File size: 42,469 Bytes
7a12bf7
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""
DSPy-free utilities for KVT4 parsing and normalization.

This module contains all parsing, normalization, and validation logic
that does NOT depend on DSPy. It can be used in production pipelines
without importing the DSPy framework.

Extracted from dspy_integration.py as part of Phase 0 decomposition.
"""

import ast
import json
import os
import re
from typing import List, Optional

# =============================================================================
# REGEX PATTERNS
# =============================================================================

_MEDGEMMA_INTERNAL_TOKEN_RE = re.compile(r"<unused\d+>")
_MEDGEMMA_THOUGHT_LINE_RE = re.compile(r"^\s*(<unused\d+>\w*\s*)?thought\b.*$", re.IGNORECASE)
_DSPY_QUOTED_FACT_RE = re.compile(r"«([^»]+)»")
_PARTIAL_JSON_FACT_RE = re.compile(
    r"""\{\s*["']cluster["']\s*:\s*["']([^"']+)["']\s*,\s*"""
    r"""["']keyword["']\s*:\s*["']([^"']+)["']\s*,\s*"""
    r"""["']value["']\s*:\s*["']([^"']+)["']\s*,\s*"""
    r"""["']timestamp["']\s*:\s*["']([^"']+)["']\s*\}""",
    re.IGNORECASE,
)
_PARTIAL_GROUPED_CLUSTER_BLOCK_RE = re.compile(
    r'"(?P<cluster>DEMOGRAPHICS|VITALS|LABS|PROBLEMS|SYMPTOMS|MEDICATIONS|PROCEDURES|UTILIZATION|DISPOSITION)"\s*:\s*\[',
    re.IGNORECASE,
)
_PARTIAL_GROUPED_ITEM_RE = re.compile(
    r"""\{\s*["']K["']\s*:\s*["'](?P<k>[^"']+)["']\s*,\s*"""
    r"""["']V["']\s*:\s*(?P<v>"[^"]*"|-?\d+(?:\.\d+)?|true|false)\s*,\s*"""
    r"""["']T["']\s*:\s*["'](?P<t>[^"']+)["']\s*\}""",
    re.IGNORECASE,
)


# =============================================================================
# CANONICAL KEYWORDS (MVP)
# =============================================================================

CANONICAL_VITALS = [
    "Heart Rate",
    "Systolic BP",
    "Diastolic BP",
    "Respiratory Rate",
    "Temperature",
    "SpO2",
    "Weight",
]

CANONICAL_LABS = [
    "Hemoglobin",
    "Hematocrit",
    "WBC",
    "Platelet",
    "Sodium",
    "Potassium",
    "Creatinine",
    "BUN",
    "Glucose",
    "Bicarbonate",
]

CANONICAL_DEMOGRAPHICS = [
    "Age",
    "Sex",
]

STRICT_KEYWORDS_READMISSION: dict[str, set[str]] = {
    "DEMOGRAPHICS": set(CANONICAL_DEMOGRAPHICS),
    "VITALS": set(CANONICAL_VITALS),
    "LABS": set(CANONICAL_LABS),
    "MEDICATIONS": {
        "Medication Count",
        "New Medications Count",
        "Polypharmacy",
        "Anticoagulation",
        "Insulin Therapy",
        "Opioid Therapy",
        "Diuretic Therapy",
    },
    "PROCEDURES": {
        "Any Procedure",
        "Surgery",
        "Dialysis",
        "Mechanical Ventilation",
    },
    "UTILIZATION": {
        "Prior Admissions 12mo",
        "ED Visits 6mo",
        "Days Since Last Admission",
        "Current Length of Stay",
    },
    "DISPOSITION": {
        "Discharge Disposition",
        "Mental Status",
    },
}
READMISSION_CLUSTERS = {
    "DEMOGRAPHICS",
    "VITALS",
    "LABS",
    "PROBLEMS",
    "SYMPTOMS",
    "MEDICATIONS",
    "PROCEDURES",
    "UTILIZATION",
    "DISPOSITION",
}


# =============================================================================
# OUTPUT PARSING HELPERS
# =============================================================================

def strip_medgemma_internal_tokens(text: str) -> str:
    """Remove MedGemma internal tokens and thinking blocks from text.
    
    IMPORTANT: Only strips the internal token itself (e.g., "<unused95>").
    Does NOT consume adjacent alphanumerics, because models sometimes emit tokens
    immediately followed by a fact prefix (e.g., "<unused95>DEMOGRAPHICS|..."),
    and we must not delete "DEMOGRAPHICS".
    """
    if not text:
        return ""

    # Remove internal tokens
    cleaned = _MEDGEMMA_INTERNAL_TOKEN_RE.sub("", text)

    # Drop explicit thought lines and thinking blocks
    lines = []
    in_thinking_block = False
    for line in cleaned.splitlines():
        line_lower = line.lower().strip()

        # Check if entering thinking mode
        if any(marker in line_lower for marker in ['thought', 'the user wants', 'here\'s my plan', 'input:', 'output:', 'constraints:']):
            in_thinking_block = True
            continue

        # Check if exiting thinking mode (actual fact payload starts).
        # Support both KVT4 lines and JSON-like payload fragments.
        if in_thinking_block:
            looks_like_kvt4 = ('|' in line and line.count('|') >= 3)
            looks_like_json_payload = (
                '{"K"' in line
                or '"facts"' in line
                or line.lstrip().startswith("{")
                or line.strip().startswith("```json")
            )
            if looks_like_kvt4 or looks_like_json_payload:
                in_thinking_block = False

        # Skip if in thinking block
        if in_thinking_block:
            continue

        # Skip thought lines
        if _MEDGEMMA_THOUGHT_LINE_RE.match(line):
            continue

        lines.append(line)

    return "\n".join(lines)


def _looks_like_kvt_fact(line: str) -> bool:
    """Validate if a line looks like a valid KVT4 fact.
    
    Expects either:
    - 4-part: CLUSTER|Keyword|Value|Timestamp  (preferred)
    - 3-part: Keyword|Value|Timestamp         (legacy)
    """
    if not line:
        return False
    s = line.strip()
    if len(s) < 5 or len(s) > 400:
        return False
    
    pipe_count = s.count("|")
    if pipe_count not in (2, 3):
        return False
    parts = [p.strip() for p in s.split("|")]
    if len(parts) not in (3, 4):
        return False

    # Keep parser permissive by default (unit tests expect 3-part legacy facts too).
    allow_kvt3 = str(os.getenv("ALLOW_KVT3", "1")).strip() == "1"
    if len(parts) == 3 and not allow_kvt3:
        return False
    parts_lower = [p.lower() for p in parts]

    # Filter common headers / schema lines.
    if parts_lower == ["k", "v", "t"]:
        return False
    if parts_lower == ["category", "keyword", "value", "timestamp"]:
        return False
    if (
        len(parts_lower) == 4
        and parts_lower[0].startswith("category")
        and parts_lower[1] == "keyword"
        and parts_lower[2] == "value"
        and parts_lower[3].startswith("timestamp")
    ):
        return False
    if "format" in parts_lower[0] and parts_lower[1:3] == ["keyword", "value"]:
        return False
    if parts_lower[0].startswith(("format", "output format")) and "timestamp" in parts_lower[-1]:
        return False

    # Filter instruction lines
    if any(marker in parts_lower[1] for marker in ['any diagnosis', 'any symptom', 'any procedure', 'value:']):
        return False
    if '(' in parts[1] and ')' in parts[1]:
        return False

    # Length heuristics to avoid capturing prose with incidental pipes.
    if len(parts[0]) > 80 or len(parts[1]) > 80 or len(parts[2]) > 200:
        return False
    if len(parts) == 4 and len(parts[3]) > 40:
        return False

    # Word-count heuristics: KVT lines are short phrases, not full sentences.
    w0 = len(parts[0].split())
    w1 = len(parts[1].split())
    w2 = len(parts[2].split())
    if w0 > 8 or w1 > 8 or w2 > 14:
        return False
    if len(parts) == 4 and len(parts[3].split()) > 4:
        return False
    return all(parts)


def _normalize_kvt_fact(line: str) -> str:
    """Normalize a KVT fact line by stripping whitespace and quotes."""
    parts = [p.strip().strip("«»\"'") for p in line.strip().split("|")]
    return "|".join(parts)


def _map_category_to_cluster(category: str) -> str:
    """Map category aliases to canonical cluster names."""
    c = (category or "").strip().lower()
    if not c:
        return ""
    mapping = {
        "vitals": "VITALS",
        "vital": "VITALS",
        "labs": "LABS",
        "lab": "LABS",
        "demographics": "DEMOGRAPHICS",
        "demo": "DEMOGRAPHICS",
        "conditions": "PROBLEMS",
        "condition": "PROBLEMS",
        "problems": "PROBLEMS",
        "problem": "PROBLEMS",
        "symptoms": "SYMPTOMS",
        "symptom": "SYMPTOMS",
        "medications": "MEDICATIONS",
        "medication": "MEDICATIONS",
        "procedures": "PROCEDURES",
        "procedure": "PROCEDURES",
        "utilization": "UTILIZATION",
        "disposition": "DISPOSITION",
    }
    return mapping.get(c, category.strip())


def _infer_cluster_from_keyword(keyword: str) -> str:
    """Infer cluster from keyword using canonical lists."""
    k = (keyword or "").strip()
    if not k:
        return ""
    if k in CANONICAL_VITALS:
        return "VITALS"
    if k in CANONICAL_LABS:
        return "LABS"
    if k in CANONICAL_DEMOGRAPHICS:
        return "DEMOGRAPHICS"
    # Minimal readmission-fixed keywords
    if k in {"Prior Admissions 12mo", "ED Visits 6mo", "Days Since Last Admission", "Current Length of Stay"}:
        return "UTILIZATION"
    if k in {"Discharge Disposition", "Mental Status"}:
        return "DISPOSITION"
    if k in {"Any Procedure", "Surgery", "Dialysis", "Mechanical Ventilation"}:
        return "PROCEDURES"
    if k in {"Medication Count", "New Medications Count", "Polypharmacy", "Anticoagulation", "Insulin Therapy", "Opioid Therapy", "Diuretic Therapy"}:
        return "MEDICATIONS"
    return ""


def _kvt4_from_fact_dict(d: dict) -> Optional[str]:
    """Convert structured fact dict into CLUSTER|Keyword|Value|Timestamp."""
    if not isinstance(d, dict):
        return None

    def _first_present(*keys: str):
        for key in keys:
            if key in d and d[key] is not None:
                return d[key]
        return None

    # Accept multiple key spellings
    cluster = _first_present("cluster", "Cluster", "C", "category", "Category")
    keyword = _first_present("keyword", "Keyword", "K")
    value = _first_present("value", "Value", "V")
    timestamp = _first_present("timestamp", "Timestamp", "T")

    keyword_s = str(keyword).strip() if keyword is not None else ""
    value_s = str(value).strip() if value is not None else ""
    timestamp_s = str(timestamp).strip() if timestamp is not None else ""

    cluster_s = str(cluster).strip() if cluster is not None else ""
    cluster_s = _map_category_to_cluster(cluster_s)

    if not cluster_s:
        cluster_s = _infer_cluster_from_keyword(keyword_s)
    if not cluster_s:
        cluster_s = "UNKNOWN"

    if not keyword_s or not value_s:
        return None
    if not timestamp_s:
        timestamp_s = "Unknown"

    return f"{cluster_s}|{keyword_s}|{value_s}|{timestamp_s}"


def _fact_dict_has_explicit_cluster(d: dict) -> bool:
    if not isinstance(d, dict):
        return False
    for key in ("cluster", "Cluster", "CLUSTER", "C", "category", "Category"):
        v = d.get(key)
        if v is not None and str(v).strip():
            return True
    return False


def extract_kvt_fact_lines(text: str) -> List[str]:
    """
    Extract candidate K|V|T / Category|K|V|T lines from arbitrary model output.

    Handles common formats:
    - Plain pipe-delimited lines
    - DSPy-rendered lists like: [1] «Vitals|Temperature|37.2°C|20240110»
    - JSON objects/lists containing "facts"
    - Python literal lists of strings
    """
    if not text:
        return []

    cleaned = strip_medgemma_internal_tokens(text).strip()
    if not cleaned:
        return []

    candidates: List[str] = []
    structured_extracted = False

    def _dedupe_preserve_order(items: List[str]) -> List[str]:
        out: List[str] = []
        seen = set()
        for it in items:
            if it not in seen:
                seen.add(it)
                out.append(it)
        return out

    def add_fact(s: str) -> None:
        s2 = _normalize_kvt_fact(s)
        if _looks_like_kvt_fact(s2):
            candidates.append(s2)

    def _map_category_to_cluster(category: str) -> str:
        c = (category or "").strip().lower()
        if not c:
            return ""
        mapping = {
            "vitals": "VITALS",
            "vital": "VITALS",
            "labs": "LABS",
            "lab": "LABS",
            "demographics": "DEMOGRAPHICS",
            "demo": "DEMOGRAPHICS",
            "conditions": "PROBLEMS",
            "condition": "PROBLEMS",
            "problems": "PROBLEMS",
            "problem": "PROBLEMS",
            "symptoms": "SYMPTOMS",
            "symptom": "SYMPTOMS",
            "medications": "MEDICATIONS",
            "medication": "MEDICATIONS",
            "procedures": "PROCEDURES",
            "procedure": "PROCEDURES",
            "utilization": "UTILIZATION",
            "disposition": "DISPOSITION",
        }
        return mapping.get(c, category.strip())

    def _infer_cluster_from_keyword(keyword: str) -> str:
        k = (keyword or "").strip()
        if not k:
            return ""
        if k in CANONICAL_VITALS:
            return "VITALS"
        if k in CANONICAL_LABS:
            return "LABS"
        if k in CANONICAL_DEMOGRAPHICS:
            return "DEMOGRAPHICS"
        # Minimal readmission-fixed keywords (from prompts/prompt.py ontology).
        if k in {"Prior Admissions 12mo", "ED Visits 6mo", "Days Since Last Admission", "Current Length of Stay"}:
            return "UTILIZATION"
        if k in {"Discharge Disposition", "Mental Status"}:
            return "DISPOSITION"
        if k in {"Any Procedure", "Surgery", "Dialysis", "Mechanical Ventilation"}:
            return "PROCEDURES"
        if k in {"Medication Count", "New Medications Count", "Polypharmacy", "Anticoagulation", "Insulin Therapy", "Opioid Therapy", "Diuretic Therapy"}:
            return "MEDICATIONS"
        return ""

    def _kvt4_from_fact_dict(d: dict) -> Optional[str]:
        """Convert common structured fact dicts into CLUSTER|Keyword|Value|Timestamp."""
        if not isinstance(d, dict):
            return None

        def _first_present(*keys: str):
            for key in keys:
                if key in d and d[key] is not None:
                    return d[key]
            return None

        # Accept multiple key spellings (legacy + short keys).
        cluster = _first_present("cluster", "Cluster", "CLUSTER", "C", "category", "Category")
        keyword = _first_present("keyword", "Keyword", "KEYWORD", "K")
        value = _first_present("value", "Value", "VALUE", "V")
        timestamp = _first_present("timestamp", "Timestamp", "TIMESTAMP", "T")

        keyword_s = str(keyword).strip() if keyword is not None else ""
        value_s = str(value).strip() if value is not None else ""
        timestamp_s = str(timestamp).strip() if timestamp is not None else ""
        # Drop prompt-template placeholders that are not real facts.
        if keyword_s.casefold() in {"keyword", "k"} and value_s.casefold() in {"value", "v"}:
            if timestamp_s.casefold() in {"timestamp", "t", "unknown", "admission", "discharge", "past"}:
                return None

        cluster_s = str(cluster).strip() if cluster is not None else ""
        # If we got a "category" like "vitals/labs", map it into prompt-style clusters.
        cluster_s = _map_category_to_cluster(cluster_s)

        if not cluster_s:
            cluster_s = _infer_cluster_from_keyword(keyword_s)
        if not cluster_s:
            cluster_s = "UNKNOWN"

        if not keyword_s or not value_s:
            return None
        if not timestamp_s:
            timestamp_s = "Unknown"

        return f"{cluster_s}|{keyword_s}|{value_s}|{timestamp_s}"

    def _kvt4_lines_from_grouped_obj(obj: dict) -> List[str]:
        """Convert grouped JSON object into KVT4 lines.

        Supported layout:
        {
          "LABS":[{"K":"Creatinine","V":1.2,"T":"Discharge"}],
          "PROBLEMS":[{"K":"Hypertension","V":"chronic","T":"Past"}]
        }
        """
        if not isinstance(obj, dict):
            return []

        out_lines: List[str] = []
        for raw_cluster, raw_entries in obj.items():
            cluster_norm = _map_category_to_cluster(str(raw_cluster).strip())
            cluster_upper = cluster_norm.upper()
            if cluster_upper not in READMISSION_CLUSTERS:
                continue

            entries: List[dict] = []
            if isinstance(raw_entries, list):
                entries = [it for it in raw_entries if isinstance(it, dict)]
            elif isinstance(raw_entries, dict):
                entries = [raw_entries]
            else:
                continue

            for ent in entries:
                keyword = ent["K"] if "K" in ent else ent.get("keyword", ent.get("Keyword"))
                value = ent["V"] if "V" in ent else ent.get("value", ent.get("Value"))
                timestamp = ent["T"] if "T" in ent else ent.get("timestamp", ent.get("Timestamp"))
                fact_obj = {
                    "cluster": cluster_upper,
                    "keyword": keyword,
                    "value": value,
                    "timestamp": timestamp,
                }
                ln = _kvt4_from_fact_dict(fact_obj)
                if ln:
                    out_lines.append(ln)
        return out_lines

    # 1) JSON / Python list attempts (whole string + best-effort substrings)
    json_like = cleaned
    substrings: List[str] = [json_like]
    first_obj = json_like.find("{")
    last_obj = json_like.rfind("}")
    if first_obj != -1 and last_obj != -1 and last_obj > first_obj:
        substrings.append(json_like[first_obj : last_obj + 1])
    first_arr = json_like.find("[")
    last_arr = json_like.rfind("]")
    if first_arr != -1 and last_arr != -1 and last_arr > first_arr:
        substrings.append(json_like[first_arr : last_arr + 1])

    cleaned_strip = cleaned.strip()
    for s in list(dict.fromkeys(substrings)):
        s_strip = s.strip()
        if not s_strip:
            continue
        is_derived_array_substring = (
            s_strip.startswith("[") and s_strip.endswith("]") and s_strip != cleaned_strip
        )
        try:
            before = len(candidates)
            obj = json.loads(s_strip)
            if isinstance(obj, dict):
                facts = obj.get("facts")
                if isinstance(facts, list):
                    for it in facts:
                        if isinstance(it, str):
                            add_fact(it)
                        elif isinstance(it, dict):
                            ln = _kvt4_from_fact_dict(it)
                            if ln:
                                add_fact(ln)
                else:
                    grouped_lines = _kvt4_lines_from_grouped_obj(obj)
                    if grouped_lines:
                        for ln in grouped_lines:
                            add_fact(ln)
                        continue
                    # Sometimes the whole object is a single fact dict.
                    ln = _kvt4_from_fact_dict(obj)
                    if ln:
                        add_fact(ln)
            elif isinstance(obj, list):
                for it in obj:
                    if isinstance(it, str):
                        add_fact(it)
                    elif isinstance(it, dict):
                        # Avoid duplicate UNKNOWN facts when a grouped JSON object is
                        # also parsed via its inner array substring (cluster context lost).
                        if is_derived_array_substring and not _fact_dict_has_explicit_cluster(it):
                            continue
                        ln = _kvt4_from_fact_dict(it)
                        if ln:
                            add_fact(ln)
            if len(candidates) > before:
                structured_extracted = True
        except Exception:
            pass

        try:
            before = len(candidates)
            obj = ast.literal_eval(s_strip)
            if isinstance(obj, dict):
                facts = obj.get("facts") if isinstance(obj.get("facts"), list) else None
                if facts is not None:
                    for it in facts:
                        if isinstance(it, str):
                            add_fact(it)
                        elif isinstance(it, dict):
                            ln = _kvt4_from_fact_dict(it)
                            if ln:
                                add_fact(ln)
                else:
                    grouped_lines = _kvt4_lines_from_grouped_obj(obj)
                    if grouped_lines:
                        for ln in grouped_lines:
                            add_fact(ln)
                        continue
                    ln = _kvt4_from_fact_dict(obj)
                    if ln:
                        add_fact(ln)
            elif isinstance(obj, list):
                for it in obj:
                    if isinstance(it, str):
                        add_fact(it)
                    elif isinstance(it, dict):
                        if is_derived_array_substring and not _fact_dict_has_explicit_cluster(it):
                            continue
                        ln = _kvt4_from_fact_dict(it)
                        if ln:
                            add_fact(ln)
            if len(candidates) > before:
                structured_extracted = True
        except Exception:
            pass

    # If we already extracted structured facts, do not run heuristic recovery
    # branches below (they may introduce noisy duplicates on valid JSON payloads).
    if structured_extracted and candidates:
        return _dedupe_preserve_order(candidates)

    # 1b) Partial/truncated JSON recovery:
    # If the model output is cut mid-stream, json.loads fails even when many
    # complete fact objects were already emitted. Recover those complete objects.
    for m in _PARTIAL_JSON_FACT_RE.finditer(cleaned):
        c, k, v, t = [x.strip() for x in m.groups()]
        if c and k and v and t:
            add_fact(f"{c}|{k}|{v}|{t}")

    # 1c) Partial/truncated grouped JSON recovery:
    # Recover complete {"K","V","T"} entries within each cluster block even when
    # root JSON is truncated and json.loads fails.
    cluster_hits = list(_PARTIAL_GROUPED_CLUSTER_BLOCK_RE.finditer(cleaned))
    if cluster_hits:
        for idx, hit in enumerate(cluster_hits):
            cluster = str(hit.group("cluster") or "").strip().upper()
            block_start = hit.end()
            block_end = cluster_hits[idx + 1].start() if idx + 1 < len(cluster_hits) else len(cleaned)
            block = cleaned[block_start:block_end]
            for item in _PARTIAL_GROUPED_ITEM_RE.finditer(block):
                k = str(item.group("k") or "").strip()
                t = str(item.group("t") or "").strip()
                v_tok = str(item.group("v") or "").strip()
                if not k or not t:
                    continue
                if v_tok.startswith('"') and v_tok.endswith('"') and len(v_tok) >= 2:
                    v = v_tok[1:-1]
                elif v_tok.casefold() in {"true", "false"}:
                    v = v_tok.casefold()
                else:
                    v = v_tok
                if v:
                    add_fact(f"{cluster}|{k}|{v}|{t}")

    # 2) Extract between DSPy quotes «...»
    for m in _DSPY_QUOTED_FACT_RE.finditer(cleaned):
        inner = m.group(1).strip()
        if "|" in inner:
            add_fact(inner)

    # 2b) Narrative markdown recovery.
    # Some small models emit facts as multi-line markdown blocks:
    #   **CLUSTER:** DEMOGRAPHICS
    #   **Keyword:** Sex
    #   **Value:** male
    #   **Timestamp:** Admission
    # Recover these into KVT4 lines.
    _narrative_kv_re = re.compile(
        r"\*{0,2}(cluster|keyword|value|timestamp)\s*:?\s*\*{0,2}\s*(.+)",
        re.IGNORECASE,
    )
    cur: dict = {}
    for line in cleaned.splitlines():
        m = _narrative_kv_re.match(line.strip())
        if not m:
            continue
        field = m.group(1).strip().lower()
        val = m.group(2).strip().strip("*").strip()
        if field == "cluster":
            if cur.get("cluster") and cur.get("keyword") and cur.get("value"):
                ts = cur.get("timestamp", "Unknown")
                add_fact(f"{cur['cluster']}|{cur['keyword']}|{cur['value']}|{ts}")
            cur = {"cluster": val}
        elif field == "keyword":
            # Flush previous fact within the same cluster before starting a new keyword
            if cur.get("cluster") and cur.get("keyword") and cur.get("value"):
                ts = cur.get("timestamp", "Unknown")
                add_fact(f"{cur['cluster']}|{cur['keyword']}|{cur['value']}|{ts}")
            cluster_keep = cur.get("cluster", "")
            cur = {"cluster": cluster_keep, "keyword": val}
        elif field in ("value", "timestamp"):
            cur[field] = val
    # flush last accumulated fact
    if cur.get("cluster") and cur.get("keyword") and cur.get("value"):
        ts = cur.get("timestamp", "Unknown")
        add_fact(f"{cur['cluster']}|{cur['keyword']}|{cur['value']}|{ts}")

    # 2c) Cluster-heading + inline JSON item recovery.
    # Some models emit planning text like:
    #   * **VITALS:**
    #     ... -> {"K":"Heart Rate","V":54,"T":"Admission"}
    # Recover such entries by tracking the current cluster heading.
    heading_re = re.compile(r"\*{0,2}\s*([A-Z][A-Z ]{2,})\s*:\s*\*{0,2}\s*$")
    cluster_inline_re = re.compile(
        r"\b(DEMOGRAPHICS|VITALS|LABS|PROBLEMS|SYMPTOMS|MEDICATIONS|PROCEDURES|UTILIZATION|DISPOSITION)\b",
        re.IGNORECASE,
    )
    item_re = re.compile(
        r'\{\s*"K"\s*:\s*"(?P<k>[^"]+)"\s*,\s*"V"\s*:\s*(?P<v>"[^"]*"|-?\d+(?:\.\d+)?|true|false)\s*,\s*"T"\s*:\s*"(?P<t>Past|Admission|Discharge|Unknown)"\s*\}',
        re.IGNORECASE,
    )
    cur_cluster = ""
    for raw_line in cleaned.splitlines():
        line = raw_line.strip()
        if not line:
            continue
        # Avoid cluster-bleed on compact one-line JSON objects:
        # grouped payloads should be handled by structured parsing above.
        if line.startswith("{") or line.startswith("["):
            continue

        # Accept headings like "**VITALS:**", "VITALS:", "*   **VITALS:**"
        norm = re.sub(r"^[*•\-\s]+", "", line)
        norm = norm.strip("* ").strip()
        hm = heading_re.match(norm)
        if hm:
            c_raw = hm.group(1).strip().upper()
            c_norm = _map_category_to_cluster(c_raw)
            c_up = str(c_norm).strip().upper()
            if c_up in READMISSION_CLUSTERS:
                cur_cluster = c_up
            continue

        # Inline headings like:
        # "- **VITALS:** ... -> {\"K\":\"Heart Rate\",...}"
        cm = cluster_inline_re.search(norm)
        if cm:
            c_raw = cm.group(1).strip().upper()
            c_norm = _map_category_to_cluster(c_raw)
            c_up = str(c_norm).strip().upper()
            if c_up in READMISSION_CLUSTERS:
                cur_cluster = c_up

        if not cur_cluster:
            continue
        for m in item_re.finditer(line):
            k = str(m.group("k") or "").strip()
            v_tok = str(m.group("v") or "").strip()
            t = str(m.group("t") or "").strip()
            if not k or not t:
                continue
            if v_tok.startswith('"') and v_tok.endswith('"') and len(v_tok) >= 2:
                v = v_tok[1:-1]
            else:
                v = v_tok.casefold() if v_tok.casefold() in {"true", "false"} else v_tok
            if v:
                add_fact(f"{cur_cluster}|{k}|{v}|{t}")

    # 3) Line-by-line heuristics (bullets / numbering / quoted JSON fragments)
    for line in cleaned.splitlines():
        s = line.strip()
        if not s:
            continue
        s = re.sub(r"^\[\d+\]\s*", "", s)
        s = re.sub(r"^[-*•]\s*", "", s)
        s = s.strip().strip("«»\"'")
        s = s.rstrip(",")
        if "|" in s:
            add_fact(s)

    # De-duplicate while preserving order
    return _dedupe_preserve_order(candidates)

def normalize_readmission_kvt4_lines(lines: List[str]) -> tuple[List[str], dict]:
    """Normalize KVT4 lines into canonical READMISSION_MVP form.

    Goals:
    - Boost strict-format usability by deterministic canonicalization
    - Reduce drift (Blood Pressure -> SBP/DBP, Oxygen Saturation -> SpO2, etc.)
    - Enforce numeric-only values for VITALS/LABS (+ known numeric fields)
    - Enforce at most one line per (CLUSTER, Keyword) via timestamp-priority dedupe

    Returns: (normalized_lines, stats)
    """

    def _parse_line(line: str) -> Optional[tuple[str, str, str, str]]:
        if not isinstance(line, str):
            return None
        s = line.strip()
        if s.count("|") != 3:
            return None
        parts = [p.strip() for p in s.split("|")]
        if len(parts) != 4:
            return None
        c, k, v, t = parts
        if not c or not k or not v:
            return None
        return c, k, v, t or "Unknown"

    def _normalize_timestamp(t: str) -> str:
        tt = (t or "").strip()
        if tt in {"Admission", "Discharge", "Past", "Unknown"}:
            return tt
        return "Unknown"

    def _fill_unknown_timestamp(cluster: str, keyword: str, value: str) -> str:
        """Best-effort timestamp fill for strict-eval stability.

        Policy is ontology-driven (not note-section heuristics):
        - DEMOGRAPHICS/VITALS/LABS/SYMPTOMS/MEDICATIONS/PROCEDURES: Admission
        - DISPOSITION: Discharge
        - UTILIZATION: Past
        - PROBLEMS: Past if chronic, Discharge if acute, else Past
        """
        c = (cluster or "").strip().upper()
        v = (value or "").strip().lower()

        if c == "DISPOSITION":
            return "Discharge"
        if c == "UTILIZATION":
            return "Past"
        if c == "PROBLEMS":
            if v == "acute":
                return "Discharge"
            if v == "chronic":
                return "Past"
            # Default: history-like framing
            return "Past"
        if c in {"DEMOGRAPHICS", "VITALS", "LABS", "SYMPTOMS", "MEDICATIONS", "PROCEDURES"}:
            return "Admission"
        return "Admission"

    def _first_number(value: str) -> Optional[str]:
        m = re.search(r"-?\d+(?:\.\d+)?", value or "")
        return m.group(0) if m else None

    # Keyword aliases (strict clusters).
    vital_alias = {
        "HR": "Heart Rate",
        "Pulse": "Heart Rate",
        "Temp": "Temperature",
        "O2 Sat": "SpO2",
        "Oxygen Saturation": "SpO2",
        "SpO2": "SpO2",
        "Resp": "Respiratory Rate",
        "RR": "Respiratory Rate",
        "Blood Pressure": "Blood Pressure",  # special-case splitter
        "BP": "Blood Pressure",
        "Systolic": "Systolic BP",
        "Diastolic": "Diastolic BP",
        "SBP": "Systolic BP",
        "DBP": "Diastolic BP",
    }
    lab_alias = {
        "Hgb": "Hemoglobin",
        "Hct": "Hematocrit",
        "Plt": "Platelet",
        "Platelets": "Platelet",
        "Na": "Sodium",
        "K": "Potassium",
        "Cr": "Creatinine",
        "HCO3": "Bicarbonate",
        "Bicarb": "Bicarbonate",
        "WBC": "WBC",
        "BUN": "BUN",
    }
    sex_alias = {"m": "male", "male": "male", "f": "female", "female": "female"}

    # Dedupe priority can be configured per mode.
    # For full readmission feature set we generally care about discharge/most-recent.
    ts_priority = [s.strip() for s in os.getenv("MEDGEMMA_TIMESTAMP_PRIORITY", "Discharge,Admission,Past,Unknown").split(",") if s.strip()]
    ts_rank = {t: i for i, t in enumerate(ts_priority)}

    stats = {
        "input_lines": len(lines or []),
        "parsed_kvt4": 0,
        "dropped_placeholders": 0,
        "dropped_noncanonical": 0,
        "dropped_by_allowed_clusters": 0,
        "expanded_bp": 0,
        "dedup_dropped": 0,
        "output_lines": 0,
        "canonical_keyword_rate_strict": None,
        "numeric_only_rate_vitals_labs": None,
        "duplicates_after_dedup": 0,
    }

    allowed_clusters_env = os.getenv("MEDGEMMA_ALLOWED_CLUSTERS", "").strip()
    allowed_clusters = None
    if allowed_clusters_env:
        allowed_clusters = {c.strip().upper() for c in allowed_clusters_env.split(",") if c.strip()}

    # First pass: normalize + expand BP
    normalized_candidates: List[tuple[str, str, str, str]] = []
    fill_unknown = os.getenv("MEDGEMMA_TIMESTAMP_FILL_UNKNOWN", "1").strip().lower() in {"1", "true", "yes"}
    for line in lines or []:
        parsed = _parse_line(line)
        if not parsed:
            continue
        c, k, v, t = parsed
        stats["parsed_kvt4"] += 1

        c_up = str(c).strip().upper()
        if allowed_clusters is not None and c_up not in allowed_clusters:
            stats["dropped_by_allowed_clusters"] += 1
            continue
        t_norm = _normalize_timestamp(t)
        k_norm = k.strip()
        v_norm = v.strip()

        # Drop obvious placeholders
        if v_norm in {"___", "__", "_", "N/A", "NA", "null", "None"}:
            stats["dropped_placeholders"] += 1
            continue

        # Cluster-specific normalization
        if c_up == "DEMOGRAPHICS":
            if k_norm == "Sex":
                vv = sex_alias.get(v_norm.strip().lower())
                if not vv:
                    stats["dropped_noncanonical"] += 1
                    continue
                v_norm = vv
            elif k_norm == "Age":
                num = _first_number(v_norm)
                if not num:
                    stats["dropped_noncanonical"] += 1
                    continue
                v_norm = num

        elif c_up == "VITALS":
            k_norm = vital_alias.get(k_norm, k_norm)
            if k_norm == "Blood Pressure":
                # Expand 120/80 -> SBP + DBP
                m = re.search(r"(\d+(?:\.\d+)?)\s*/\s*(\d+(?:\.\d+)?)", v_norm)
                if not m:
                    stats["dropped_noncanonical"] += 1
                    continue
                sbp, dbp = m.group(1), m.group(2)
                normalized_candidates.append(("VITALS", "Systolic BP", sbp, t_norm))
                normalized_candidates.append(("VITALS", "Diastolic BP", dbp, t_norm))
                stats["expanded_bp"] += 1
                continue

            # Enforce numeric-only for vitals
            num = _first_number(v_norm)
            if not num:
                stats["dropped_noncanonical"] += 1
                continue
            v_norm = num

        elif c_up == "LABS":
            k_norm = lab_alias.get(k_norm, k_norm)
            num = _first_number(v_norm)
            if not num:
                stats["dropped_noncanonical"] += 1
                continue
            v_norm = num

        elif c_up == "UTILIZATION":
            num = _first_number(v_norm)
            if not num:
                stats["dropped_noncanonical"] += 1
                continue
            v_norm = num

        elif c_up == "MEDICATIONS":
            if k_norm in {"Medication Count", "New Medications Count"}:
                num = _first_number(v_norm)
                if not num:
                    stats["dropped_noncanonical"] += 1
                    continue
                v_norm = num
            elif k_norm in {"Polypharmacy", "Anticoagulation", "Insulin Therapy", "Opioid Therapy", "Diuretic Therapy"}:
                vv = v_norm.strip().lower()
                if vv in {"yes", "y", "true", "1"}:
                    v_norm = "yes"
                elif vv in {"no", "n", "false", "0"}:
                    v_norm = "no"
                else:
                    stats["dropped_noncanonical"] += 1
                    continue

        elif c_up == "PROCEDURES":
            if k_norm in {"Any Procedure", "Surgery"}:
                vv = v_norm.strip().lower()
                if vv in {"yes", "y", "true", "1"}:
                    v_norm = "yes"
                elif vv in {"no", "n", "false", "0"}:
                    v_norm = "no"
                else:
                    stats["dropped_noncanonical"] += 1
                    continue
            elif k_norm == "Dialysis":
                vv = v_norm.strip().lower()
                allowed = {"decided", "started", "done", "cancelled", "no"}
                if vv not in allowed:
                    stats["dropped_noncanonical"] += 1
                    continue
                v_norm = vv
            elif k_norm == "Mechanical Ventilation":
                vv = v_norm.strip().lower()
                if "no" == vv:
                    v_norm = "no"
                else:
                    num = _first_number(v_norm)
                    if not num:
                        stats["dropped_noncanonical"] += 1
                        continue
                    v_norm = num

        elif c_up == "DISPOSITION":
            if k_norm == "Discharge Disposition":
                vv = v_norm.strip().lower()
                # Normalize into the prompt enums.
                if "home with" in vv or "home w" in vv or "services" in vv:
                    v_norm = "Home with Services"
                elif vv == "home" or vv.startswith("home "):
                    v_norm = "Home"
                elif "snf" in vv or "skilled nursing" in vv:
                    v_norm = "SNF"
                elif "rehab" in vv:
                    v_norm = "Rehab"
                elif "ltac" in vv:
                    v_norm = "LTAC"
                elif "hospice" in vv:
                    v_norm = "Hospice"
                elif "ama" in vv or "against medical advice" in vv:
                    v_norm = "AMA"
                else:
                    stats["dropped_noncanonical"] += 1
                    continue
            elif k_norm == "Mental Status":
                vv = v_norm.strip().lower()
                if "confus" in vv:
                    v_norm = "confused"
                elif "letharg" in vv:
                    v_norm = "lethargic"
                elif "alert" in vv:
                    v_norm = "alert"
                elif "orient" in vv:
                    v_norm = "oriented"
                else:
                    stats["dropped_noncanonical"] += 1
                    continue

        elif c_up == "PROBLEMS":
            vv = re.sub(r"\s+", " ", v_norm.strip().lower())
            if vv in {"chronic", "acute", "exist", "not exist"}:
                v_norm = vv
            elif vv in {"past", "history", "historical", "pmh", "chronic condition", "chronic disease"}:
                v_norm = "chronic"
            elif vv in {"discharge", "discharged", "active", "current"}:
                v_norm = "acute"
            elif vv in {"present", "yes", "true", "1", "positive", "confirmed", "exists"}:
                v_norm = "exist"
            elif vv in {"no", "none", "false", "0", "absent", "negative", "not present", "ruled out"}:
                v_norm = "not exist"
            else:
                stats["dropped_noncanonical"] += 1
                continue

        elif c_up == "SYMPTOMS":
            vv = re.sub(r"\s+", " ", v_norm.strip().lower())
            if vv in {"yes", "no", "severe"}:
                v_norm = vv
            elif vv in {"present", "positive", "true", "1", "y", "symptomatic"}:
                v_norm = "yes"
            elif vv in {"none", "absent", "negative", "false", "0", "n", "denied", "denies"}:
                v_norm = "no"
            elif "severe" in vv or vv in {"marked", "significant"}:
                v_norm = "severe"
            else:
                stats["dropped_noncanonical"] += 1
                continue

        # Drop non-canonical keywords for strict clusters (objective ones).
        if c_up in STRICT_KEYWORDS_READMISSION:
            if k_norm not in STRICT_KEYWORDS_READMISSION[c_up]:
                stats["dropped_noncanonical"] += 1
                continue

        if fill_unknown and t_norm == "Unknown":
            t_norm = _fill_unknown_timestamp(c_up, k_norm, v_norm)

        normalized_candidates.append((c_up, k_norm, v_norm, t_norm))

    # Second pass: dedupe by (CLUSTER, Keyword) using timestamp priority.
    best: dict[tuple[str, str], tuple[str, str, str, str]] = {}
    for c, k, v, t in normalized_candidates:
        key = (c, k)
        cur = best.get(key)
        if cur is None:
            best[key] = (c, k, v, t)
            continue
        _, _, _, t_prev = cur
        r_new = ts_rank.get(t, 999)
        r_prev = ts_rank.get(t_prev, 999)
        if r_new < r_prev:
            best[key] = (c, k, v, t)
        else:
            stats["dedup_dropped"] += 1

    out_lines = [f"{c}|{k}|{v}|{t}" for (c, k), (c, k, v, t) in best.items()]
    out_lines.sort(key=lambda s: (s.split("|")[0], s.split("|")[1]))

    # Metrics: canonical + numeric-only compliance for VITALS/LABS.
    strict_total = 0
    strict_ok = 0
    vitlab_total = 0
    vitlab_numeric = 0
    key_counts: dict[tuple[str, str], int] = {}
    for ln in out_lines:
        parsed = _parse_line(ln)
        if not parsed:
            continue
        c, k, v, _t = parsed
        key_counts[(c, k)] = key_counts.get((c, k), 0) + 1
        if c in STRICT_KEYWORDS_READMISSION:
            strict_total += 1
            if k in STRICT_KEYWORDS_READMISSION[c]:
                strict_ok += 1
        if c in {"VITALS", "LABS"}:
            vitlab_total += 1
            if re.fullmatch(r"-?\d+(?:\.\d+)?", v.strip()):
                vitlab_numeric += 1

    stats["duplicates_after_dedup"] = sum(1 for cnt in key_counts.values() if cnt > 1)
    stats["output_lines"] = len(out_lines)
    stats["canonical_keyword_rate_strict"] = (strict_ok / strict_total) if strict_total else 1.0
    stats["numeric_only_rate_vitals_labs"] = (vitlab_numeric / vitlab_total) if vitlab_total else 1.0

    return out_lines, stats


def _normalize_mode(mode: Optional[str]) -> str:
    """Normalize mode string to canonical format."""
    if not mode:
        return "READMISSION_DISCHARGE"
    mode = mode.upper().replace("-", "_")
    if mode in {"CCDE", "CCDE_ADMISSION"}:
        return "CCDE_ADMISSION"
    elif mode in {"TABULAR", "READMISSION_TABULAR", "MVP_TABULAR", "TOON_TABULAR"}:
        return "READMISSION_TABULAR"
    elif mode in {"STRUCTURED", "READMISSION_STRUCTURED", "PYDANTIC", "STRUCTURED_OUTPUT"}:
        return "READMISSION_STRUCTURED"
    else:
        return "READMISSION_DISCHARGE"


# =============================================================================
# CUSTOM DSPy ADAPTER FOR MEDGEMMA (Local Transformers)
# =============================================================================