Fix broken source links caused by LLM-generated letter suffixes (e.g. 1383a)

The LLM occasionally generates source references with letter suffixes
(e.g. [1383a], [1396b]) despite being instructed not to. This caused
broken links because the sources array only contained integer nr values.

Backend: Add _sanitize_sources() to strip letter suffixes after parsing
and deduplicate, preferring entries with valid URLs.

Frontend: Add fallback in citation renderer - when a suffix reference
like [1383a] has no matching source with URL, fall back to the base
number [1383].

Also cleaned up 99 broken suffix entries and 44 suffix references in
the Irankonflikt incident (ID 6) database records.

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

src/agents/analyzer.py

@@ -242,6 +242,7 @@ class AnalyzerAgent:
             result, usage = await call_claude(prompt)
             analysis = self._parse_response(result)
             if analysis:
+                analysis = self._sanitize_sources(analysis)
                 logger.info(f"Erstanalyse abgeschlossen: {len(analysis.get('sources', []))} Quellen referenziert")
             return analysis, usage
         except Exception as e:
@@ -303,6 +304,8 @@ class AnalyzerAgent:
         try:
             result, usage = await call_claude(prompt)
             analysis = self._parse_response(result)
+            if analysis:
+                analysis = self._sanitize_sources(analysis)
             if analysis and self._all_previous_sources:
                 # Merge: alte Quellen beibehalten, neue hinzufuegen
                 returned_sources = analysis.get("sources", [])
@@ -325,6 +328,51 @@ class AnalyzerAgent:
             logger.error(f"Inkrementelle Analyse-Fehler: {e}")
             return None, None
 
+    def _sanitize_sources(self, analysis: dict) -> dict:
+        """Entfernt Buchstaben-Suffixe aus Quellennummern (z.B. '1383a' -> 1383).
+
+        Das LLM erzeugt trotz Anweisung gelegentlich Suffix-Nummern.
+        Diese werden hier auf die Basisnummer normalisiert.
+        Duplikate werden entfernt, wobei Eintraege mit URL bevorzugt werden.
+        """
+        sources = analysis.get("sources", [])
+        if not sources:
+            return analysis
+
+        cleaned = {}
+        suffix_count = 0
+        for s in sources:
+            nr = s.get("nr", "")
+            nr_str = str(nr)
+            # Prüfe auf Buchstaben-Suffix (z.B. "1383a", "1383b")
+            m = re.match(r"^(\d+)[a-z]$", nr_str)
+            if m:
+                base_nr = int(m.group(1))
+                suffix_count += 1
+                # Nur übernehmen wenn Basisnummer noch nicht existiert oder
+                # dieser Eintrag eine URL hat und der bisherige nicht
+                if base_nr not in cleaned:
+                    s_copy = dict(s)
+                    s_copy["nr"] = base_nr
+                    cleaned[base_nr] = s_copy
+                elif s.get("url") and not cleaned[base_nr].get("url"):
+                    s_copy = dict(s)
+                    s_copy["nr"] = base_nr
+                    cleaned[base_nr] = s_copy
+            else:
+                nr_int = int(nr) if isinstance(nr, (int, float)) or (isinstance(nr, str) and nr.isdigit()) else nr
+                if nr_int not in cleaned:
+                    cleaned[nr_int] = s
+                elif s.get("url") and not cleaned[nr_int].get("url"):
+                    cleaned[nr_int] = s
+
+        if suffix_count > 0:
+            logger.info(f"Quellen-Sanitierung: {suffix_count} Buchstaben-Suffixe entfernt")
+            analysis["sources"] = sorted(cleaned.values(),
+                key=lambda s: s.get("nr", 0) if isinstance(s.get("nr"), int) else 9999)
+
+        return analysis
+
     def _parse_response(self, response: str) -> dict | None:
         """Parst die Claude-Antwort als JSON-Objekt mit robustem Fallback."""
         # Markdown-Code-Fences entfernen

src/static/js/components.js

@@ -444,10 +444,17 @@ const UI = {
         html = html.replace(/<\/ul>(<br>)+/g, '</ul>');
         html = html.replace(/(<br>){2,}/g, '<br>');
 
-        // Inline-Zitate [1], [2] etc. als klickbare Links rendern
+        // Inline-Zitate [1], [2], [1383a] etc. als klickbare Links rendern
         if (sources.length > 0) {
             html = html.replace(/\[(\d+[a-z]?)\]/g, (match, num) => {
-                const src = sources.find(s => String(s.nr) === num || Number(s.nr) === Number(num));
+                // Exakte Suche (auch mit Buchstaben-Suffix)
+                let src = sources.find(s => String(s.nr) === num || Number(s.nr) === Number(num));
+                // Fallback: Bei Suffix wie "1383a" auf Basisnummer 1383 zurueckfallen
+                if ((!src || !src.url) && /[a-z]$/.test(num)) {
+                    const baseNum = num.replace(/[a-z]$/, '');
+                    const baseSrc = sources.find(s => String(s.nr) === baseNum || Number(s.nr) === Number(baseNum));
+                    if (baseSrc && baseSrc.url) src = baseSrc;
+                }
                 if (src && src.url) {
                     return `<a href="${this.escape(src.url)}" target="_blank" rel="noopener" class="citation" title="${this.escape(src.name)}">[${num}]</a>`;
                 }