Hybrid-Zitate: verified/unverified statt drop + UI-Labels

reconstruct_zitate droppt Zitate nicht mehr bei No-Match, sondern
markiert sie als verified=false. Das ist ehrlicher: paraphrasierte
Zitate sind wertvoller Kontext, sie brauchen nur ein visuelles
Unterscheidungsmerkmal.

UI:
- Verifizierte Zitate: grüner solid Border, "✓ verifiziert"
- Paraphrasierte Zitate: gelber dashed Border, "~ paraphrasiert
  (nicht wörtlich im Programm)"
- Warning-Text: "Zu diesem Themenkomplex konnten keine konkreten
  Formulierungen im Wahlprogramm gefunden werden"
- Antragsteller:in / Landesregierung als farbige Badges

Zitat-Model: neues Optional[bool] Feld "verified".

Tests: 206 passed (test_drops angepasst auf neues Verhalten).

app/embeddings.py

@@ -797,20 +797,18 @@ def find_chunk_for_text(text: str, chunks: list[dict]) -> Optional[dict]:
 
 
 def reconstruct_zitate(data: dict, semantic_quotes: dict) -> dict:
-    """Replace LLM-emitted quelle/url with canonical chunk values; drop unbacked.
+    """Verify and reconstruct LLM-emitted zitate against retrieved chunks.
 
-    Walks over ``data['wahlprogrammScores'][i][kind]['zitate']`` (the raw
+    For each Zitat:
-    LLM-output dict, not the Pydantic model). For each Zitat:
+    * **verified** (substring/4-word-anchor match): overwrite quelle/url
+      with canonical chunk values, set ``verified: true``.
+    * **unverified** (no match found): keep the Zitat but set
+      ``verified: false``. The UI shows it with a different style so the
+      user knows it's an LLM-Paraphrase, not a wörtliches Zitat.
 
-    * Locate the chunk whose text contains the snippet (or a 5-word anchor
+    This replaces the old drop-on-no-match behavior (ADR 0001 Option B)
-      from it). Search across **all** retrieved chunks regardless of party,
+    with a more honest approach: paraphrased citations are still valuable
-      so cross-mixes between Q-IDs become invisible to the persisted output.
+    context, they just need to be marked as such.
-    * If found: overwrite ``quelle`` and ``url`` with values derived from
-      the matching chunk's ``programm_id`` + ``seite``. The LLM is no longer
-      trusted for these fields.
-    * If not found: drop the Zitat entirely.
-
-    Returns the same ``data`` dict (mutated in place) for chaining.
     """
     if not semantic_quotes:
         return data
@@ -830,12 +828,16 @@ def reconstruct_zitate(data: dict, semantic_quotes: dict) -> dict:
             for z in zitate:
                 text = z.get("text", "")
                 matched = find_chunk_for_text(text, all_chunks)
-                if matched is None:
+                if matched is not None:
-                    continue
                     z["quelle"] = _chunk_source_label(matched)
                     url = _chunk_pdf_url(matched)
                     if url:
                         z["url"] = url
+                    z["verified"] = True
+                else:
+                    # Kein Match — Zitat behalten aber als unverified markieren.
+                    # Die LLM-emittierte quelle/url bleibt (best effort).
+                    z["verified"] = False
                 cleaned.append(z)
             blk["zitate"] = cleaned
     return data

app/models.py

@@ -45,6 +45,7 @@ class Zitat(BaseModel):
     text: str
     quelle: str
     url: Optional[str] = None
+    verified: Optional[bool] = None  # True=wörtlich im Chunk, False=paraphrasiert, None=pre-#97
 
 
 class ProgrammScore(BaseModel):

app/templates/index.html

@@ -1619,14 +1619,22 @@
 
             const wahlprogrammHtml = (item.wahlprogrammScores || []).map(wp => {
                 // Zitate formatieren mit klickbaren Links + Highlighting
-                const zitateHtml = (wp.wahlprogramm?.zitate || []).map(z => `
+                const zitateHtml = (wp.wahlprogramm?.zitate || []).map(z => {
-                    <div style="margin: 0.5rem 0; padding: 0.5rem; background: #f8f9fa; border-left: 3px solid #889e33; font-size: 0.85rem;">
+                    const isVerified = z.verified !== false;
+                    const borderColor = isVerified ? '#889e33' : '#ffc107';
+                    const bgColor = isVerified ? '#f8f9fa' : '#fffbf0';
+                    const badge = isVerified
+                        ? '<span style="font-size:0.7rem;color:#889e33;">✓ verifiziert</span>'
+                        : '<span style="font-size:0.7rem;color:#b8860b;">~ paraphrasiert (nicht wörtlich im Programm)</span>';
+                    return `
+                    <div style="margin: 0.5rem 0; padding: 0.5rem; background: ${bgColor}; border-left: 3px ${isVerified ? 'solid' : 'dashed'} ${borderColor}; font-size: 0.85rem;">
                         <em>"${z.text}"</em><br>
-                        <a href="${makeCiteUrl(z, item.drucksache, item.bundesland)}" target="_blank" style="color: #009da5; font-size: 0.8rem;">
+                        ${z.quelle ? `<a href="${makeCiteUrl(z, item.drucksache, item.bundesland)}" target="_blank" style="color: #009da5; font-size: 0.8rem;">
                             📄 ${z.quelle}
-                        </a>
+                        </a>` : ''}
-                    </div>
+                        ${badge}
-                `).join('');
+                    </div>`;
+                }).join('');
 
                 // Issue #63: Transparenz-Warnung bei Score > 0 ohne Zitate.
                 // Differenziert zwischen "Score 0 = keine Quellen" (LLM hat

tests/test_embeddings.py

@@ -316,8 +316,15 @@ class TestReconstructZitate:
         }
         out = reconstruct_zitate(data, semantic_quotes)
         zitate = out["wahlprogrammScores"][0]["wahlprogramm"]["zitate"]
-        assert len(zitate) == 1
+        # Beide Zitate bleiben erhalten — das nicht-matchende wird als
-        assert "geschlechtersensiblen" in zitate[0]["text"]
+        # unverified markiert statt gedroppt (Hybrid-Ansatz).
+        assert len(zitate) == 2
+        # Das halluzinierte Zitat ist unverified
+        halluziniert = [z for z in zitate if "Rechtsextremismus" in z["text"]]
+        assert halluziniert[0]["verified"] is False
+        # Das echte Zitat ist verified
+        echt = [z for z in zitate if "geschlechtersensiblen" in z["text"]]
+        assert echt[0]["verified"] is True
 
     def test_empty_semantic_quotes_is_noop(self):
         data = {"wahlprogrammScores": [{