Stop the slugifier from deleting the letter D-stroke

ingestion/ingestion/segment/assembler.py

@@ -92,9 +92,21 @@ _Event = Union[Heading, _SectionEvent, _TextEvent]  # Heading == a title event
 
 
 def _slugify(text: str) -> str:
-    normalized = unicodedata.normalize("NFKD", text)
+    # `đ`/`Đ` (U+0111/U+0110) are standalone Vietnamese letters, not a base
+    # letter plus a combining mark, so NFKD leaves them whole and the ASCII
+    # encode below then discards them silently -- turning `ĐIỆN GIẢI` into
+    # `ien_giai`. Exactly three monographs contain `Đ` and all three carried a
+    # damaged drug_id because of this: GIẢI ĐỘC TỐ UỐN VÁN, KHÁNG ĐỘC TỐ BẠCH
+    # HẦU, and THUỐC UỐNG BÙ NƯỚC VÀ ĐIỆN GIẢI. Found via the adversarial eval
+    # suite, where both Oresol cases failed to resolve their monograph.
+    #
+    # Casefold before replacing so one pass covers both cases: `Đ` casefolds
+    # to `đ`. This matches what `entities/catalog.py::normalize_name` already
+    # does -- that function got it right and this one did not.
+    folded = text.casefold().replace("đ", "d")
+    normalized = unicodedata.normalize("NFKD", folded)
     ascii_text = normalized.encode("ascii", "ignore").decode("ascii")
-    return re.sub(r"[^a-z0-9]+", "_", ascii_text.lower()).strip("_")
+    return re.sub(r"[^a-z0-9]+", "_", ascii_text).strip("_")
 
 
 def _starts_its_visual_line(span: Span, previous: Span | None) -> bool:

ingestion/tests/test_segment_assembler.py

@@ -3,7 +3,7 @@ from dataclasses import replace
 import pytest
 
 from ingestion.extract.models import Span
-from ingestion.segment.assembler import DuplicateDrugIdError, assemble
+from ingestion.segment.assembler import DuplicateDrugIdError, _slugify, assemble
 
 
 def _span(text, page, y0, bold=True, size=9.5, printed=None, column="left"):
@@ -449,3 +449,37 @@ def test_a_section_name_opening_its_own_line_is_still_a_heading():
     ]
     monograph = list(assemble(spans))[0]
     assert monograph.sections["chong_chi_dinh"].text == "Suy tủy nặng."
+
+
+@pytest.mark.parametrize(
+    ("name", "expected"),
+    [
+        ("THUỐC UỐNG BÙ NƯỚC VÀ ĐIỆN GIẢI", "thuoc_uong_bu_nuoc_va_dien_giai"),
+        ("KHÁNG ĐỘC TỐ BẠCH HẦU", "khang_doc_to_bach_hau"),
+        ("Điện giải", "dien_giai"),
+        ("đường huyết", "duong_huyet"),
+    ],
+)
+def test_slugify_keeps_d_with_stroke(name, expected):
+    """`đ`/`Đ` must become `d`, not disappear.
+
+    They are standalone letters (U+0111/U+0110), not a base letter plus a
+    combining mark, so NFKD leaves them whole and a following ASCII encode
+    drops them outright. That silently produced `ien_giai` from `ĐIỆN GIẢI`,
+    and all three monographs in the formulary whose names contain `Đ` carried
+    a damaged drug_id as a result.
+    """
+    assert _slugify(name) == expected
+
+
+def test_slugify_unchanged_for_names_without_d_stroke():
+    """The fix must not move any id that was already correct -- 681 of the
+    684 monographs were fine, and changing one of those would orphan its
+    chunks in the vector store."""
+    for name, expected in [
+        ("PARACETAMOL (Acetaminophen)", "paracetamol_acetaminophen"),
+        ("Acid ioxaglic", "acid_ioxaglic"),
+        ("METFORMIN", "metformin"),
+        ("Vắc xin uốn ván hấp phụ", "vac_xin_uon_van_hap_phu"),
+    ]:
+        assert _slugify(name) == expected