feat(rag): regulation_short Casing-Normalisierung am Ingest-Rand

Der Re-Ingest leitet regulation_short z.T. via title()-Casing aus Dateinamen ab
('dsgvo'->'Dsgvo', 'osha otm'->'Osha Otm') -> falsche Akronyme im Payload UND im
article_label ('Art. 37 Dsgvo'). NEU: normalize_regulation_short() in legal_metadata,
token-basiert mit kuratiertem Akronym-Set -> nur gelistete Akronyme werden gross,
legitimes Mixed-Case (GeschGehG, MuSchG, GoBD, MiCA, eIDAS, EuGH) bleibt unberuehrt.
Angewandt am Ingest-Rand in documents.py (greift fuer Payload-Feld + display_name).
+13 Tests gruen. Bestandsdaten brauchen separaten einmaligen Qdrant-Patch.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

rag-service/api/documents.py

@@ -10,7 +10,12 @@ from embedding_client import embedding_client
 from html_utils import decode_html_bytes, looks_like_html, strip_html
 from minio_client_wrapper import minio_wrapper
 from qdrant_client_wrapper import qdrant_wrapper
-from legal_metadata import build_legal_fields, compute_chunk_hash, deterministic_point_id
+from legal_metadata import (
+    build_legal_fields,
+    compute_chunk_hash,
+    deterministic_point_id,
+    normalize_regulation_short,
+)
 
 logger = logging.getLogger("rag-service.api.documents")
 
@@ -155,6 +160,13 @@ async def upload_document(
         except json.JSONDecodeError:
             logger.warning("Invalid metadata_json, ignoring")
 
+    # Casing-Normalisierung am Ingest-Rand: title-caste Akronyme korrigieren
+    # ('Dsgvo'->'DSGVO'), damit Payload-Feld UND article_label sauber sind.
+    if extra_metadata.get("regulation_short"):
+        extra_metadata["regulation_short"] = normalize_regulation_short(
+            extra_metadata["regulation_short"]
+        )
+
     # --- Build payloads (rag_reingest_spec.md §2/§3: zitierfaehige Legal-Metadaten) ---
     reg_code = (
         extra_metadata.get("regulation_code")

rag-service/legal_metadata.py

@@ -25,6 +25,40 @@ _SUB_RE = re.compile(r"(lit\.\s*[a-z]+|Satz\s*\d+|Nr\.\s*\d+)", re.IGNORECASE)
 # Zitiert wird das Aktenzeichen (display_name / regulation_short).
 _RULING_TYPES = {"urteil", "ruling", "court_decision", "beschluss"}
 
+# Akronyme, deren KORREKTE Schreibweise GROSS ist. Der Ingest leitet regulation_short
+# z.T. via title()-Casing aus Dateinamen ab ("dsgvo" -> "Dsgvo", "osha otm" -> "Osha Otm").
+# Token-basiert: NUR Tokens, deren Upper-Form hier steht, werden gross geschrieben — damit
+# bleiben legitime Mixed-Case-Kuerzel (GeschGehG, MuSchG, GoBD, MiCA, eIDAS, EuGH) unberuehrt.
+# Bewusst NICHT enthalten: EUGH/BVERFG/BVERWG (korrekt sind EuGH/BVerfG/BVerwG).
+# Konservativ + erweiterbar.
+_REG_SHORT_ACRONYMS = {
+    "DSGVO", "GDPR", "DSA", "DMA", "DORA", "AMLR", "GPSR", "DPF", "IFRS", "DSM", "MDR",
+    "CRA", "NIS2", "EU", "US", "PPE", "ICS", "SCADA", "OSHA", "OTM", "NIST", "NISTIR",
+    "ENISA", "EDPB", "EDPS", "DSK", "BFDI", "OWASP", "API", "GPAI", "EUCC", "ECCG",
+    "CISA", "CVSS", "CVD", "SRP", "PF", "SCHUFA", "SLSA", "BGH", "OGH", "BAG",
+}
+# Ganz-Wert-Overrides (Token-Logik kann diese nicht herleiten).
+_REG_SHORT_OVERRIDES = {
+    "Dataact": "Data Act",
+    "Ecommerce": "E-Commerce",
+}
+
+
+def normalize_regulation_short(name: str) -> str:
+    """Korrigiert title-caste Akronyme im druckbaren reg-short ('Dsgvo'->'DSGVO',
+    'Osha Otm ...'->'OSHA OTM ...', 'EU Mdr'->'EU MDR') und erhaelt legitimes Mixed-Case
+    (GeschGehG, MuSchG, MiCA, eIDAS, EuGH). Nur Tokens in _REG_SHORT_ACRONYMS werden
+    gross geschrieben; alles andere bleibt unveraendert."""
+    s = (name or "").strip()
+    if not s:
+        return s
+    if s in _REG_SHORT_OVERRIDES:
+        return _REG_SHORT_OVERRIDES[s]
+    return " ".join(
+        tok.upper() if tok.upper() in _REG_SHORT_ACRONYMS else tok
+        for tok in s.split()
+    )
+
 
 def detect_citation_style(section: str) -> str:
     """'§ 38' -> 'paragraph' (DE-Gesetze); 'Artikel/Art. 13' -> 'article' (EU-VO)."""

rag-service/tests/test_legal_metadata.py

@@ -113,3 +113,27 @@ def test_build_legal_fields_ruling_uses_aktenzeichen_not_paragraph():
     assert out["article"] == ""
     assert out["article_label"] == "BAG, 1 ABR 22/21"
     assert out["regulation_code"] == "BAG_1_ABR_22_21"
+
+
+def test_normalize_regulation_short():
+    # title-caste Akronyme -> korrekt GROSS (Token-basiert)
+    assert L.normalize_regulation_short("Dsgvo") == "DSGVO"
+    assert L.normalize_regulation_short("Gpsr") == "GPSR"
+    assert L.normalize_regulation_short("Ifrs") == "IFRS"
+    assert L.normalize_regulation_short("EU Mdr") == "EU MDR"
+    assert L.normalize_regulation_short("OWASP Api TOP10") == "OWASP API TOP10"
+    assert (
+        L.normalize_regulation_short("Osha Otm Section-4-safety-hazards Chapter-1")
+        == "OSHA OTM Section-4-safety-hazards Chapter-1"
+    )
+    # legitimes Mixed-Case bleibt unberuehrt (NICHT in der Akronym-Liste)
+    assert L.normalize_regulation_short("GeschGehG") == "GeschGehG"
+    assert L.normalize_regulation_short("MuSchG") == "MuSchG"
+    assert L.normalize_regulation_short("GoBD") == "GoBD"
+    assert L.normalize_regulation_short("MiCA") == "MiCA"
+    assert L.normalize_regulation_short("eIDAS") == "eIDAS"
+    assert L.normalize_regulation_short("EuGH C-252/21 (Meta)") == "EuGH C-252/21 (Meta)"
+    # Ganz-Wert-Overrides + Leerfaelle
+    assert L.normalize_regulation_short("Dataact") == "Data Act"
+    assert L.normalize_regulation_short("") == ""
+    assert L.normalize_regulation_short(None) == ""