows with its own items and its
            // trailing content stays in the postamble. Bridging a lacking
            // side to "the gap after its nearest prior item" used to pull the
            // postamble (or a mid-sequence gap) in early, duplicating it —
            // the heddle#484 Bug 1 (`// MARK` woven twice) / Bug 2 (module
            // duplicated) class.
            if let Some(&r) = side_idx_maps[s].get(key)
                && emitted[s].insert(r)
            {
                segs[s] = Some(inter_slice(side_sources[s], &side_ranges[s], r));
            }
        }
        let (seg_bytes, seg_conflicts) = merge_segment(segs[0], segs[1], segs[2], markers);
        output.extend_from_slice(&seg_bytes);
        total_conflicts += seg_conflicts;

        if let Some((Some(item_bytes), _)) = resolved.get(key) {
            output.extend_from_slice(item_bytes);
        }
    }

    // Postamble: each side's last range (a container body's closing brace +
    // indentation lives here), but only if it wasn't already pulled in as a
    // bridge above and adds bytes (avoids duplicating a trailing newline
    // already in the last item's bytes — the top-level P1 #2 shape).
    let mut post: [Option<&str>; N_SIDES] = [None, None, None];
    for s in 0..N_SIDES {
        let last = side_ranges[s].len() - 1;
        if emitted[s].insert(last) {
            post[s] = Some(inter_slice(side_sources[s], &side_ranges[s], last));
        }
    }
    let (post_bytes, post_conflicts) = merge_segment(post[0], post[1], post[2], markers);
    if !post_bytes.is_empty() {
        output.extend_from_slice(&post_bytes);
    }
    total_conflicts += post_conflicts;

    (output, total_conflicts)
}

/// Walk a side's items in source order and tag each with its per-key
/// occurrence index — 0 for the first item with that key, 1 for the second,
/// and so on. Length matches `items.len()`.
fn build_match_keys(items: &[Item]) -> Vec<MatchKey> {
    let mut counters: BTreeMap<ItemKey, usize> = BTreeMap::new();
    items
        .iter()
        .map(|item| {
            let n = counters.entry(item.key.clone()).or_insert(0);
            let occurrence = *n;
            *n += 1;
            (item.key.clone(), occurrence)
        })
        .collect()
}

/// Immediate child-key set of a container item (empty for leaves). Used to
/// align container instances across sides by content overlap.
fn child_key_set(item: &Item) -> BTreeSet<&ItemKey> {
    match &item.body {
        Some(body) => body.items.iter().map(|c| &c.key).collect(),
        None => BTreeSet::new(),
    }
}

/// Build the three sides' `MatchKey` lists with cross-side-consistent
/// discriminators.
///
/// Leaves get a per-side positional occurrence index — base's first `foo`
/// pairs with ours's first `foo`, the existing heddle#68 scheme. **Container
/// instances** (same `(kind, name, scope)`, multiple blocks) anchor to base in
/// two passes:
///
/// 1. **Content-overlap (primary).** Each ours/theirs block is paired with the
///    unused base block whose immediate child-keys it most overlaps (overlap
///    `> 0`), inheriting that base block's discriminator. Positional occurrence
///    alone mis-pairs a prepended `impl Foo` (occurrence 0) with base's
///    `impl Foo` (occurrence 0) — the r3 collapse; overlap alignment pairs by
///    what the block actually contains, so identity survives reordering.
/// 2. **Header-anchored positional fallback (no content signal).** A block that
///    found no overlap — an EMPTY container, or one whose children were fully
///    replaced — aligns to the UNUSED base block of its key whose *header*
///    bytes (`[start_byte, body.inner_start)`, which absorb leading metadata:
///    attributes, decorators, and separator comments — see
///    [`super::items::leading_metadata_start`]) byte-match it; failing a header
///    match, the next unused base block in source order. Only when no unused
///    base block of that key remains is it a genuinely-new container → fresh
///    discriminator above the base range. The header anchor is what pins a
///    *surviving* zero-overlap block to its TRUE base occurrence even when an
///    earlier same-key block was deleted on that side: "next unused" alone
///    grabs the earliest free slot (0), so deleting the first block and editing
///    the second mis-mapped the survivor to slot 0, treated slot 1 as deleted,
///    and wove its separator/trivia onto the wrong block (heddle#490 r2 /
///    Codex P1). Without *any* fallback an empty same-key container minted
///    fresh, its base slot resolved as deleted, and a clean one-sided edit
///    corrupted (heddle#490 r1).
///
/// For leaves (no children) every overlap is 0 and the leaf path applies a
/// plain positional occurrence index directly.
fn build_aligned_match_keys(
    base: &[Item],
    ours: &[Item],
    theirs: &[Item],
    sides: SideSources<'_>,
) -> (Vec<MatchKey>, Vec<MatchKey>, Vec<MatchKey>) {
    // A key is a "container key" if any instance on any side carries a body.
    let mut container_keys: BTreeSet<ItemKey> = BTreeSet::new();
    for items in [base, ours, theirs] {
        for it in items {
            if it.body.is_some() {
                container_keys.insert(it.key.clone());
            }
        }
    }

    // Base is the anchor: plain positional occurrence (its discriminator for a
    // container is its source-order index within the key group).
    let base_mks = build_match_keys(base);

    let align = |side: &[Item], side_src: &str| -> Vec<MatchKey> {
        // base container instances per key: (base discriminator, &base item),
        // in base source order (so the positional fallback below scans the
        // *earliest* unused base slot first).
        let mut base_by_key: BTreeMap<&ItemKey, Vec<(usize, &Item)>> = BTreeMap::new();
        for (i, it) in base.iter().enumerate() {
            if container_keys.contains(&it.key) {
                base_by_key
                    .entry(&it.key)
                    .or_default()
                    .push((base_mks[i].1, it));
            }
        }
        let mut used: BTreeMap<&ItemKey, BTreeSet<usize>> = BTreeMap::new();
        let mut leaf_occ: BTreeMap<&ItemKey, usize> = BTreeMap::new();
        let mut fresh: BTreeMap<&ItemKey, usize> = BTreeMap::new();

        // Container discriminators are decided in TWO passes so that
        // content-overlap stays the PRIMARY signal: a zero-overlap (e.g. empty)
        // block must never greedily claim a base slot that a later
        // content-overlap block would match. `disc_of[pos]` holds the
        // pass-1 verdict per side position; `None` = deferred to pass 2.
        let mut disc_of: Vec<Option<usize>> = vec![None; side.len()];

        // Pass 1 — content-overlap alignment (PRIMARY). Each container claims
        // the unused base candidate of its key with the greatest immediate
        // child-key overlap, when that overlap is > 0. This is the heddle#484
        // r3 mechanism: a reordered/edited non-empty block matches by what it
        // actually contains.
        for (pos, it) in side.iter().enumerate() {
            if !container_keys.contains(&it.key) {
                continue;
            }
            let childset = child_key_set(it);
            let used_set = used.entry(&it.key).or_default();
            let mut best: Option<(usize, usize)> = None; // (overlap, base disc)
            if let Some(cands) = base_by_key.get(&it.key) {
                for (disc, bitem) in cands {
                    if used_set.contains(disc) {
                        continue;
                    }
                    let overlap = childset.intersection(&child_key_set(bitem)).count();
                    if overlap > 0 && best.is_none_or(|(o, _)| overlap > o) {
                        best = Some((overlap, *disc));
                    }
                }
            }
            if let Some((_, d)) = best {
                used_set.insert(d);
                disc_of[pos] = Some(d);
            }
        }

        // Pass 2a — header-anchored alignment for the no-content-signal case.
        // A container left unresolved by pass 1 (an EMPTY container, or one
        // whose children were fully replaced so it overlaps no base block)
        // claims the earliest UNUSED base candidate of its key whose HEADER
        // bytes byte-match it. The header is `[start_byte, body.inner_start)`,
        // which absorbs the block's leading metadata — attributes, decorators,
        // and separator comments (see [`super::items::leading_metadata_start`])
        // — so a SURVIVING block that kept its preceding comment re-anchors to
        // the exact base occurrence that comment belonged to. This pins the
        // survivor to its TRUE base slot even when an earlier same-key block was
        // deleted on this side, and even under a reorder: "next unused" alone
        // grabbed slot 0 and wove the deleted slot's separator onto the survivor
        // (heddle#490 r2). Header matching runs as a PRIORITY pass — before the
        // source-order scan below — so a newly-prepended block (no header match)
        // can't greedily steal the slot a surviving block needs.
        for (pos, it) in side.iter().enumerate() {
            if !container_keys.contains(&it.key) || disc_of[pos].is_some() {
                continue;
            }
            let it_header = align_header_bytes(it, side_src);
            let used_set = used.entry(&it.key).or_default();
            if let Some(cands) = base_by_key.get(&it.key)
                && let Some((d, _)) = cands.iter().find(|(d, b)| {
                    !used_set.contains(d) && align_header_bytes(b, sides.base) == it_header
                })
            {
                used_set.insert(*d);
                disc_of[pos] = Some(*d);
            }
        }

        // Pass 2b — source-order fallback + leaves. Any container still
        // unresolved (indistinguishable headers, or no header match) claims the
        // next UNUSED base candidate of its key in base source order; only when
        // none remains is it a genuinely-new container beyond base's count →
        // fresh discriminator above the base range. When headers are
        // indistinguishable, either slot choice is byte-equivalent (identical
        // separators), so no corruption can result. This stays a narrow fallback
        // for the zero-overlap case, NOT a return to the per-everything ordinal
        // model. Leaves get a plain positional occurrence index.
        let mut out = Vec::with_capacity(side.len());
        for (pos, it) in side.iter().enumerate() {
            if container_keys.contains(&it.key) {
                let disc = if let Some(d) = disc_of[pos] {
                    d
                } else {
                    let used_set = used.entry(&it.key).or_default();
                    let next_unused = base_by_key.get(&it.key).and_then(|cands| {
                        cands
                            .iter()
                            .map(|(d, _)| *d)
                            .find(|d| !used_set.contains(d))
                    });
                    if let Some(d) = next_unused {
                        used_set.insert(d);
                        d
                    } else {
                        let base_count = base_by_key.get(&it.key).map_or(0, Vec::len);
                        let f = fresh.entry(&it.key).or_insert(0);
                        let d = base_count + *f;
                        *f += 1;
                        d
                    }
                };
                out.push((it.key.clone(), disc));
            } else {
                let occ = leaf_occ.entry(&it.key).or_insert(0);
                let d = *occ;
                *occ += 1;
                out.push((it.key.clone(), d));
            }
        }
        out
    };

    let ours_mks = align(ours, sides.ours);
    let theirs_mks = align(theirs, sides.theirs);
    (base_mks, ours_mks, theirs_mks)
}

fn match_key_index(mks: &[MatchKey]) -> BTreeMap<MatchKey, usize> {
    mks.iter()
        .enumerate()
        .map(|(i, mk)| (mk.clone(), i))
        .collect()
}

fn inter_slice<'a>(source: &'a str, ranges: &[(usize, usize)], idx: usize) -> &'a str {
    let (start, end) = ranges[idx];
    &source[start..end]
}

/// 3-way merge a single inter-item segment. Handles "side doesn't have
/// this segment" by promoting the present side(s).
fn merge_segment(
    base: Option<&str>,
    ours: Option<&str>,
    theirs: Option<&str>,
    markers: ConflictMarkers<'_>,
) -> (Vec<u8>, usize) {
    match (base, ours, theirs) {
        (None, None, None) => (Vec::new(), 0),
        (Some(b), Some(o), Some(t)) => materialize_segment(
            text_hunk_merge_with_markers(b.as_bytes(), o.as_bytes(), t.as_bytes(), markers),
            b,
        ),
        // Base has it; one side doesn't carry this segment (item missing
        // there). Treat the missing side as "no change" against base.
        (Some(b), Some(o), None) => materialize_segment(
            text_hunk_merge_with_markers(b.as_bytes(), o.as_bytes(), b.as_bytes(), markers),
            b,
        ),
        (Some(b), None, Some(t)) => materialize_segment(
            text_hunk_merge_with_markers(b.as_bytes(), b.as_bytes(), t.as_bytes(), markers),
            b,
        ),
        (Some(b), None, None) => (b.as_bytes().to_vec(), 0),
        // Added item — only the adding side(s) contribute a segment.
        (None, Some(o), Some(t)) => {
            if o == t {
                (o.as_bytes().to_vec(), 0)
            } else {
                materialize_segment(
                    text_hunk_merge_with_markers(&[], o.as_bytes(), t.as_bytes(), markers),
                    "",
                )
            }
        }
        (None, Some(o), None) => (o.as_bytes().to_vec(), 0),
        (None, None, Some(t)) => (t.as_bytes().to_vec(), 0),
    }
}

fn materialize_segment(outcome: MergeOutcome, fallback: &str) -> (Vec<u8>, usize) {
    match outcome {
        MergeOutcome::Clean(bytes) => (bytes, 0),
        MergeOutcome::Conflicts {
            merged_bytes_with_markers,
            conflict_count,
        } => (merged_bytes_with_markers, conflict_count),
        // Binary / DeleteVsModify shouldn't fire on a text subset, but
        // carry through with base bytes rather than nothing.
        MergeOutcome::Binary | MergeOutcome::DeleteVsModify => (fallback.as_bytes().to_vec(), 0),
    }
}

/// Whole-file source bundle threaded through item resolution. Lets
/// `resolve_item` slice item bytes per side AND carries the
/// whole-file `EolPolicy` used by `reconcile_trailing_newline` and as
/// the zero-observation fallback by `emit_addadd_conflict`. The
/// marker path's primary policy is per-item — see
/// `emit_addadd_conflict` — but reuses this whole-file policy when
/// the conflicting items contribute no `\n` of their own
/// (single-line items, Codex r8 cid 3256283857).
#[derive(Clone, Copy)]
struct SideSources<'a> {
    base: &'a str,
    ours: &'a str,
    theirs: &'a str,
    eol_policy: EolPolicy,
}

impl<'a> SideSources<'a> {
    fn new(base: &'a str, ours: &'a str, theirs: &'a str) -> Self {
        let eol_policy = EolPolicy::detect(&[base.as_bytes(), ours.as_bytes(), theirs.as_bytes()]);
        SideSources {
            base,
            ours,
            theirs,
            eol_policy,
        }
    }
}

/// Dominant line-ending across a set of byte samples. Built via
/// [`EolPolicy::detect`] once over the whole-file sources (see
/// [`SideSources::new`]) and reused everywhere downstream that needs
/// to emit a newline. Counts `\r\n` occurrences vs bare `\n` (LF not
/// preceded by CR); the strict majority wins, and ties fall back to
/// the first sample's own dominant style — by convention callers pass
/// `base` first — then to LF.
///
/// Earlier revisions returned CRLF as soon as ANY sample contained
/// one `\r\n`; that wrongly flipped a majority-LF file to CRLF when a
/// single side happened to be CRLF (Codex r7 P2, cid 3256225712).
/// Majority voting respects the file's actual style without
/// overweighting a single divergent side.
#[derive(Clone, Copy)]
struct EolPolicy {
    crlf: usize,
    lf: usize,
    first_crlf: usize,
    first_lf: usize,
}

impl EolPolicy {
    fn detect(samples: &[&[u8]]) -> Self {
        let mut crlf = 0usize;
        let mut lf = 0usize;
        let mut first_crlf = 0usize;
        let mut first_lf = 0usize;
        for (i, s) in samples.iter().enumerate() {
            let (c, l) = count_eols(s);
            crlf += c;
            lf += l;
            if i == 0 {
                first_crlf = c;
                first_lf = l;
            }
        }
        EolPolicy {
            crlf,
            lf,
            first_crlf,
            first_lf,
        }
    }

    fn eol(self) -> &'static [u8] {
        if self.crlf > self.lf {
            return b"\r\n";
        }
        if self.lf > self.crlf {
            return b"\n";
        }
        if self.first_crlf > self.first_lf {
            return b"\r\n";
        }
        b"\n"
    }
}

/// Whole byte span of an item on a side.
fn whole_bytes<'a>(item: &Item, src: &'a str) -> &'a [u8] {
    &src.as_bytes()[item.start_byte..item.end_byte]
}

/// Header bytes of a container item: `[start_byte, body.inner_start)` — e.g.
/// `impl Foo ` (everything before the body delimiter).
fn header_bytes<'a>(item: &Item, src: &'a str) -> &'a [u8] {
    let inner_start = item.body.as_ref().expect("container").inner_start;
    &src.as_bytes()[item.start_byte..inner_start]
}

/// Header bytes used to anchor a zero-overlap container to its base occurrence
/// in pass 2 of [`build_aligned_match_keys`]. For a real container this is
/// [`header_bytes`] — everything before the body delimiter, which absorbs the
/// item's leading metadata (attributes / decorators / separator comments) and
/// so identifies which base block a surviving block came from. For an opaque
/// (too-deeply-nested) container carried as a `body: None` leaf, fall back to
/// the whole byte span so the comparison stays total instead of panicking on
/// the missing body.
fn align_header_bytes<'a>(item: &Item, src: &'a str) -> &'a [u8] {
    if item.body.is_some() {
        header_bytes(item, src)
    } else {
        whole_bytes(item, src)
    }
}

/// Footer bytes of a container item: `[body.inner_end, end_byte)` — usually
/// empty (the body node ends at the item end).
fn footer_bytes<'a>(item: &Item, src: &'a str) -> &'a [u8] {
    let inner_end = item.body.as_ref().expect("container").inner_end;
    &src.as_bytes()[inner_end..item.end_byte]
}

/// Dispatch a match-key resolution to the container or leaf path.
///
/// The structural container path ([`resolve_container`] → [`merge_container_3way`])
/// is entered ONLY when *every* side that carries this matched item is a real
/// container WITH a body. That is a hard precondition: `merge_container_3way`
/// (and the `header_bytes` / `footer_bytes` it calls) read `body.inner_start` /
/// `body.inner_end`, which a leaf does not have.
///
/// A key is *usually* consistently a container or a leaf across sides, but it
/// can be MIXED: the same `ItemKey` (kind, name, scope) names a container on one
/// side and a leaf on another, because two distinct syntactic forms share a key.
/// Concretely — a Python `class C` (container) wrapped in a decorator becomes a
/// `decorated_definition` whose `container_body` is forced to `None` (a leaf)
/// while keeping the inner class's key; a Rust `mod foo { … }` (container)
/// rewritten to `mod foo;` (a leaf, no body) keeps the same module key. Such a
/// kind-mismatch cannot merge on the tree, so it routes to a whole-item 3-way
/// text merge ([`resolve_item`], over each side's full `[start_byte, end_byte)`
/// span) — clean when the edits are disjoint, a normal conflict when they
/// overlap. Making body-presence a CHECKED precondition of structural entry —
/// rather than an `unwrap` deep inside `merge_container_3way` — closes the whole
/// "structural-merge precondition violated" panic class (heddle#490 r3 / Codex
/// P2).
fn resolve_node(
    sides: SideSources<'_>,
    base_item: Option<&Item>,
    ours_item: Option<&Item>,
    theirs_item: Option<&Item>,
    markers: ConflictMarkers<'_>,
) -> (Option<Vec<u8>>, usize) {
    let mut present = [base_item, ours_item, theirs_item]
        .into_iter()
        .flatten()
        .peekable();
    let all_containers = present.peek().is_some() && present.all(|i| i.body.is_some());
    if all_containers {
        resolve_container(sides, base_item, ours_item, theirs_item, markers)
    } else {
        resolve_item(sides, base_item, ours_item, theirs_item, markers)
    }
}

/// Resolve a matched *container* by merging on the tree: header + recursively
/// merged body + footer. The byte-identical fast paths (unchanged side defers,
/// both-sides-identical dedup, clean delete) avoid recursion and keep the
/// container's bytes verbatim; only a genuine cross-side divergence recurses
/// into the body. Because two distinct same-name containers are distinct
/// `MatchKey`s (by occurrence), an added/prepended/appended container is never
/// conflated with a matched one — the heddle#484 collapse class is impossible
/// by construction.
fn resolve_container(
    sides: SideSources<'_>,
    base_item: Option<&Item>,
    ours_item: Option<&Item>,
    theirs_item: Option<&Item>,
    markers: ConflictMarkers<'_>,
) -> (Option<Vec<u8>>, usize) {
    match (base_item, ours_item, theirs_item) {
        (None, None, None) => (None, 0),
        // Added on one side only — take it verbatim.
        (None, Some(o), None) => (Some(whole_bytes(o, sides.ours).to_vec()), 0),
        (None, None, Some(t)) => (Some(whole_bytes(t, sides.theirs).to_vec()), 0),
        // Both sides added a same-key container with NO base to anchor against.
        // A recursive body merge is only sound when a BASE container exists to
        // diff each side against; with no anchor, recursing mis-weaves the
        // header/delimiters — it attaches each side's opening `{`/preamble to
        // that side's first added child, emitting `{` before BOTH children and
        // duplicating the delimiter, and the empty-base safety fallback can then
        // duplicate the whole module (heddle#490 r5). So compare the two added
        // containers as WHOLE units (header + body + footer): byte-identical →
        // both sides added the same thing, take one copy; any divergence → an
        // irreconcilable whole-container conflict. This single rule subsumes the
        // r4 divergent-header case (different header ⇒ different whole content ⇒
        // conflict). The recursive structural merge below stays for the
        // base-anchored case, where diffing against the base makes it sound.
        (None, Some(o), Some(t)) => {
            let ow = whole_bytes(o, sides.ours);
            let tw = whole_bytes(t, sides.theirs);
            if ow == tw {
                (Some(ow.to_vec()), 0)
            } else {
                (Some(emit_addadd_conflict(ow, tw, markers, sides)), 1)
            }
        }
        // Existed in base, removed on both sides → clean delete.
        (Some(_), None, None) => (None, 0),
        // Modify/delete: clean delete when the modifying side preserved base;
        // conflict (whole container) otherwise.
        (Some(b), Some(o), None) => {
            let bw = whole_bytes(b, sides.base);
            let ow = whole_bytes(o, sides.ours);
            if bw == ow {
                (None, 0)
            } else {
                materialize_outcome(text_hunk_merge_with_markers(bw, ow, &[], markers))
            }
        }
        (Some(b), None, Some(t)) => {
            let bw = whole_bytes(b, sides.base);
            let tw = whole_bytes(t, sides.theirs);
            if bw == tw {
                (None, 0)
            } else {
                materialize_outcome(text_hunk_merge_with_markers(bw, &[], tw, markers))
            }
        }
        // 3-way modify. Unchanged side defers; both-identical dedups;
        // otherwise merge header + body + footer structurally.
        (Some(b), Some(o), Some(t)) => {
            let bw = whole_bytes(b, sides.base);
            let ow = whole_bytes(o, sides.ours);
            let tw = whole_bytes(t, sides.theirs);
            if ow == bw {
                (Some(tw.to_vec()), 0)
            } else if tw == bw || ow == tw {
                (Some(ow.to_vec()), 0)
            } else {
                merge_container_3way(sides, b, o, t, markers)
            }
        }
    }
}

/// Merge a container that genuinely diverged across sides *against a base
/// anchor*: 3-way merge its header text, recurse [`merge_region`] over its body
/// children (base body vs ours vs theirs), 3-way merge its footer text, and
/// concatenate. Entered ONLY for the base-anchored 3-way-modify case — a real
/// base container exists to diff each side against, which is what makes the
/// recursive body merge sound. The no-base add/add case never reaches here: it
/// is resolved by a whole-container comparison in [`resolve_container`] (no
/// recursion without an anchor — heddle#490 r5).
fn merge_container_3way(
    sides: SideSources<'_>,
    base: &Item,
    ours: &Item,
    theirs: &Item,
    markers: ConflictMarkers<'_>,
) -> (Option<Vec<u8>>, usize) {
    // Precondition (guaranteed by `resolve_node`'s `all_containers` gate): every
    // participating side is a container WITH a body. The `header_bytes` /
    // `footer_bytes` / `body.as_ref()` reads below depend on it; a mixed
    // container/leaf key never reaches here (it routes to whole-item text merge).
    debug_assert!(
        base.body.is_some() && ours.body.is_some() && theirs.body.is_some(),
        "merge_container_3way entered with a leaf side — structural precondition violated"
    );
    let (header, hc) = merge3_text(
        header_bytes(base, sides.base),
        header_bytes(ours, sides.ours),
        header_bytes(theirs, sides.theirs),
        markers,
    );

    let bb = base.body.as_ref().expect("container");
    let ob = ours.body.as_ref().expect("container");
    let tb = theirs.body.as_ref().expect("container");

    // The body's STRUCTURAL opening/closing delimiters (`{` / `}` for brace
    // languages; empty for delimiter-less Python `block`s) are merged ONCE
    // here and emitted around the woven children — never folded into the child
    // weave. Folding them in let an empty base body, whose only inter-item
    // range is the whole `{}`, re-emit each side's opening `{` in that side's
    // first added-child slot: `mod foo {}` + ours adds `fn a` + theirs adds
    // `fn b` produced two `{` (heddle#490 r6). With the delimiters peeled off,
    // `merge_region` only ever weaves the inter-child content between
    // `content_start` and `content_end`, so the opening delimiter is emitted
    // exactly once for the region.
    let (open, oc) = merge3_text(
        &sides.base.as_bytes()[bb.inner_start..bb.content_start],
        &sides.ours.as_bytes()[ob.inner_start..ob.content_start],
        &sides.theirs.as_bytes()[tb.inner_start..tb.content_start],
        markers,
    );
    let (body, bc) = merge_region(
        sides,
        &bb.items,
        &ob.items,
        &tb.items,
        (bb.content_start, bb.content_end),
        (ob.content_start, ob.content_end),
        (tb.content_start, tb.content_end),
        markers,
    );
    let (close, cc) = merge3_text(
        &sides.base.as_bytes()[bb.content_end..bb.inner_end],
        &sides.ours.as_bytes()[ob.content_end..ob.inner_end],
        &sides.theirs.as_bytes()[tb.content_end..tb.inner_end],
        markers,
    );

    let (footer, fc) = merge3_text(
        footer_bytes(base, sides.base),
        footer_bytes(ours, sides.ours),
        footer_bytes(theirs, sides.theirs),
        markers,
    );

    let mut out = header;
    out.extend_from_slice(&open);
    out.extend_from_slice(&body);
    out.extend_from_slice(&close);
    out.extend_from_slice(&footer);
    (Some(out), hc + oc + bc + cc + fc)
}

/// 3-way merge a slice of bytes (a base-anchored container header/footer).
/// Equal-sides dedup and unchanged-side defer short-circuit; otherwise fall
/// through to the text hunk merge. Only ever called with a real base anchor —
/// the no-base add/add case is conflicted as a whole container upstream in
/// [`resolve_container`] and never recurses into header/footer here.
fn merge3_text(
    base: &[u8],
    ours: &[u8],
    theirs: &[u8],
    markers: ConflictMarkers<'_>,
) -> (Vec<u8>, usize) {
    if ours == theirs {
        return (ours.to_vec(), 0);
    }
    if ours == base {
        return (theirs.to_vec(), 0);
    }
    if theirs == base {
        return (ours.to_vec(), 0);
    }
    materialize_segment(
        text_hunk_merge_with_markers(base, ours, theirs, markers),
        std::str::from_utf8(base).unwrap_or(""),
    )
}

/// Resolve a single item's 3-way merge. Returns `(Some(bytes), n_conflicts)`
/// when the item survives, `(None, n_conflicts)` when both sides removed
/// it.
fn resolve_item(
    sides: SideSources<'_>,
    base_item: Option<&Item>,
    ours_item: Option<&Item>,
    theirs_item: Option<&Item>,
    markers: ConflictMarkers<'_>,
) -> (Option<Vec<u8>>, usize) {
    let base_bytes = base_item.map(|i| &sides.base.as_bytes()[i.start_byte..i.end_byte]);
    let ours_bytes = ours_item.map(|i| &sides.ours.as_bytes()[i.start_byte..i.end_byte]);
    let theirs_bytes = theirs_item.map(|i| &sides.theirs.as_bytes()[i.start_byte..i.end_byte]);

    match (base_bytes, ours_bytes, theirs_bytes) {
        (None, None, None) => (None, 0),
        // Added on one side only — take it.
        (None, Some(o), None) => (Some(o.to_vec()), 0),
        (None, None, Some(t)) => (Some(t.to_vec()), 0),
        // Both sides added the same item. Clean only if bytes match.
        // For diverging add-add we MUST surface a conflict directly rather
        // than delegating to text_hunk_merge — the engine's "same-anchor
        // insertion" path concatenates both insertions, which produces a
        // syntactically invalid file when both sides added a function /
        // method with the same name. heddle#68 calls this out as a conflict.
        (None, Some(o), Some(t)) => {
            // The ONLY clean add/add is byte-identical. `use` items never
            // reach this arm — they are resolved as whole leaf-components by
            // `resolve_use_component` (set-valued, not positional). This arm
            // now governs only non-`use` items (e.g. two top-level functions
            // with the same name added on both sides): byte-identical → dedup,
            // anything else → conflict (heddle#68).
            if o == t {
                (Some(o.to_vec()), 0)
            } else {
                (Some(emit_addadd_conflict(o, t, markers, sides)), 1)
            }
        }
        // Existed in base, removed on both sides → clean delete.
        (Some(_), None, None) => (None, 0),
        // Modify/delete: clean delete when the modifying side preserved
        // base; conflict otherwise.
        (Some(b), Some(o), None) => {
            if b == o {
                (None, 0)
            } else {
                // Encode the modify-vs-delete conflict as a synthetic
                // 3-way merge where the deleting side is empty.
                let outcome = text_hunk_merge_with_markers(b, o, &[], markers);
                materialize_outcome(outcome)
            }
        }
        (Some(b), None, Some(t)) => {
            if b == t {
                (None, 0)
            } else {
                let outcome = text_hunk_merge_with_markers(b, &[], t, markers);
                materialize_outcome(outcome)
            }
        }
        // 3-way modify.
        (Some(b), Some(o), Some(t)) => {
            if o == b {
                (Some(t.to_vec()), 0)
            } else if t == b || o == t {
                (Some(o.to_vec()), 0)
            } else {
                let outcome = text_hunk_merge_with_markers(b, o, t, markers);
                materialize_outcome(outcome)
            }
        }
    }
}

fn materialize_outcome(outcome: MergeOutcome) -> (Option<Vec<u8>>, usize) {
    match outcome {
        MergeOutcome::Clean(bytes) => (Some(bytes), 0),
        MergeOutcome::Conflicts {
            merged_bytes_with_markers,
            conflict_count,
        } => (Some(merged_bytes_with_markers), conflict_count),
        // Binary / DeleteVsModify shouldn't fire on UTF-8 source we already
        // parsed, but carry through safely.
        MergeOutcome::Binary | MergeOutcome::DeleteVsModify => (None, 1),
    }
}

/// Resolve one canonical leaf-component of `use` items as a single
/// set-valued unit. `base_items` / `ours_items` / `theirs_items` are every
/// declaration each side contributes to the component, in source order;
/// any of them may be empty (component absent on that side) or hold more
/// than one declaration (the heddle#468 r5 base-widened-grouped shape).
///
/// The component's text on a side is the byte-exact concatenation of its
/// declarations (one EOL between consecutive lines). The 3-way verdict is
/// taken over those WHOLE-component texts — never per declaration by
/// positional occurrence — and reduces to exactly three outcomes:
///
/// * **one side left the component byte-identical to base** → take the
///   other side (the standard "unchanged side defers" rule, generalized
///   from [`resolve_item`]'s 3-way-modify arm to the set);
/// * **both sides produced byte-identical text** → dedup to one copy;
/// * **everything else** — a widened/regrouped base item, divergent
///   additions, alias / `cfg` / visibility drift, or any multi-occurrence
///   ambiguity within the component → **conflict** the whole component as
///   one `<<<<<<< / ======= / >>>>>>>` block.
///
/// Because the comparison is over complete leaf-SETS rather than
/// occurrence positions, the r5 class (base `use a::Bar;`; ours adds a
/// separate `use a::Baz;`; theirs widens to `use a::{Bar, Baz};`) lands in
/// the conflict outcome instead of silently emitting both `{Bar, Baz}` and
/// `Baz` — a duplicate import (Rust E0252). No future regroup / widen /
/// multi-occurrence shape can drip the same way.
fn resolve_use_component(
    sides: SideSources<'_>,
    base_items: &[&Item],
    ours_items: &[&Item],
    theirs_items: &[&Item],
    markers: ConflictMarkers<'_>,
) -> (Option<Vec<u8>>, usize) {
    let eol = sides.eol_policy.eol();
    let base_bytes = join_component(base_items, sides.base, eol);
    let ours_bytes = join_component(ours_items, sides.ours, eol);
    let theirs_bytes = join_component(theirs_items, sides.theirs, eol);

    let non_empty = |v: Vec<u8>| if v.is_empty() { None } else { Some(v) };

    if ours_bytes == base_bytes {
        // ours left the component untouched → take theirs (which may be a
        // clean delete when theirs is empty).
        return (non_empty(theirs_bytes), 0);
    }
    if theirs_bytes == base_bytes || ours_bytes == theirs_bytes {
        // theirs left it untouched → take ours; or both sides made the
        // byte-identical change → dedup to a single copy.
        return (non_empty(ours_bytes), 0);
    }
    // Both sides changed the component, differently. Conflict the whole
    // unit — see the outcome list above.
    (
        Some(emit_addadd_conflict(
            &ours_bytes,
            &theirs_bytes,
            markers,
            sides,
        )),
        1,
    )
}

/// Concatenate a `use` component's declarations into one byte-exact text,
/// separating consecutive lines with `eol`. A single declaration yields its
/// own bytes verbatim (so single-occurrence components compare and emit
/// exactly as [`resolve_item`] did before the set-valued path existed).
fn join_component(items: &[&Item], source: &str, eol: &[u8]) -> Vec<u8> {
    let mut out = Vec::new();
    for (i, item) in items.iter().enumerate() {
        if i > 0 {
            out.extend_from_slice(eol);
        }
        out.extend_from_slice(&source.as_bytes()[item.start_byte..item.end_byte]);
    }
    out
}

/// Determine the order items should appear in the output. Strategy:
///
/// 1. Start with base's order.
/// 2. Items missing in base (added on a side) are spliced in after their
///    "left neighbour" in their originating side. The left neighbour is
///    found by walking left from the added item until a key common with
///    base is reached.
/// 3. If both ours and theirs added the same key independently and base
///    doesn't have it, we use ours's neighbour; ties go to ours.
fn compute_item_emit_order(
    base_mks: &[MatchKey],
    ours_mks: &[MatchKey],
    theirs_mks: &[MatchKey],
    all_keys: &BTreeSet<&MatchKey>,
) -> Vec<MatchKey> {
    let mut order: Vec<MatchKey> = base_mks.to_vec();

    for side_mks in [ours_mks, theirs_mks] {
        for (idx, key) in side_mks.iter().enumerate() {
            if order.contains(key) {
                continue;
            }
            let mut insert_at = 0usize;
            for i in (0..idx).rev() {
                if let Some(pos) = order.iter().position(|k| *k == side_mks[i]) {
                    insert_at = pos + 1;
                    break;
                }
            }
            order.insert(insert_at, key.clone());
        }
    }

    order.into_iter().filter(|k| all_keys.contains(k)).collect()
}

/// Append `eol` to `out` unless `out` already ends with a `\n` (which
/// covers both LF and CRLF terminations). Used to keep conflict-marker
/// blocks well-formed when a body doesn't end with its own newline.
fn ensure_trailing_newline(out: &mut Vec<u8>, eol: &[u8]) {
    if !out.is_empty() && *out.last().unwrap() != b'\n' {
        out.extend_from_slice(eol);
    }
}

/// Count (`\r\n`, bare `\n`) occurrences in `s`. A `\n` is "bare" iff
/// it is not preceded by `\r`.
fn count_eols(s: &[u8]) -> (usize, usize) {
    let mut crlf = 0usize;
    let mut lf = 0usize;
    let mut prev = 0u8;
    for &b in s {
        if b == b'\n' {
            if prev == b'\r' {
                crlf += 1;
            } else {
                lf += 1;
            }
        }
        prev = b;
    }
    (crlf, lf)
}

/// Match the trailing-newline state of `output` to the majority of the
/// three input sides. `text_hunk_merge` preserves whatever its line
/// splitter sees on the last line; the semantic path used to force a
/// trailing `\n` unconditionally, which dirtied files that ended
/// without one on every side (Codex r3 P2 #2).
///
/// Rule: count how many of `base`, `ours`, `theirs` end with `\n`. If
/// the majority do, ensure output ends with `\n`; otherwise strip any
/// `\n` we may have inherited from a single side's content. Empty
/// inputs are not counted (they have no opinion on trailing-newline
/// state).
///
/// CRLF is treated as a single unit on BOTH the pop and push paths:
/// when popping a trailing `\n`, an immediately-preceding `\r` is
/// popped along with it (Codex r5 P1 #4); when pushing a trailing
/// newline back, the dominant EOL of the inputs is pushed so a
/// CRLF-canonical file doesn't gain a bare LF (heddle#114 r7 self-
/// audit prediction P1, same hazard class as the r6 P2 #1 markers
/// finding).
fn reconcile_trailing_newline(out: &mut Vec<u8>, sides: SideSources<'_>) {
    if out.is_empty() {
        return;
    }
    let want_newline = majority_ends_with_newline(sides.base, sides.ours, sides.theirs);
    let has_newline = *out.last().unwrap() == b'\n';
    match (want_newline, has_newline) {
        (true, false) => {
            out.extend_from_slice(sides.eol_policy.eol());
        }
        (false, true) => {
            out.pop();
            if out.last() == Some(&b'\r') {
                out.pop();
            }
        }
        _ => {}
    }
}

fn majority_ends_with_newline(base: &str, ours: &str, theirs: &str) -> bool {
    let mut with = 0u8;
    let mut total = 0u8;
    for s in [base, ours, theirs] {
        if s.is_empty() {
            continue;
        }
        total += 1;
        if s.as_bytes().last() == Some(&b'\n') {
            with += 1;
        }
    }
    // Default to "yes" when nothing has an opinion (all sides empty —
    // unreachable in practice since we'd have returned Clean(empty)
    // before reconstruction), and require strict majority otherwise.
    total == 0 || with * 2 > total
}

/// Emit a `<<<<<<< / ======= / >>>>>>>` conflict block wrapping two
/// insertion bodies. Mirrors the marker shape `heddle-merge::markers`
/// produces so external validators (heddle#78) and IDE conflict tools
/// parse it identically.
///
/// Line endings on the marker lines come from a per-item [`EolPolicy`]
/// computed over the two conflicting item bodies, NOT the whole-file
/// policy carried by `sides`. The markers and the body they bracket
/// are derived from the same sample, so the r8 invariant (markers +
/// body cannot disagree) holds — but they now reflect the item's own
/// EOL discipline rather than the surrounding file. In a mixed-EOL
/// file where the LF context outnumbers a CRLF item, the whole-file
/// policy would vote LF and wrap a CRLF body with bare-LF markers,
/// reintroducing the mixed-EOL hunk shape (Codex r2 P2, PR #193 cid
/// 3291860840).
///
/// When both items carry zero EOL observations — single-line bodies
/// in a CRLF file — the per-item policy ties to LF by default, which
/// reintroduces Codex r8 P2 (cid 3256283857). The whole-file
/// `sides.eol_policy` fills that case: it counts the surrounding
/// file context, so a CRLF file resolves to CRLF markers even when
/// the items contribute no observations of their own.
fn emit_addadd_conflict(
    ours: &[u8],
    theirs: &[u8],
    markers: ConflictMarkers<'_>,
    sides: SideSources<'_>,
) -> Vec<u8> {
    let items_policy = EolPolicy::detect(&[ours, theirs]);
    let eol = if items_policy.crlf + items_policy.lf > 0 {
        items_policy.eol()
    } else {
        sides.eol_policy.eol()
    };
    let mut out = Vec::with_capacity(ours.len() + theirs.len() + 64);
    out.extend_from_slice(b"<<<<<<< ");
    out.extend_from_slice(markers.ours.as_bytes());
    out.extend_from_slice(eol);
    out.extend_from_slice(ours);
    ensure_trailing_newline(&mut out, eol);
    out.extend_from_slice(b"=======");
    out.extend_from_slice(eol);
    out.extend_from_slice(theirs);
    ensure_trailing_newline(&mut out, eol);
    out.extend_from_slice(b">>>>>>> ");
    out.extend_from_slice(markers.theirs.as_bytes());
    out.extend_from_slice(eol);
    out
}