feat(rag): Legal-Metadaten — article_label + deterministische IDs + chunk_hash

Neues pures Modul legal_metadata.py (nur stdlib, lokal+CI testbar): §3-Normalisierung
section->article, strikte Header-Extraktion (Datum/Seiten-Rauschen -> kein Falsch-Zitat),
citation_style pro Regulierung (EU/CH=article, DE=paragraph), Urteil=Aktenzeichen statt §,
camelCase-Klarnamen (ProdHaftG), deterministische uuid5-Point-ID + chunk_hash (sha256).
documents.py verdrahtet build_legal_fields in den Payload-Build + document_version.
10 Tests gruen. Vertrag: rag_reingest_spec.md (§2/§3).

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

rag-service/api/documents.py

@@ -10,6 +10,7 @@ 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
 
 logger = logging.getLogger("rag-service.api.documents")
 
@@ -49,6 +50,7 @@ async def upload_document(
     chunk_size: int = Form(default=512),
     chunk_overlap: int = Form(default=50),
     metadata_json: Optional[str] = Form(default=None),
+    document_version: str = Form(default="1"),
 ):
     """
     Upload a document:
@@ -153,36 +155,55 @@ async def upload_document(
         except json.JSONDecodeError:
             logger.warning("Invalid metadata_json, ignoring")
 
-    # --- Build payloads ---
+    # --- Build payloads (rag_reingest_spec.md §2/§3: zitierfaehige Legal-Metadaten) ---
+    reg_code = (
+        extra_metadata.get("regulation_code")
+        or extra_metadata.get("regulation_short")
+        or extra_metadata.get("regulation_id")
+        or ""
+    ).strip()
     payloads = []
+    ids = []
     for i, chunk in enumerate(chunks):
+        meta = chunks_meta[i] if i < len(chunks_meta) else {}
+        legal = build_legal_fields(
+            meta,
+            reg_code,
+            chunk,
+            citation_style=extra_metadata.get("citation_style"),
+            display_name=extra_metadata.get("regulation_short"),
+            source_type=extra_metadata.get("source_type"),
+        )
         payload = {
             "document_id": document_id,
+            "document_version": document_version,
             "object_name": object_name,
             "filename": filename,
             "chunk_index": i,
             "chunk_text": chunk,
+            "chunk_hash": compute_chunk_hash(chunk),
             "data_type": data_type,
             "bundesland": bundesland,
             "use_case": use_case,
             "year": year,
             **extra_metadata,
+            **{k: v for k, v in legal.items() if v not in (None, "")},
         }
-        # Merge structural metadata from embedding service (D2)
-        if i < len(chunks_meta):
-            meta = chunks_meta[i]
-            for field in _STRUCT_FIELDS:
-                value = meta.get(field)
-                if value is not None and value != "":
-                    payload[field] = value
+        # Seite aus den Struktur-Metadaten uebernehmen (nicht Teil von legal)
+        if meta.get("page") not in (None, ""):
+            payload["page"] = meta["page"]
         payloads.append(payload)
+        ids.append(
+            deterministic_point_id(reg_code, legal["article"], legal["paragraph"], i, document_version)
+        )
 
-    # --- Index in Qdrant ---
+    # --- Index in Qdrant (deterministische IDs fuer stabilen Re-Link) ---
     try:
         indexed = await qdrant_wrapper.index_documents(
             collection=collection,
             vectors=embeddings,
             payloads=payloads,
+            ids=ids,
         )
     except Exception as exc:
         logger.error("Qdrant indexing failed: %s", exc)