feat(cookie): Layer-3 sufficiency-judge — Haiku re-judges embedding/boost rescues

The embedding/boost auto-rescue is intentionally optimistic (finds the topic, not
fulfilment) -> 159 FN over-rescues vs Opus-GT (recall 0.13). Layer-3 re-judges
exactly the rescued passes with the validated Haiku judge (cohort
cookie_sufficiency_v1 P0.89/R0.91) -- NOT the Qwen-first cascade (local is
disproven as a sufficiency judge) -- and un-passes them when the obligation is
not concretely met. Gated to the full check (not skip_llm).

Measured (5-firm Opus-GT, engine+L3): FN 159->12, recall 0.13->0.93,
precision 0.96->0.90 (276 rescues corrected). "Embedding finds, Claude decides."

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

backend-compliance/tests/test_cookie_sufficiency_judge.py

@@ -0,0 +1,67 @@
+"""Layer-3 cookie sufficiency-judge: only embedding/boost-RESCUED passes are
+re-judged by Haiku; keyword passes are untouched; a FEHLT verdict un-passes."""
+import pytest
+from unittest.mock import AsyncMock, patch
+
+from compliance.services.specialist_agents.cookie_policy._sufficiency_judge import (
+    judge_rescued,
+)
+
+_ANTHROPIC = "compliance.services.llm_cascade._call_anthropic"
+
+
+def _r(cid, source, passed=True):
+    return {"control_id": cid, "source": source, "passed": passed,
+            "label": cid, "_pass_criteria": ["konkrete Angabe nötig"]}
+
+
+@pytest.mark.asyncio
+async def test_rescued_unpassed_when_judge_fehlt():
+    results = [_r("A", "keyword+embedding")]
+    fake = AsyncMock(return_value='{"erfuellt": false, "confidence": 0.9, "begruendung": "fehlt"}')
+    with patch(_ANTHROPIC, new=fake):
+        n = await judge_rescued("text", results)
+    assert n == 1
+    assert results[0]["passed"] is False
+    assert "+llm_failed" in results[0]["source"]
+
+
+@pytest.mark.asyncio
+async def test_rescued_kept_when_judge_erfuellt():
+    results = [_r("A", "keyword+embedding")]
+    fake = AsyncMock(return_value='{"erfuellt": true, "confidence": 0.9}')
+    with patch(_ANTHROPIC, new=fake):
+        n = await judge_rescued("text", results)
+    assert n == 0
+    assert results[0]["passed"] is True
+
+
+@pytest.mark.asyncio
+async def test_keyword_pass_not_judged():
+    """Deterministisch (keyword) bestandene Controls werden NICHT befragt."""
+    results = [_r("A", "keyword")]
+    fake = AsyncMock(return_value='{"erfuellt": false}')
+    with patch(_ANTHROPIC, new=fake):
+        n = await judge_rescued("text", results)
+    assert n == 0
+    assert results[0]["passed"] is True
+    assert fake.call_count == 0
+
+
+@pytest.mark.asyncio
+async def test_boost_rescue_is_judged():
+    results = [_r("A", "keyword+regex_boost")]
+    fake = AsyncMock(return_value='{"erfuellt": false}')
+    with patch(_ANTHROPIC, new=fake):
+        n = await judge_rescued("text", results)
+    assert n == 1 and results[0]["passed"] is False
+
+
+@pytest.mark.asyncio
+async def test_failed_controls_ignored():
+    """Nicht-bestandene (failed) Controls sind nicht Sache dieser Schicht."""
+    results = [_r("A", "keyword+embedding", passed=False)]
+    fake = AsyncMock(return_value='{"erfuellt": false}')
+    with patch(_ANTHROPIC, new=fake):
+        n = await judge_rescued("text", results)
+    assert n == 0 and fake.call_count == 0