feat(checkers): platform router + Haiku sufficiency tier; cookie is first consumer

Generalise "Embedding finds, Claude decides" into the shared Pruefer-Library:
- router.route_and_check dispatches control -> sensor_classification -> Checker.
- build_spec reads sensor_classification (CONTENT/LLM -> judge=haiku, the
  validated sufficiency tier; the Qwen-first cascade is disproven for sufficiency).
- LLMChecker gains a Haiku-direct tier (reuses the validated deep_check prompt).
- Cookie Layer-3 now routes through route_and_check instead of bespoke code, so
  cookie is the first real router consumer -- proves the architecture end-to-end.

Reproduces the validated result via the shared path: FN 159->14, recall
0.13->0.92, precision 0.89 (vs bespoke 12/0.93/0.90 -- within Haiku noise).
Tests: 10/10 (router dispatch + build_spec + haiku tier + cookie rewire).

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

backend-compliance/compliance/services/checkers/llm_checker.py

@@ -45,6 +45,11 @@ class LLMChecker:
         text = doc.text or ""
         if len(text) < 50:
             return CheckResult(present=None, source="llm")
+        # decision_method=LLM mit judge='haiku': Sufficiency-Pfad (validiert
+        # P0.89/R0.91). Der Qwen-first-Cascade ist als Sufficiency-Judge
+        # widerlegt -> hier Haiku direkt, kriteriengeführte Subsumtion.
+        if (ctrl.extra or {}).get("judge") == "haiku":
+            return await self._haiku(ctrl, text)
         secs = _sections(text)
         if ctrl.topic_regex:
             rel = [s for s in secs if re.search(ctrl.topic_regex, s, re.I)][:6] or secs[:6]
@@ -71,3 +76,31 @@ class LLMChecker:
         except Exception as e:
             logger.info("llm checker fail %s: %s", ctrl.control_id, str(e)[:80])
             return CheckResult(present=None, source="error")
+
+    async def _haiku(self, ctrl: ControlSpec, text: str) -> CheckResult:
+        """Sufficiency via Haiku direkt (validierter Judge). Kriteriengeführt:
+        die Rechts-Elemente stehen in ctrl.paraphrases; wiederverwendet den
+        validierten deep_check-Sufficiency-Prompt."""
+        try:
+            from compliance.services.llm_cascade import _call_anthropic
+            from compliance.services.specialist_agents.dse.deep_check import (
+                _JUDGE_SYS, _build_user, _parse as _parse_judge,
+            )
+            crit = ctrl.paraphrases or [ctrl.label or ctrl.control_id]
+            user = _build_user(text, ctrl.label or ctrl.control_id, crit)
+            obj = None
+            for _ in range(2):
+                obj = _parse_judge(await _call_anthropic(_JUDGE_SYS, user, max_tokens=400))
+                if obj:
+                    break
+            if not obj:
+                return CheckResult(present=None, source="haiku")
+            return CheckResult(
+                present=bool(obj.get("erfuellt")),
+                evidence=(obj.get("begruendung") or "")[:120],
+                confidence=float(obj.get("confidence") or 0.0),
+                source="haiku",
+            )
+        except Exception as e:
+            logger.info("llm haiku checker fail %s: %s", ctrl.control_id, str(e)[:80])
+            return CheckResult(present=None, source="error")