File size: 46,122 Bytes
1ef1de7
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""Manifest audit and leakage-safe split helpers."""

from __future__ import annotations

import json
import random
from collections import Counter, defaultdict
from dataclasses import dataclass, field
from pathlib import Path
from typing import Any

from .errors import DataValidationError
from .manifest import ManifestEntry, load_manifest, read_jsonl, sha256_path, sha256_text, write_manifest
from .normalization import script_counts, unsupported_characters
from .validation import check_leakage


@dataclass(slots=True)
class ManifestAudit:
    manifest_path: str
    passed: bool
    sample_count: int
    dataset_counts: dict[str, int]
    split_counts: dict[str, int]
    slice_counts: dict[str, int]
    script_counts: dict[str, int]
    issues: list[str] = field(default_factory=list)
    warnings: list[str] = field(default_factory=list)

    def to_dict(self) -> dict[str, Any]:
        return {
            "manifest_path": self.manifest_path,
            "passed": self.passed,
            "sample_count": self.sample_count,
            "dataset_counts": self.dataset_counts,
            "split_counts": self.split_counts,
            "slice_counts": self.slice_counts,
            "script_counts": self.script_counts,
            "issues": self.issues,
            "warnings": self.warnings,
        }


@dataclass(slots=True)
class SplitSummary:
    train_manifest: str
    eval_manifest: str
    train_count: int
    eval_count: int
    group_count: int
    group_by: str
    seed: int
    leakage_passed: bool
    warnings: list[str] = field(default_factory=list)

    def to_dict(self) -> dict[str, Any]:
        return {
            "train_manifest": self.train_manifest,
            "eval_manifest": self.eval_manifest,
            "train_count": self.train_count,
            "eval_count": self.eval_count,
            "group_count": self.group_count,
            "group_by": self.group_by,
            "seed": self.seed,
            "leakage_passed": self.leakage_passed,
            "warnings": self.warnings,
        }


@dataclass(slots=True)
class MergeSummary:
    output_manifest: str
    input_manifests: list[str]
    sample_count: int
    dataset_counts: dict[str, int]
    warnings: list[str] = field(default_factory=list)

    def to_dict(self) -> dict[str, Any]:
        return {
            "output_manifest": self.output_manifest,
            "input_manifests": self.input_manifests,
            "sample_count": self.sample_count,
            "dataset_counts": self.dataset_counts,
            "warnings": self.warnings,
        }


@dataclass(slots=True)
class RebalanceSummary:
    output_manifest: str
    source_manifest: str
    original_count: int
    rebalanced_count: int
    target_count: int
    slice_counts_before: dict[str, int]
    slice_counts_after: dict[str, int]
    warnings: list[str] = field(default_factory=list)

    def to_dict(self) -> dict[str, Any]:
        return {
            "output_manifest": self.output_manifest,
            "source_manifest": self.source_manifest,
            "original_count": self.original_count,
            "rebalanced_count": self.rebalanced_count,
            "target_count": self.target_count,
            "slice_counts_before": self.slice_counts_before,
            "slice_counts_after": self.slice_counts_after,
            "warnings": self.warnings,
        }


@dataclass(slots=True)
class FilterSummary:
    output_manifest: str
    source_manifest: str
    source_count: int
    selected_count: int
    sample_id_filters: list[str] = field(default_factory=list)
    slice_filters: list[str] = field(default_factory=list)
    document_type_filters: list[str] = field(default_factory=list)
    degradation_filters: list[str] = field(default_factory=list)
    warnings: list[str] = field(default_factory=list)

    def to_dict(self) -> dict[str, Any]:
        return {
            "output_manifest": self.output_manifest,
            "source_manifest": self.source_manifest,
            "source_count": self.source_count,
            "selected_count": self.selected_count,
            "sample_id_filters": self.sample_id_filters,
            "slice_filters": self.slice_filters,
            "document_type_filters": self.document_type_filters,
            "degradation_filters": self.degradation_filters,
            "warnings": self.warnings,
        }


@dataclass(slots=True)
class RecognizerCorpusProfile:
    manifest_path: str
    sample_count: int
    dataset_counts: dict[str, int]
    slice_counts: dict[str, int]
    script_counts: dict[str, int]
    script_sample_counts: dict[str, int]

    def to_dict(self) -> dict[str, Any]:
        return {
            "manifest_path": self.manifest_path,
            "sample_count": self.sample_count,
            "dataset_counts": self.dataset_counts,
            "slice_counts": self.slice_counts,
            "script_counts": self.script_counts,
            "script_sample_counts": self.script_sample_counts,
        }


@dataclass(slots=True)
class RecognizerCorpusAudit:
    train_manifest: str
    eval_manifest: str
    passed: bool
    train_profile: RecognizerCorpusProfile
    eval_profile: RecognizerCorpusProfile
    leakage_passed: bool
    issues: list[str] = field(default_factory=list)
    warnings: list[str] = field(default_factory=list)
    summary_json_path: str | None = None
    summary_md_path: str | None = None

    def to_dict(self) -> dict[str, Any]:
        return {
            "train_manifest": self.train_manifest,
            "eval_manifest": self.eval_manifest,
            "passed": self.passed,
            "train_profile": self.train_profile.to_dict(),
            "eval_profile": self.eval_profile.to_dict(),
            "leakage_passed": self.leakage_passed,
            "issues": self.issues,
            "warnings": self.warnings,
            "summary_json_path": self.summary_json_path,
            "summary_md_path": self.summary_md_path,
        }


@dataclass(slots=True)
class RecognizerCorpusBuildSummary:
    output_manifest: str
    input_manifests: list[str]
    source_count: int
    output_count: int
    target_counts: dict[str, int]
    bucket_counts_before: dict[str, int]
    bucket_counts_after: dict[str, int]
    warnings: list[str] = field(default_factory=list)

    def to_dict(self) -> dict[str, Any]:
        return {
            "output_manifest": self.output_manifest,
            "input_manifests": self.input_manifests,
            "source_count": self.source_count,
            "output_count": self.output_count,
            "target_counts": self.target_counts,
            "bucket_counts_before": self.bucket_counts_before,
            "bucket_counts_after": self.bucket_counts_after,
            "warnings": self.warnings,
        }


def audit_manifest(
    manifest_path: str | Path,
    output_dir: str | Path | None = None,
    *,
    verify_hashes: bool = True,
    require_slices: list[str] | None = None,
    strict_chars: bool = False,
) -> ManifestAudit:
    path = Path(manifest_path)
    rows = list(read_jsonl(path))
    entries: list[ManifestEntry] = []
    issues: list[str] = []
    warnings: list[str] = []
    sample_ids: set[str] = set()
    image_paths: set[str] = set()
    image_hashes: set[str] = set()
    sample_hashes: set[str] = set()
    dataset_counts: Counter[str] = Counter()
    split_counts: Counter[str] = Counter()
    slice_counts: Counter[str] = Counter()
    aggregate_scripts: Counter[str] = Counter()

    for row_index, row in enumerate(rows):
        try:
            entry = ManifestEntry.from_dict(row)
        except DataValidationError as exc:
            issues.append(f"row {row_index}: {exc}")
            continue
        entries.append(entry)
        if entry.sample_id in sample_ids:
            issues.append(f"duplicate sample_id: {entry.sample_id}")
        sample_ids.add(entry.sample_id)
        if entry.image_path in image_paths:
            warnings.append(f"duplicate image_path: {entry.image_path}")
        image_paths.add(entry.image_path)
        input_path = Path(entry.image_path)
        if not input_path.exists():
            issues.append(f"missing image_path for {entry.sample_id}: {input_path}")
        elif verify_hashes:
            actual = sha256_path(input_path)
            if entry.sha256 and entry.sha256 != actual:
                issues.append(f"sha256 mismatch for {entry.sample_id}: manifest={entry.sha256} actual={actual}")
            if actual in image_hashes:
                warnings.append(f"duplicate image sha256: {actual}")
            image_hashes.add(actual)
        elif entry.sha256:
            if entry.sha256 in image_hashes:
                warnings.append(f"duplicate image sha256: {entry.sha256}")
            image_hashes.add(entry.sha256)

        text_sha = sha256_text(entry.text)
        metadata_text_sha = entry.metadata.get("text_sha256") if entry.metadata else None
        if metadata_text_sha and metadata_text_sha != text_sha:
            issues.append(f"text_sha256 mismatch for {entry.sample_id}")
        sample_sha = _entry_sample_signature(entry)
        if sample_sha in sample_hashes:
            warnings.append(f"duplicate sample signature: {sample_sha}")
        sample_hashes.add(sample_sha)
        if strict_chars:
            unsupported = unsupported_characters(entry.text)
            if unsupported:
                issues.append(f"unsupported characters for {entry.sample_id}: {unsupported}")
        dataset_counts[entry.dataset] += 1
        split_counts[entry.split] += 1
        for slice_name in entry_slices(entry):
            slice_counts[slice_name] += 1
        aggregate_scripts.update(script_counts(entry.text))

    for required in require_slices or []:
        if slice_counts.get(required, 0) == 0:
            issues.append(f"required slice missing: {required}")

    audit = ManifestAudit(
        manifest_path=str(path),
        passed=not issues,
        sample_count=len(entries),
        dataset_counts=dict(sorted(dataset_counts.items())),
        split_counts=dict(sorted(split_counts.items())),
        slice_counts=dict(sorted(slice_counts.items())),
        script_counts=dict(sorted(aggregate_scripts.items())),
        issues=issues,
        warnings=warnings,
    )
    if output_dir is not None:
        _write_audit(audit, Path(output_dir))
    return audit


def build_recognizer_corpus(
    manifest_paths: list[str | Path],
    output_manifest: str | Path,
    *,
    target_latin_only: int = 0,
    target_devanagari_only: int = 0,
    target_mixed: int = 0,
    target_real: int = 0,
    target_synthetic: int = 0,
    seed: int = 13,
) -> RecognizerCorpusBuildSummary:
    if not manifest_paths:
        raise DataValidationError("build-recognizer-corpus requires at least one input manifest")
    targets = {
        "latin_only": target_latin_only,
        "devanagari_only": target_devanagari_only,
        "mixed_devanagari_latin": target_mixed,
        "real": target_real,
        "synthetic": target_synthetic,
    }
    invalid = [name for name, value in targets.items() if value < 0]
    if invalid:
        raise DataValidationError(f"recognizer corpus targets must be non-negative: {', '.join(invalid)}")
    if all(value == 0 for value in targets.values()):
        raise DataValidationError("build-recognizer-corpus requires at least one positive target")

    entries: list[ManifestEntry] = []
    seen_sample_ids: set[str] = set()
    seen_signatures: set[str] = set()
    warnings: list[str] = []
    for manifest_path in manifest_paths:
        for entry in load_manifest(manifest_path):
            if entry.sample_id in seen_sample_ids:
                raise DataValidationError(f"duplicate sample_id across recognizer corpus inputs: {entry.sample_id}")
            seen_sample_ids.add(entry.sample_id)
            signature = _entry_sample_signature(entry)
            if signature in seen_signatures:
                warnings.append(f"duplicate sample signature across recognizer corpus inputs: {entry.sample_id}")
            seen_signatures.add(signature)
            entries.append(entry)
    if not entries:
        raise DataValidationError("build-recognizer-corpus requires non-empty input manifests")

    output_entries = list(entries)
    rng = random.Random(seed)
    duplicate_index = 0
    counts_before = _recognizer_bucket_counts(output_entries)
    for bucket_name, target in targets.items():
        if target <= 0:
            continue
        current = _recognizer_bucket_counts(output_entries).get(bucket_name, 0)
        if current >= target:
            continue
        bucket = [entry for entry in output_entries if _entry_in_recognizer_bucket(entry, bucket_name)]
        if not bucket:
            raise DataValidationError(f"recognizer corpus cannot satisfy missing bucket: {bucket_name}")
        deficit = target - current
        warnings.append(f"upsampled recognizer bucket {bucket_name} by {deficit} samples")
        for _ in range(deficit):
            source = rng.choice(bucket)
            duplicate_index += 1
            output_entries.append(_duplicate_recognizer_corpus_entry(source, bucket_name=bucket_name, duplicate_index=duplicate_index))

    out_path = Path(output_manifest)
    write_manifest(output_entries, out_path)
    summary = RecognizerCorpusBuildSummary(
        output_manifest=str(out_path),
        input_manifests=[str(Path(path)) for path in manifest_paths],
        source_count=len(entries),
        output_count=len(output_entries),
        target_counts={key: value for key, value in targets.items() if value > 0},
        bucket_counts_before=_recognizer_bucket_counts(entries),
        bucket_counts_after=_recognizer_bucket_counts(output_entries),
        warnings=warnings,
    )
    _write_recognizer_corpus_build_summary(summary, out_path)
    return summary


def audit_recognizer_corpus(
    train_manifest: str | Path,
    eval_manifest: str | Path,
    output_dir: str | Path | None = None,
    *,
    min_train_samples: int = 1000,
    min_eval_samples: int = 100,
    min_train_english: int = 200,
    min_train_devanagari: int = 200,
    min_train_mixed: int = 25,
    min_eval_english: int = 50,
    min_eval_devanagari: int = 50,
    min_eval_mixed: int = 10,
    min_train_latin_only: int = 50,
    min_train_devanagari_only: int = 50,
    min_eval_latin_only: int = 10,
    min_eval_devanagari_only: int = 10,
    min_train_real: int = 100,
    min_train_synthetic: int = 100,
) -> RecognizerCorpusAudit:
    thresholds = {
        "min_train_samples": min_train_samples,
        "min_eval_samples": min_eval_samples,
        "min_train_english": min_train_english,
        "min_train_devanagari": min_train_devanagari,
        "min_train_mixed": min_train_mixed,
        "min_eval_english": min_eval_english,
        "min_eval_devanagari": min_eval_devanagari,
        "min_eval_mixed": min_eval_mixed,
        "min_train_latin_only": min_train_latin_only,
        "min_train_devanagari_only": min_train_devanagari_only,
        "min_eval_latin_only": min_eval_latin_only,
        "min_eval_devanagari_only": min_eval_devanagari_only,
        "min_train_real": min_train_real,
        "min_train_synthetic": min_train_synthetic,
    }
    invalid_thresholds = [name for name, value in thresholds.items() if value < 0]
    if invalid_thresholds:
        raise DataValidationError(f"recognizer corpus thresholds must be non-negative: {', '.join(invalid_thresholds)}")

    train_path = Path(train_manifest)
    eval_path = Path(eval_manifest)
    train_entries = load_manifest(train_path)
    eval_entries = load_manifest(eval_path)
    train_profile = _recognizer_corpus_profile(train_path, train_entries)
    eval_profile = _recognizer_corpus_profile(eval_path, eval_entries)
    leakage = check_leakage([train_path], eval_path)

    issues: list[str] = []
    warnings: list[str] = []
    _append_minimum_issue(issues, "train samples", train_profile.sample_count, min_train_samples)
    _append_minimum_issue(issues, "eval samples", eval_profile.sample_count, min_eval_samples)
    _append_minimum_issue(issues, "train latin-script samples", train_profile.script_sample_counts.get("contains_latin", 0), min_train_english)
    _append_minimum_issue(issues, "train devanagari-script samples", train_profile.script_sample_counts.get("contains_devanagari", 0), min_train_devanagari)
    _append_minimum_issue(issues, "train mixed-script samples", train_profile.script_sample_counts.get("mixed_devanagari_latin", 0), min_train_mixed)
    _append_minimum_issue(issues, "eval latin-script samples", eval_profile.script_sample_counts.get("contains_latin", 0), min_eval_english)
    _append_minimum_issue(issues, "eval devanagari-script samples", eval_profile.script_sample_counts.get("contains_devanagari", 0), min_eval_devanagari)
    _append_minimum_issue(issues, "eval mixed-script samples", eval_profile.script_sample_counts.get("mixed_devanagari_latin", 0), min_eval_mixed)
    _append_minimum_issue(issues, "train latin-only samples", train_profile.script_sample_counts.get("latin_only", 0), min_train_latin_only)
    _append_minimum_issue(issues, "train devanagari-only samples", train_profile.script_sample_counts.get("devanagari_only", 0), min_train_devanagari_only)
    _append_minimum_issue(issues, "eval latin-only samples", eval_profile.script_sample_counts.get("latin_only", 0), min_eval_latin_only)
    _append_minimum_issue(issues, "eval devanagari-only samples", eval_profile.script_sample_counts.get("devanagari_only", 0), min_eval_devanagari_only)
    _append_minimum_issue(issues, "train real slice samples", train_profile.slice_counts.get("real", 0), min_train_real)
    _append_minimum_issue(issues, "train synthetic slice samples", train_profile.slice_counts.get("synthetic", 0), min_train_synthetic)
    if not leakage.passed:
        issues.append(f"train/eval leakage detected: {leakage.overlaps}")
    if eval_profile.slice_counts.get("hard_eval", 0) == 0:
        warnings.append("eval manifest has no hard_eval slice; recognizer admission should still use a separate hard gate")
    if train_profile.script_sample_counts.get("latin_only", 0) == 0:
        warnings.append("train manifest has no latin-only samples; English preservation is unlikely")
    if train_profile.script_sample_counts.get("devanagari_only", 0) == 0:
        warnings.append("train manifest has no devanagari-only samples; Nepali isolation is unlikely")

    audit = RecognizerCorpusAudit(
        train_manifest=str(train_path),
        eval_manifest=str(eval_path),
        passed=not issues,
        train_profile=train_profile,
        eval_profile=eval_profile,
        leakage_passed=leakage.passed,
        issues=issues,
        warnings=warnings,
    )
    if output_dir is not None:
        _write_recognizer_corpus_audit(audit, Path(output_dir))
    return audit


def split_manifest(
    manifest_path: str | Path,
    output_dir: str | Path,
    *,
    eval_ratio: float = 0.15,
    seed: int = 13,
    stratify_by: str = "slices",
    group_by: str = "sample_signature",
    train_name: str = "train.jsonl",
    eval_name: str = "eval.jsonl",
) -> SplitSummary:
    if eval_ratio <= 0 or eval_ratio >= 1:
        raise DataValidationError("eval_ratio must be between 0 and 1")
    entries = load_manifest(manifest_path)
    if len(entries) < 2:
        raise DataValidationError("split-manifest requires at least two entries")
    groups = _group_entries(entries, group_by=group_by)
    if len(groups) < 2:
        raise DataValidationError("split-manifest requires at least two unique sample groups")
    rng = random.Random(seed)
    strata: dict[str, list[tuple[str, list[ManifestEntry]]]] = defaultdict(list)
    for group_key, group_entries in groups.items():
        strata[_stratum_key(group_entries, stratify_by)].append((group_key, group_entries))

    eval_group_keys: set[str] = set()
    for stratum, stratum_groups in sorted(strata.items()):
        shuffled = list(stratum_groups)
        rng.shuffle(shuffled)
        target = max(1, round(len(shuffled) * eval_ratio)) if len(shuffled) > 1 else 0
        for group_key, _ in shuffled[:target]:
            eval_group_keys.add(group_key)

    if not eval_group_keys:
        eval_group_keys.add(sorted(groups)[0])
    if len(eval_group_keys) == len(groups):
        eval_group_keys.remove(sorted(eval_group_keys)[-1])

    train_entries: list[ManifestEntry] = []
    eval_entries: list[ManifestEntry] = []
    for group_key, group_entries in groups.items():
        target = eval_entries if group_key in eval_group_keys else train_entries
        split_name = "eval" if group_key in eval_group_keys else "train"
        target.extend(_with_split(entry, split_name) for entry in group_entries)

    out = Path(output_dir)
    out.mkdir(parents=True, exist_ok=True)
    train_path = out / train_name
    eval_path = out / eval_name
    write_manifest(train_entries, train_path)
    write_manifest(eval_entries, eval_path)
    leakage = check_leakage([train_path], eval_path)
    warnings: list[str] = []
    if not leakage.passed:
        raise DataValidationError(f"generated split has leakage: {leakage.to_dict()}")
    if not eval_entries:
        warnings.append("eval split is empty")
    if not train_entries:
        warnings.append("train split is empty")
    summary = SplitSummary(
        train_manifest=str(train_path),
        eval_manifest=str(eval_path),
        train_count=len(train_entries),
        eval_count=len(eval_entries),
        group_count=len(groups),
        group_by=group_by,
        seed=seed,
        leakage_passed=leakage.passed,
        warnings=warnings,
    )
    _write_split_summary(summary, out)
    return summary


def merge_manifests(
    manifest_paths: list[str | Path],
    output_manifest: str | Path,
) -> MergeSummary:
    if len(manifest_paths) < 2:
        raise DataValidationError("merge-manifests requires at least two input manifests")
    merged: list[ManifestEntry] = []
    sample_ids: set[str] = set()
    sample_signatures: set[str] = set()
    dataset_counts: Counter[str] = Counter()
    warnings: list[str] = []
    ordered_inputs = [str(Path(path)) for path in manifest_paths]

    for manifest_path in manifest_paths:
        entries = load_manifest(manifest_path)
        for entry in entries:
            if entry.sample_id in sample_ids:
                raise DataValidationError(f"duplicate sample_id across merged manifests: {entry.sample_id}")
            sample_ids.add(entry.sample_id)
            signature = _entry_sample_signature(entry)
            if signature in sample_signatures:
                warnings.append(f"duplicate sample signature across merged manifests: {entry.sample_id}")
            sample_signatures.add(signature)
            dataset_counts[entry.dataset] += 1
            merged.append(entry)
    out_path = Path(output_manifest)
    write_manifest(merged, out_path)
    summary = MergeSummary(
        output_manifest=str(out_path),
        input_manifests=ordered_inputs,
        sample_count=len(merged),
        dataset_counts=dict(sorted(dataset_counts.items())),
        warnings=warnings,
    )
    _write_merge_summary(summary, out_path)
    return summary


def rebalance_manifest(
    manifest_path: str | Path,
    output_manifest: str | Path,
    *,
    slices: list[str],
    target_count: int | None = None,
    seed: int = 13,
) -> RebalanceSummary:
    if not slices:
        raise DataValidationError("rebalance-manifest requires at least one --slice")
    requested_slices = [slice_name.strip() for slice_name in slices if slice_name.strip()]
    if not requested_slices:
        raise DataValidationError("rebalance-manifest requires at least one non-empty --slice")
    entries = load_manifest(manifest_path)
    if not entries:
        raise DataValidationError("rebalance-manifest requires a non-empty manifest")
    buckets: dict[str, list[ManifestEntry]] = {slice_name: [] for slice_name in requested_slices}
    for entry in entries:
        entry_slice_set = set(entry_slices(entry))
        for slice_name in requested_slices:
            if slice_name in entry_slice_set:
                buckets[slice_name].append(entry)
    missing = [slice_name for slice_name, bucket in buckets.items() if not bucket]
    if missing:
        raise DataValidationError(f"rebalance-manifest missing requested slices: {', '.join(missing)}")
    before_counts = {slice_name: len(bucket) for slice_name, bucket in buckets.items()}
    target = target_count if target_count is not None else max(before_counts.values())
    if target <= 0:
        raise DataValidationError("rebalance-manifest target_count must be positive")
    rng = random.Random(seed)
    rebalanced = list(entries)
    duplicate_index = 0
    warnings: list[str] = []
    for slice_name, bucket in buckets.items():
        deficit = target - len(bucket)
        if deficit <= 0:
            continue
        warnings.append(f"upsampled slice {slice_name} by {deficit} samples")
        for _ in range(deficit):
            source = rng.choice(bucket)
            duplicate_index += 1
            rebalanced.append(_duplicate_entry(source, slice_name=slice_name, duplicate_index=duplicate_index))
    out_path = Path(output_manifest)
    write_manifest(rebalanced, out_path)
    after_counts = Counter[str]()
    for entry in rebalanced:
        for slice_name in requested_slices:
            if slice_name in set(entry_slices(entry)):
                after_counts[slice_name] += 1
    summary = RebalanceSummary(
        output_manifest=str(out_path),
        source_manifest=str(Path(manifest_path)),
        original_count=len(entries),
        rebalanced_count=len(rebalanced),
        target_count=target,
        slice_counts_before=dict(sorted(before_counts.items())),
        slice_counts_after=dict(sorted(after_counts.items())),
        warnings=warnings,
    )
    _write_rebalance_summary(summary, out_path)
    return summary


def filter_manifest(
    manifest_path: str | Path,
    output_manifest: str | Path,
    *,
    sample_ids: list[str] | None = None,
    slices: list[str] | None = None,
    document_types: list[str] | None = None,
    degradations: list[str] | None = None,
    limit: int | None = None,
) -> FilterSummary:
    entries = load_manifest(manifest_path)
    requested_sample_ids = [value for value in (sample_ids or []) if value]
    requested_slices = [value for value in (slices or []) if value]
    requested_document_types = [value for value in (document_types or []) if value]
    requested_degradations = [value for value in (degradations or []) if value]
    if not any([requested_sample_ids, requested_slices, requested_document_types, requested_degradations, limit is not None]):
        raise DataValidationError("filter-manifest requires at least one filter or --limit")

    selected = [
        entry
        for entry in entries
        if _matches_entry_filters(
            entry,
            sample_ids=requested_sample_ids,
            slices=requested_slices,
            document_types=requested_document_types,
            degradations=requested_degradations,
        )
    ]
    if requested_sample_ids:
        selected_by_id = {entry.sample_id for entry in selected}
        missing = [sample_id for sample_id in requested_sample_ids if sample_id not in selected_by_id]
        if missing:
            raise DataValidationError(f"requested sample_ids not found after filtering: {', '.join(missing)}")
    if limit is not None:
        if limit <= 0:
            raise DataValidationError("limit must be positive")
        selected = selected[:limit]
    if not selected:
        raise DataValidationError("filter-manifest produced zero samples")

    out_path = Path(output_manifest)
    write_manifest(selected, out_path)
    summary = FilterSummary(
        output_manifest=str(out_path),
        source_manifest=str(Path(manifest_path)),
        source_count=len(entries),
        selected_count=len(selected),
        sample_id_filters=requested_sample_ids,
        slice_filters=requested_slices,
        document_type_filters=requested_document_types,
        degradation_filters=requested_degradations,
    )
    _write_filter_summary(summary, out_path)
    return summary


def entry_slices(entry: ManifestEntry) -> list[str]:
    metadata = entry.metadata or {}
    values: set[str] = set()
    for key in ("slice", "script", "document_type", "degradation", "language", "input_format"):
        value = metadata.get(key)
        if isinstance(value, str) and value:
            values.add(value)
    for key in ("slices", "scripts", "tags"):
        value = metadata.get(key)
        if isinstance(value, list):
            values.update(str(item) for item in value if str(item))
        elif isinstance(value, str) and value:
            values.add(value)
    return sorted(values)


def _recognizer_bucket_counts(entries: list[ManifestEntry]) -> dict[str, int]:
    counts: Counter[str] = Counter()
    for entry in entries:
        for bucket_name in ("latin_only", "devanagari_only", "mixed_devanagari_latin", "real", "synthetic"):
            if _entry_in_recognizer_bucket(entry, bucket_name):
                counts[bucket_name] += 1
    return dict(sorted(counts.items()))


def _entry_in_recognizer_bucket(entry: ManifestEntry, bucket_name: str) -> bool:
    counts = script_counts(entry.text)
    has_devanagari = counts.get("devanagari", 0) > 0
    has_latin = counts.get("latin", 0) > 0
    if bucket_name == "latin_only":
        return has_latin and not has_devanagari
    if bucket_name == "devanagari_only":
        return has_devanagari and not has_latin
    if bucket_name == "mixed_devanagari_latin":
        return has_latin and has_devanagari
    if bucket_name in {"real", "synthetic"}:
        return bucket_name in set(entry_slices(entry))
    raise DataValidationError(f"unknown recognizer corpus bucket: {bucket_name}")


def _duplicate_recognizer_corpus_entry(entry: ManifestEntry, *, bucket_name: str, duplicate_index: int) -> ManifestEntry:
    metadata = dict(entry.metadata or {})
    metadata["recognizer_corpus_source_sample_id"] = entry.sample_id
    metadata["recognizer_corpus_bucket"] = bucket_name
    metadata["recognizer_corpus_duplicate"] = True
    return ManifestEntry(
        sample_id=f"{entry.sample_id}~recognizer-{bucket_name}-{duplicate_index:06d}",
        dataset=entry.dataset,
        split=entry.split,
        image_path=entry.image_path,
        text=entry.text,
        sha256=entry.sha256,
        metadata=metadata,
    )


def _recognizer_corpus_profile(path: Path, entries: list[ManifestEntry]) -> RecognizerCorpusProfile:
    dataset_counts: Counter[str] = Counter()
    slice_counts: Counter[str] = Counter()
    aggregate_scripts: Counter[str] = Counter()
    script_sample_counts: Counter[str] = Counter()
    for entry in entries:
        dataset_counts[entry.dataset] += 1
        for slice_name in entry_slices(entry):
            slice_counts[slice_name] += 1
        counts = script_counts(entry.text)
        aggregate_scripts.update(counts)
        has_devanagari = counts.get("devanagari", 0) > 0
        has_latin = counts.get("latin", 0) > 0
        if has_latin:
            script_sample_counts["contains_latin"] += 1
        if has_devanagari:
            script_sample_counts["contains_devanagari"] += 1
        if has_latin and has_devanagari:
            script_sample_counts["mixed_devanagari_latin"] += 1
        elif has_latin:
            script_sample_counts["latin_only"] += 1
        elif has_devanagari:
            script_sample_counts["devanagari_only"] += 1
        else:
            script_sample_counts["other_only"] += 1
    return RecognizerCorpusProfile(
        manifest_path=str(path),
        sample_count=len(entries),
        dataset_counts=dict(sorted(dataset_counts.items())),
        slice_counts=dict(sorted(slice_counts.items())),
        script_counts=dict(sorted(aggregate_scripts.items())),
        script_sample_counts=dict(sorted(script_sample_counts.items())),
    )


def _append_minimum_issue(issues: list[str], label: str, actual: int, required: int) -> None:
    if actual < required:
        issues.append(f"{label} {actual} < required {required}")


def _group_entries(entries: list[ManifestEntry], *, group_by: str = "sample_signature") -> dict[str, list[ManifestEntry]]:
    if group_by not in {"sample_signature", "text_sha256", "image_sha256", "image_path", "sample_id"}:
        raise DataValidationError("group_by must be one of: sample_signature, text_sha256, image_sha256, image_path, sample_id")
    groups: dict[str, list[ManifestEntry]] = defaultdict(list)
    for entry in entries:
        groups[_entry_group_key(entry, group_by)].append(entry)
    return dict(groups)


def _entry_group_key(entry: ManifestEntry, group_by: str) -> str:
    metadata = entry.metadata or {}
    if group_by == "sample_signature":
        return _entry_sample_signature(entry)
    if group_by == "text_sha256":
        value = metadata.get("text_sha256")
        return str(value) if value else sha256_text(entry.text)
    if group_by == "image_sha256":
        return str(entry.sha256 or entry.image_path)
    if group_by == "image_path":
        return entry.image_path
    if group_by == "sample_id":
        return entry.sample_id
    raise DataValidationError("group_by must be one of: sample_signature, text_sha256, image_sha256, image_path, sample_id")


def _entry_sample_signature(entry: ManifestEntry) -> str:
    metadata = entry.metadata or {}
    sample_hash = metadata.get("sample_sha256")
    if isinstance(sample_hash, str) and sample_hash:
        return sample_hash
    image_hash = entry.sha256 or entry.image_path
    text_hash = metadata.get("text_sha256") if isinstance(metadata.get("text_sha256"), str) else sha256_text(entry.text)
    return sha256_text(f"{image_hash}\n{text_hash}")


def _stratum_key(entries: list[ManifestEntry], stratify_by: str) -> str:
    if stratify_by == "none":
        return "all"
    if stratify_by == "dataset":
        return entries[0].dataset
    if stratify_by == "document_type":
        return str((entries[0].metadata or {}).get("document_type") or "unknown")
    if stratify_by != "slices":
        raise DataValidationError("stratify_by must be one of: slices, dataset, document_type, none")
    slices = sorted({slice_name for entry in entries for slice_name in entry_slices(entry)})
    return ",".join(slices) if slices else "unsliced"


def _with_split(entry: ManifestEntry, split: str) -> ManifestEntry:
    metadata = dict(entry.metadata)
    metadata["original_split"] = entry.split
    return ManifestEntry(
        sample_id=entry.sample_id,
        dataset=entry.dataset,
        split=split,
        image_path=entry.image_path,
        text=entry.text,
        sha256=entry.sha256,
        metadata=metadata,
    )


def _duplicate_entry(entry: ManifestEntry, *, slice_name: str, duplicate_index: int) -> ManifestEntry:
    metadata = dict(entry.metadata or {})
    metadata["rebalance_source_sample_id"] = entry.sample_id
    metadata["rebalance_slice"] = slice_name
    metadata["rebalance_duplicate"] = True
    return ManifestEntry(
        sample_id=f"{entry.sample_id}~rebalance-{slice_name}-{duplicate_index:06d}",
        dataset=entry.dataset,
        split=entry.split,
        image_path=entry.image_path,
        text=entry.text,
        sha256=entry.sha256,
        metadata=metadata,
    )


def _matches_entry_filters(
    entry: ManifestEntry,
    *,
    sample_ids: list[str],
    slices: list[str],
    document_types: list[str],
    degradations: list[str],
) -> bool:
    metadata = entry.metadata or {}
    if sample_ids and entry.sample_id not in set(sample_ids):
        return False
    entry_slice_set = set(entry_slices(entry))
    if slices and not set(slices).issubset(entry_slice_set):
        return False
    document_type = str(metadata.get("document_type") or "")
    if document_types and document_type not in set(document_types):
        return False
    degradation = str(metadata.get("degradation") or "")
    if degradations and degradation not in set(degradations):
        return False
    return True


def _write_audit(audit: ManifestAudit, output_dir: Path) -> None:
    output_dir.mkdir(parents=True, exist_ok=True)
    (output_dir / "manifest-audit.json").write_text(json.dumps(audit.to_dict(), ensure_ascii=False, indent=2) + "\n", encoding="utf-8")
    lines = [
        "# Manifest Audit",
        "",
        f"Passed: `{'yes' if audit.passed else 'no'}`",
        f"Samples: `{audit.sample_count}`",
        "",
        "## Issues",
        "",
    ]
    lines.extend(f"- {item}" for item in audit.issues) if audit.issues else lines.append("- none")
    lines.extend(["", "## Warnings", ""])
    lines.extend(f"- {item}" for item in audit.warnings) if audit.warnings else lines.append("- none")
    (output_dir / "manifest-audit.md").write_text("\n".join(lines) + "\n", encoding="utf-8")


def _write_recognizer_corpus_audit(audit: RecognizerCorpusAudit, output_dir: Path) -> None:
    output_dir.mkdir(parents=True, exist_ok=True)
    json_path = output_dir / "recognizer-corpus-audit.json"
    md_path = output_dir / "recognizer-corpus-audit.md"
    audit.summary_json_path = str(json_path)
    audit.summary_md_path = str(md_path)
    json_path.write_text(json.dumps(audit.to_dict(), ensure_ascii=False, indent=2) + "\n", encoding="utf-8")
    lines = [
        "# Recognizer Corpus Audit",
        "",
        f"Passed: `{'yes' if audit.passed else 'no'}`",
        f"Train manifest: `{audit.train_manifest}`",
        f"Eval manifest: `{audit.eval_manifest}`",
        f"Leakage check: `{'pass' if audit.leakage_passed else 'fail'}`",
        "",
        "## Profiles",
        "",
        "| split | samples | latin | devanagari | mixed | real | synthetic |",
        "| --- | ---: | ---: | ---: | ---: | ---: | ---: |",
        _recognizer_profile_row("train", audit.train_profile),
        _recognizer_profile_row("eval", audit.eval_profile),
        "",
        "## Issues",
        "",
    ]
    lines.extend(f"- {item}" for item in audit.issues) if audit.issues else lines.append("- none")
    lines.extend(["", "## Warnings", ""])
    lines.extend(f"- {item}" for item in audit.warnings) if audit.warnings else lines.append("- none")
    lines.extend(
        [
            "",
            "## Next Actions",
            "",
        ]
    )
    lines.extend(_recognizer_corpus_next_actions(audit))
    md_path.write_text("\n".join(lines) + "\n", encoding="utf-8")


def _write_recognizer_corpus_build_summary(summary: RecognizerCorpusBuildSummary, output_manifest: Path) -> None:
    summary_json = output_manifest.with_name(f"{output_manifest.stem}-recognizer-corpus-build.json")
    summary_md = output_manifest.with_name(f"{output_manifest.stem}-recognizer-corpus-build.md")
    summary_json.write_text(json.dumps(summary.to_dict(), ensure_ascii=False, indent=2) + "\n", encoding="utf-8")
    lines = [
        "# Recognizer Corpus Build",
        "",
        f"Output manifest: `{summary.output_manifest}`",
        f"Source samples: `{summary.source_count}`",
        f"Output samples: `{summary.output_count}`",
        "",
        "## Targets",
        "",
    ]
    lines.extend(f"- `{key}`: {value}" for key, value in summary.target_counts.items()) if summary.target_counts else lines.append("- none")
    lines.extend(["", "## Bucket Counts Before", ""])
    lines.extend(f"- `{key}`: {value}" for key, value in summary.bucket_counts_before.items()) if summary.bucket_counts_before else lines.append("- none")
    lines.extend(["", "## Bucket Counts After", ""])
    lines.extend(f"- `{key}`: {value}" for key, value in summary.bucket_counts_after.items()) if summary.bucket_counts_after else lines.append("- none")
    lines.extend(["", "## Warnings", ""])
    lines.extend(f"- {item}" for item in summary.warnings) if summary.warnings else lines.append("- none")
    summary_md.write_text("\n".join(lines) + "\n", encoding="utf-8")


def _recognizer_profile_row(split: str, profile: RecognizerCorpusProfile) -> str:
    scripts = profile.script_sample_counts
    slices = profile.slice_counts
    return (
        f"| {split} | {profile.sample_count} | {scripts.get('contains_latin', 0)} | "
        f"{scripts.get('contains_devanagari', 0)} | {scripts.get('mixed_devanagari_latin', 0)} | "
        f"{slices.get('real', 0)} | {slices.get('synthetic', 0)} |"
    )


def _recognizer_corpus_next_actions(audit: RecognizerCorpusAudit) -> list[str]:
    if audit.passed:
        return ["- corpus passes the configured recognizer training gate"]
    actions: list[str] = []
    issue_text = "\n".join(audit.issues)
    if "latin-script" in issue_text:
        actions.append("- add or upsample English/Latin line crops before training")
    if "devanagari-script" in issue_text:
        actions.append("- add or upsample Nepali/Devanagari line crops before training")
    if "mixed-script" in issue_text:
        actions.append("- add mixed Nepali/English samples so the recognizer sees script transitions")
    if "latin-only" in issue_text:
        actions.append("- add Latin-only English samples to protect English preservation")
    if "devanagari-only" in issue_text:
        actions.append("- add Devanagari-only Nepali samples to prevent mixed-script-only overfitting")
    if "real slice" in issue_text:
        actions.append("- add real scanned/captured Nepali data; synthetic-only training is not enough")
    if "synthetic slice" in issue_text:
        actions.append("- add synthetic coverage for controllable character and layout diversity")
    if "leakage" in issue_text:
        actions.append("- regenerate train/eval split with leakage-safe grouping before training")
    if not actions:
        actions.append("- inspect failed thresholds and rebuild the training manifest")
    return actions


def _write_split_summary(summary: SplitSummary, output_dir: Path) -> None:
    (output_dir / "split-summary.json").write_text(json.dumps(summary.to_dict(), ensure_ascii=False, indent=2) + "\n", encoding="utf-8")
    lines = [
        "# Manifest Split",
        "",
        f"Train manifest: `{summary.train_manifest}`",
        f"Eval manifest: `{summary.eval_manifest}`",
        f"Train samples: `{summary.train_count}`",
        f"Eval samples: `{summary.eval_count}`",
        f"Sample groups: `{summary.group_count}`",
        f"Group by: `{summary.group_by}`",
        f"Leakage check: `{'pass' if summary.leakage_passed else 'fail'}`",
        "",
        "## Warnings",
        "",
    ]
    lines.extend(f"- {item}" for item in summary.warnings) if summary.warnings else lines.append("- none")
    (output_dir / "split-summary.md").write_text("\n".join(lines) + "\n", encoding="utf-8")


def _write_merge_summary(summary: MergeSummary, output_manifest: Path) -> None:
    summary_json = output_manifest.with_name(f"{output_manifest.stem}-merge.json")
    summary_md = output_manifest.with_name(f"{output_manifest.stem}-merge.md")
    summary_json.write_text(json.dumps(summary.to_dict(), ensure_ascii=False, indent=2) + "\n", encoding="utf-8")
    lines = [
        "# Manifest Merge",
        "",
        f"Output manifest: `{summary.output_manifest}`",
        f"Samples: `{summary.sample_count}`",
        "",
        "## Input Manifests",
        "",
    ]
    lines.extend(f"- `{path}`" for path in summary.input_manifests)
    lines.extend(["", "## Warnings", ""])
    lines.extend(f"- {item}" for item in summary.warnings) if summary.warnings else lines.append("- none")
    summary_md.write_text("\n".join(lines) + "\n", encoding="utf-8")


def _write_rebalance_summary(summary: RebalanceSummary, output_manifest: Path) -> None:
    summary_json = output_manifest.with_name(f"{output_manifest.stem}-rebalance.json")
    summary_md = output_manifest.with_name(f"{output_manifest.stem}-rebalance.md")
    summary_json.write_text(json.dumps(summary.to_dict(), ensure_ascii=False, indent=2) + "\n", encoding="utf-8")
    lines = [
        "# Manifest Rebalance",
        "",
        f"Source manifest: `{summary.source_manifest}`",
        f"Output manifest: `{summary.output_manifest}`",
        f"Original samples: `{summary.original_count}`",
        f"Rebalanced samples: `{summary.rebalanced_count}`",
        f"Target count: `{summary.target_count}`",
        "",
        "## Slice Counts Before",
        "",
    ]
    lines.extend(f"- `{key}`: {value}" for key, value in summary.slice_counts_before.items())
    lines.extend(["", "## Slice Counts After", ""])
    lines.extend(f"- `{key}`: {value}" for key, value in summary.slice_counts_after.items())
    lines.extend(["", "## Warnings", ""])
    lines.extend(f"- {item}" for item in summary.warnings) if summary.warnings else lines.append("- none")
    summary_md.write_text("\n".join(lines) + "\n", encoding="utf-8")


def _write_filter_summary(summary: FilterSummary, output_manifest: Path) -> None:
    summary_json = output_manifest.with_name(f"{output_manifest.stem}-filter.json")
    summary_md = output_manifest.with_name(f"{output_manifest.stem}-filter.md")
    summary_json.write_text(json.dumps(summary.to_dict(), ensure_ascii=False, indent=2) + "\n", encoding="utf-8")
    lines = [
        "# Manifest Filter",
        "",
        f"Source manifest: `{summary.source_manifest}`",
        f"Output manifest: `{summary.output_manifest}`",
        f"Source samples: `{summary.source_count}`",
        f"Selected samples: `{summary.selected_count}`",
        f"Sample IDs: `{', '.join(summary.sample_id_filters)}`",
        f"Slices: `{', '.join(summary.slice_filters)}`",
        f"Document types: `{', '.join(summary.document_type_filters)}`",
        f"Degradations: `{', '.join(summary.degradation_filters)}`",
        "",
        "## Warnings",
        "",
    ]
    lines.extend(f"- {item}" for item in summary.warnings) if summary.warnings else lines.append("- none")
    summary_md.write_text("\n".join(lines) + "\n", encoding="utf-8")