feat: Browser-Matrix Stufe 1.a + 2 weitere GT-Findings + Plausibility-LLM-Härtung

Stage 1.a Browser-Matrix (Task #15) — Multi-Engine Scaffolding:
  - consent-tester/Dockerfile: firefox + webkit + Xvfb deps
  - playwright install chromium firefox webkit
  - services/browser_profiles.py: Registry mit DEFAULT_PROFILES
    (Chromium-Headed/Firefox-Headed/WebKit-Headed/Mobile-Safari) +
    EXTRA_PROFILES (Chrome-Channel, Edge, Brave)
  - services/multi_browser_scanner.py: run_matrix() orchestriert N
    parallele Scans + worst-of-Aggregation + 3 Sub-Scores
    (Pre-Consent 50%, Reject-Respekt 30%, Banner-Design 20%) +
    Hard-Fail-Cap auf <60% bei Pre-Consent/Reject-Verstoß
  - routes_matrix.py: POST /scan-matrix Endpoint (eigenes Modul,
    damit main.py unter 500 LOC bleibt)
  KNOWN: Stage 1.a-Shim ruft alle Profile auf demselben Chromium,
    echte Engine-Diversität in Stage 1.b (consent_scanner.py Param)

Coverage-Gap 3 (Task #17): 2/3 verbleibende GT-Lücken geschlossen:
  - B9 impressum_multi_entity_check (IMPRESSUM-001): erkennt
    USt-IdNr/HR/GF-Fehlen pro Entity bei multi-entity Impressen
    (Elli: USt-IdNr nur bei Elli Mobility, fehlt bei VW Group Charging)
  - B10 transfer_mechanism_check (TRANSFER-001): pro Non-EU-Vendor
    in cmp_vendors prüft DSE auf DPF/SCCs/BCRs/Einwilligung im
    ±400-char-Window. Findet Vendors ohne benannten Mechanismus.
  - TH-RETENTION-002 (AI-Datenkategorie-Differenzierung) bleibt
    semantisch-tief, vorgesehen für Specialist-Agents Task #18.

Plausibility-LLM Empty-Response-Härtung (Task #16):
  - BATCH_SIZE 8 → 4, EXCERPT 4000 → 1500 chars, TIMEOUT 60 → 45s
  - Single-retry mit halbierter Batch wenn LLM empty content
    zurückgibt — qwen3:30b-a3b rejektiert manchmal ≥6-Item-Prompts
    unter format='json'. Falls auch Half-Batch empty: log + skip.
  - Pipeline läuft jetzt nicht mehr 10min in Timeouts.

GT-Coverage Sprung: 10/13 → 11/13 (85%). 4/4 HIGH ✓, 5/6 MEDIUM ✓,
2/3 LOW ✓.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

backend-compliance/compliance/services/finding_plausibility_check.py

@@ -51,8 +51,13 @@ logger = logging.getLogger(__name__)
 
 OLLAMA_URL = os.getenv("OLLAMA_URL", "http://host.docker.internal:11434")
 MODEL = os.getenv("PLAUSIBILITY_LLM_MODEL", "qwen3:30b-a3b")
-BATCH_SIZE = int(os.getenv("PLAUSIBILITY_BATCH_SIZE", "8"))
-TIMEOUT = float(os.getenv("PLAUSIBILITY_TIMEOUT_S", "60.0"))
+# Reduced from 8 → 4 to fight qwen3 empty-response-on-large-prompts bug.
+# 4 items × ~500 token/item + 2000 system + 1500 excerpt = ~5500 token total,
+# well within qwen3's safe range for format='json'.
+BATCH_SIZE = int(os.getenv("PLAUSIBILITY_BATCH_SIZE", "4"))
+TIMEOUT = float(os.getenv("PLAUSIBILITY_TIMEOUT_S", "45.0"))
+# Reduced excerpt 4000 → 1500 chars (same reason).
+DOC_EXCERPT_CHARS = int(os.getenv("PLAUSIBILITY_DOC_EXCERPT", "1500"))
 
 # In-memory cache: (input_hash) -> result_dict. Survives one run.
 _CACHE: dict[str, dict] = {}
@@ -121,7 +126,8 @@ def _build_user_prompt(items: list[dict], doc_title: str,
     )
     return (
         f"DOKUMENT: {doc_title}\n\n"
-        f"DOKUMENT-AUSZUG (max 4000 Zeichen):\n{doc_excerpt[:4000]}\n\n"
+        f"DOKUMENT-AUSZUG (max {DOC_EXCERPT_CHARS} Zeichen):\n"
+        f"{doc_excerpt[:DOC_EXCERPT_CHARS]}\n\n"
         f"FINDINGS ZU BEWERTEN:\n{findings_block}"
     )
 
@@ -149,6 +155,23 @@ async def _ask_llm_batch(items: list[dict], doc_title: str,
             r.raise_for_status()
             content = (r.json().get("message") or {}).get("content", "")
             if not content:
+                # Single retry with smaller batch — qwen3 sometimes
+                # rejects ≥6-item prompts under format='json'.
+                if len(items) > 2:
+                    half = len(items) // 2
+                    logger.info(
+                        "plausibility empty → retry split %d → %dx2",
+                        len(items), half,
+                    )
+                    first = await _ask_llm_batch(
+                        items[:half], doc_title, doc_excerpt,
+                    )
+                    second = await _ask_llm_batch(
+                        items[half:], doc_title, doc_excerpt,
+                    )
+                    out.update(first)
+                    out.update(second)
+                    return out
                 logger.warning("plausibility LLM returned empty content")
                 return out
             try: