feat: Mehrstufige Deep-Research-Pipeline mit Quellenkontext

- DEEP_RESEARCH_PROMPT: 4-Phasen-Strategie (Breite Erfassung → Lückenanalyse → Gezielte Tiefenrecherche → Verifikation)
- Ziel 15-25 Quellen aus 5+ Quellentypen statt 8-15 aus Mainstream
- researcher.search(): Neuer Parameter existing_articles — bereits bekannte Quellen werden als Kontext übergeben, damit Claude gezielt neue Perspektiven findet
- orchestrator: DB-Abfrage vor Pipeline verschoben, bestehende Artikel als Kontext an Researcher übergeben (nur Research-Typ)

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

src/agents/orchestrator.py

@@ -571,6 +571,13 @@ class AgentOrchestrator:
                     "data": {"status": research_status, "detail": research_detail, "started_at": now_utc},
                 }, visibility, created_by, tenant_id)
 
+            # Bestehende Artikel vorladen (für Dedup UND Kontext)
+            cursor = await db.execute(
+                "SELECT id, source_url, headline, source FROM articles WHERE incident_id = ?",
+                (incident_id,),
+            )
+            existing_db_articles_full = await cursor.fetchall()
+
             # Schritt 1+2: RSS-Feeds und Claude-Recherche parallel ausführen
             async def _rss_pipeline():
                 """RSS-Feed-Suche (Feed-Selektion + dynamische Keywords + Parsing)."""
@@ -617,7 +624,20 @@ class AgentOrchestrator:
             async def _web_search_pipeline():
                 """Claude WebSearch-Recherche."""
                 researcher = ResearcherAgent()
-                results, usage = await researcher.search(title, description, incident_type, international=international, user_id=user_id)
+                # Bei Research: bestehende Artikel als Kontext mitgeben
+                existing_for_context = None
+                if incident_type == "research" and existing_db_articles_full:
+                    existing_for_context = [
+                        {"source": row["source"] if "source" in row.keys() else "",
+                         "headline": row["headline"],
+                         "source_url": row["source_url"]}
+                        for row in existing_db_articles_full
+                    ]
+                results, usage = await researcher.search(
+                    title, description, incident_type,
+                    international=international, user_id=user_id,
+                    existing_articles=existing_for_context,
+                )
                 logger.info(f"Claude-Recherche: {len(results)} Ergebnisse")
                 return results, usage
 
@@ -714,14 +734,10 @@ class AgentOrchestrator:
                 }, visibility, created_by, tenant_id)
 
             # --- Set-basierte DB-Deduplizierung (statt N×M Queries) ---
-            cursor = await db.execute(
-                "SELECT id, source_url, headline FROM articles WHERE incident_id = ?",
-                (incident_id,),
-            )
-            existing_db_articles = await cursor.fetchall()
+            # existing_db_articles_full wurde bereits oben geladen
             existing_urls = set()
             existing_headlines = set()
-            for row in existing_db_articles:
+            for row in existing_db_articles_full:
                 if row["source_url"]:
                     existing_urls.add(_normalize_url(row["source_url"]))
                 if row["headline"] and len(row["headline"]) > 20: