feat(cra): Datenblatt-Extraktion auf lokales 35B + llm_status-Fix

llm_cascade additiv modell-faehig (optionaler model-Param, Cache-Key kennt
model_hint → keine Kollision; Default unveraendert für alle anderen Nutzer).
Datenblatt-Extraktor nutzt jetzt qwen3.5:35b-a3b (CRA_DATASHEET_MODEL, gleiches
Modell wie der Compliance Advisor) für bessere semantische Zuordnung. Plus
llm_status (ok|empty|unavailable) + Logging statt stillem except; Frontend zeigt
bei 'unavailable' einen Hinweis statt leerer Felder (wichtig auf prod ohne
lokales Ollama → Cascade-Fallback bzw. Hinweis).

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

backend-compliance/compliance/services/cra_datasheet_extractor.py

@@ -10,8 +10,15 @@ Pure + testable: detect_signals / parse_grenzen_json / compute_followups. The
 async extract_grenzen() wraps the LLM call (llm_cascade, same as vendor extractor).
 """
 import json
+import logging
+import os
 import re
-from typing import Optional
+
+logger = logging.getLogger(__name__)
+
+# Datasheet extraction uses the local 35B (same model as the Compliance Advisor) —
+# higher-quality semantic mapping than the default cascade model. Env-overridable.
+_DATASHEET_MODEL = os.getenv("CRA_DATASHEET_MODEL", "qwen3.5:35b-a3b")
 
 # IACE Grenzen field keys (must match admin LimitsFormData). label + whether it
 # is essential for a usable risk assessment (=> asked as follow-up if empty).
@@ -150,6 +157,7 @@ async def extract_grenzen(text: str, max_chars: int = 20000) -> dict:
     signals = detect_signals(text or "")
     limits: dict = {}
     provenance: dict = {}
+    llm_status = "skipped"  # skipped | ok | empty | unavailable
     excerpt = (text or "")[:max_chars]
     if len(excerpt) >= 200:
         try:
@@ -157,20 +165,25 @@ async def extract_grenzen(text: str, max_chars: int = 20000) -> dict:
             res = await call_with_cascade(
                 system=_system_prompt(),
                 user=f"Datenblatt-Text:\n\n{excerpt}",
-                min_confidence=0.5, max_tokens=4000,
+                min_confidence=0.5, max_tokens=4000, model=_DATASHEET_MODEL,
             )
             parsed = parse_grenzen_json(res.get("text", "") if isinstance(res, dict) else "")
             for key, entry in parsed.items():
                 limits[key] = entry["value"]
                 provenance[key] = entry.get("source", "")
-        except Exception:
-            pass  # extraction is best-effort; fall back to detector + follow-ups
+            llm_status = "ok" if parsed else "empty"
+        except Exception as e:
+            # best-effort: keep the deterministic facts, but surface the failure so
+            # a cold-start/timeout doesn't masquerade as "nothing on the datasheet".
+            logger.warning("datasheet LLM extraction failed: %s (%s)", e, type(e).__name__)
+            llm_status = "unavailable"
 
     _merge_detected(limits, provenance, signals)
     return {
         "limits": limits,
         "provenance": provenance,
         "detected": signals,
+        "llm_status": llm_status,
         "filled": sorted(limits.keys()),
         "missing": [k for k in _FIELD_KEYS if not (limits.get(k) or "").strip()],
         "followup": compute_followups(limits),