fix(llm-review): Diagnose-Logging + think=false + <think>-Tag-Stripping

- think: false in Ollama API Request (qwen3 disables CoT nativ)
- <think>...</think> Stripping in _parse_llm_json_array (Fallback falls
  think:false nicht greift)
- INFO-Logging: wie viele Einträge gesendet werden, Response-Länge,
  Anzahl geparster Einträge
- DEBUG-Logging: erste 3 Eingabe-Einträge, ersten 500 Zeichen der Antwort
- Bessere Fehlermeldung wenn JSON-Parsing fehlschlägt

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

klausur-service/backend/cv_vocab_pipeline.py

@@ -5562,6 +5562,10 @@ async def llm_review_entries(
         for e in review_entries
     ]
 
+    logger.info("LLM review: sending %d/%d entries to %s (skipped %d without digit-pattern)",
+                len(review_entries), len(entries), model, len(entries) - len(reviewable))
+    logger.debug("LLM review input: %s", _json.dumps(table_lines[:3], ensure_ascii=False))
+
     prompt = _build_llm_prompt(table_lines)
 
     t0 = time.time()
@@ -5572,6 +5576,7 @@ async def llm_review_entries(
                 "model": model,
                 "messages": [{"role": "user", "content": prompt}],
                 "stream": False,
+                "think": False,   # qwen3: disable chain-of-thought (Ollama >=0.6)
                 "options": {"temperature": 0.1, "num_predict": 8192},
             },
         )
@@ -5579,7 +5584,11 @@ async def llm_review_entries(
         content = resp.json().get("message", {}).get("content", "")
     duration_ms = int((time.time() - t0) * 1000)
 
+    logger.info("LLM review: response in %dms, raw length=%d chars", duration_ms, len(content))
+    logger.debug("LLM review raw response (first 500): %.500s", content)
+
     corrected = _parse_llm_json_array(content)
+    logger.info("LLM review: parsed %d corrected entries, applying diff...", len(corrected))
     changes, corrected_entries = _diff_batch(review_entries, corrected)
 
     # Merge corrected entries back into the full list
@@ -5696,15 +5705,19 @@ async def llm_review_entries_streaming(
 
 
 def _parse_llm_json_array(text: str) -> List[Dict]:
-    """Extract JSON array from LLM response (may contain markdown fences)."""
+    """Extract JSON array from LLM response (handles markdown fences and qwen3 think-tags)."""
+    # Strip qwen3 <think>...</think> blocks (present even with think=False on some builds)
+    text = _re.sub(r'<think>.*?</think>', '', text, flags=_re.DOTALL)
     # Strip markdown code fences
     text = _re.sub(r'```json\s*', '', text)
     text = _re.sub(r'```\s*', '', text)
-    # Find array
+    # Find first [ ... last ]  (non-greedy would miss nested structures, greedy is correct here)
     match = _re.search(r'\[.*\]', text, _re.DOTALL)
     if match:
         try:
             return _json.loads(match.group())
-        except (ValueError, _json.JSONDecodeError):
-            pass
+        except (ValueError, _json.JSONDecodeError) as e:
+            logger.warning("LLM review: JSON parse failed: %s | raw snippet: %.200s", e, match.group()[:200])
+    else:
+        logger.warning("LLM review: no JSON array found in response (%.200s)", text[:200])
     return []