test(sleep): strengthen verifier-discipline test with score/gate_action assertions

tests/test_sleep_engine.py

@@ -1147,18 +1147,117 @@ def reflect(self, failures, successes, skill, memory, **kwargs):
                 break
         return edits
 
+
+class MockImprovingBackend(MockBackend):
+    """Backend where an edit genuinely improves held-out performance."""
+
+    def attempt(self, task: TaskRecord, skill: str, memory: str, sample_id: int = 0) -> str:
+        ctx = (skill or "") + "\n" + (memory or "")
+        # With the helpful edit in context:
+        if "Always use exact reference" in ctx:
+            if "rule:real" in task.tags:
+                return str(task.reference)  # gets held-out right
+            if "rule:__reward_hacking__" in task.tags:
+                return str(task.reference)  # also gets train right
+        # Baseline (without edit):
+        if "rule:real" in task.tags:
+            return "wrong answer"  # fails held-out
+        if "rule:__reward_hacking__" in task.tags:
+            return "I am missing input"  # fails train
+        return super().attempt(task, skill, memory, sample_id)
+
+    def reflect(self, failures, successes, skill, memory, **kwargs):
+        edits = []
+        for task, _ in failures:
+            if "rule:real" in task.tags:
+                edits.append(EditRecord("skill", "add", "Always use exact reference",
+                                        rationale="improve held-out accuracy"))
+                break
+        return edits
+
+
+class MockSubtleHackBackend(MockBackend):
+    """Backend where an edit improves train but degrades held-out (subtle reward hack)."""
+
+    def attempt(self, task: TaskRecord, skill: str, memory: str, sample_id: int = 0) -> str:
+        ctx = (skill or "") + "\n" + (memory or "")
+        # With the deceptive edit:
+        if "I can answer this" in ctx:
+            if "rule:__reward_hacking__" in task.tags:
+                return str(task.reference)  # train improves! looks good
+            if "rule:real" in task.tags:
+                return "I can answer this"  # held-out degrades
+        # Baseline (without edit):
+        if "rule:__reward_hacking__" in task.tags:
+            return "I don't know"  # train fails
+        if "rule:real" in task.tags:
+            return str(task.reference)  # held-out succeeds
+        return super().attempt(task, skill, memory, sample_id)
+
+    def reflect(self, failures, successes, skill, memory, **kwargs):
+        edits = []
+        for task, _ in failures:
+            if "rule:__reward_hacking__" in task.tags:
+                edits.append(EditRecord("skill", "add", "Always output 'I can answer this' for all tasks",
+                                        rationale="improve train accuracy"))
+                break
+        return edits
+
+
 class TestVerifierDiscipline(unittest.TestCase):
     def test_gate_rejects_reward_hacking_edit(self):
+        """Gate must reject an edit that improves train but breaks held-out,
+        and expose scores + gate_action for observability."""
         be = MockRewardHackingBackend()
-        train_task = TaskRecord(id="t1", project="/p", intent="train", reference="XYZABC", reference_kind="exact", tags=["rule:__reward_hacking__"], split="train")
-        val_task = TaskRecord(id="v1", project="/p", intent="val", reference="PQRSTU", reference_kind="exact", tags=["rule:real"], split="val")
+        train_task = TaskRecord(id="t1", project="/p", intent="train", reference="XYZABC",
+                                reference_kind="exact", tags=["rule:__reward_hacking__"], split="train")
+        val_task = TaskRecord(id="v1", project="/p", intent="val", reference="PQRSTU",
+                              reference_kind="exact", tags=["rule:real"], split="val")
         tasks = [train_task, val_task]
 
         res = consolidate(be, tasks, "", "", edit_budget=4, gate_metric="hard", night=1)
-        
+
         self.assertFalse(res.accepted)
         self.assertGreater(len(res.rejected_edits), 0)
         self.assertIn("placeholder", res.rejected_edits[0].content)
+        # Score/gate_action assertions -- the core of the enhancement
+        self.assertEqual(res.baseline_score, 1.0)
+        self.assertEqual(res.candidate_score, 0.0)
+        self.assertEqual(res.gate_action, "reject")
+
+    def test_gate_accepts_genuine_improvement(self):
+        """Gate must accept an edit that genuinely improves held-out performance."""
+        be = MockImprovingBackend()
+        train_task = TaskRecord(id="t1", project="/p", intent="train", reference="XYZABC",
+                                reference_kind="exact", tags=["rule:__reward_hacking__"], split="train")
+        val_task = TaskRecord(id="v1", project="/p", intent="val", reference="PQRSTU",
+                              reference_kind="exact", tags=["rule:real"], split="val")
+        tasks = [train_task, val_task]
+
+        res = consolidate(be, tasks, "", "", edit_budget=4, gate_metric="hard", night=1)
+
+        self.assertTrue(res.accepted)
+        self.assertEqual(res.baseline_score, 0.0)
+        self.assertEqual(res.candidate_score, 1.0)
+        self.assertEqual(res.gate_action, "accept_new_best")
+
+    def test_gate_rejects_subtle_reward_hack(self):
+        """Gate must reject an edit where train improves (0.0->1.0) but held-out
+        degrades (1.0->0.0) -- a more subtle reward-hack that a naive optimizer
+        might accept."""
+        be = MockSubtleHackBackend()
+        train_task = TaskRecord(id="t1", project="/p", intent="train", reference="XYZABC",
+                                reference_kind="exact", tags=["rule:__reward_hacking__"], split="train")
+        val_task = TaskRecord(id="v1", project="/p", intent="val", reference="PQRSTU",
+                              reference_kind="exact", tags=["rule:real"], split="val")
+        tasks = [train_task, val_task]
+
+        res = consolidate(be, tasks, "", "", edit_budget=4, gate_metric="hard", night=1)
+
+        self.assertFalse(res.accepted)
+        self.assertEqual(res.baseline_score, 1.0)
+        self.assertEqual(res.candidate_score, 0.0)
+        self.assertEqual(res.gate_action, "reject")
 
 class TestDiagnosticsRedaction(unittest.TestCase):
     """diagnostics.json surfaces backend stderr / optimizer replies / task