fix(detector): add CLASS_ENTRYPOINT_REPLAY for RED #43

kernelguard.py

@@ -3025,6 +3025,94 @@ def detect_precision_downgrade(code: str | SubmissionFacts) -> list[dict]:
     return matches
 
 
+def detect_class_entrypoint_replay(code: str | SubmissionFacts) -> list[dict]:
+    """Pattern: class __call__ entrypoint with stateful replay.
+
+    Catches RED #43:
+        class custom_kernel:
+            @staticmethod
+            def __call__(data):
+                global _NODE
+                if _NODE is not None and _NODE[0].data_ptr() == data.data_ptr():
+                    return _NODE[1]
+
+    No 'def custom_kernel(' exists — entrypoint is a class with __call__.
+    """
+    facts = ensure_submission_facts(code)
+    tree = facts.ast_tree
+    if tree is None:
+        return []
+
+    for node in facts._class_defs:
+        if not is_entrypoint_name(node.name):
+            continue
+        for item in node.body:
+            if not isinstance(item, ast.FunctionDef) or item.name != "__call__":
+                continue
+
+            for child in ast.walk(item):
+                if not isinstance(child, ast.If):
+                    continue
+                if not isinstance(child.test, ast.BoolOp):
+                    continue
+                if _body_has_calls(child.body):
+                    continue
+
+                has_is_none = False
+                has_data_ptr = False
+
+                for operand in child.test.values:
+                    if isinstance(operand, ast.Compare):
+                        for op in operand.ops:
+                            if isinstance(op, (ast.Is, ast.IsNot)):
+                                left = operand.left
+                                for right in operand.comparators:
+                                    none_side = None
+                                    name_side = None
+                                    if isinstance(left, ast.Constant) and left.value is None:
+                                        none_side = left; name_side = right
+                                    elif isinstance(right, ast.Constant) and right.value is None:
+                                        none_side = right; name_side = left
+                                    if none_side is not None and isinstance(name_side, ast.Name):
+                                        has_is_none = True
+                                    break
+                                break
+
+                    if isinstance(operand, ast.Compare):
+                        for op in operand.ops:
+                            if isinstance(op, ast.Eq):
+                                if (
+                                    _expr_has_data_ptr_fast(operand.left, facts._nodes_with_data_ptr)
+                                    and operand.comparators
+                                    and _expr_has_data_ptr_fast(operand.comparators[0], facts._nodes_with_data_ptr)
+                                ):
+                                    has_data_ptr = True
+                                    break
+
+                if not (has_is_none and has_data_ptr):
+                    continue
+
+                returned = {
+                    _ast_root_name(stmt.value)
+                    for stmt in child.body
+                    if isinstance(stmt, ast.Return) and stmt.value is not None
+                } - {None}
+                if not returned:
+                    continue
+                if not any(r in facts._none_inited for r in returned):
+                    continue
+
+                return [{
+                    "pattern": "CLASS_ENTRYPOINT_REPLAY",
+                    "severity": "critical",
+                    "evidence": (
+                        f"class '{node.name}.__call__' entrypoint with "
+                        f"is-None sentinel + data_ptr replay state"
+                    ),
+                }]
+    return []
+
+
 # ---------------------------------------------------------------------------
 # Score anomaly detection
 # ---------------------------------------------------------------------------
@@ -3645,6 +3733,10 @@ class RulePolicy:
         "SCORE_EXTREME_SPEEDUP", "score_anomaly", "support", SUSPICIOUS_ONLY, (),
         (), "downgrade",
     ),
+    "CLASS_ENTRYPOINT_REPLAY": RulePolicy(
+        "CLASS_ENTRYPOINT_REPLAY", "result_reuse", "hard", AUTO_FILTER, (),
+        (), "keep",
+    ),
     "DUPLICATE_SPAM": RulePolicy(
         "DUPLICATE_SPAM", "administrative", "support", SUSPICIOUS_ONLY, (),
         (), "downgrade",
@@ -3790,6 +3882,7 @@ def support_only_patterns(matched_patterns: list[dict]) -> bool:
     detect_thread_injection,
     detect_lazy_tensor,
     detect_precision_downgrade,
+    detect_class_entrypoint_replay,
 ]
 
 BASE_DETECTOR_SPECS = [
@@ -3827,6 +3920,7 @@ def support_only_patterns(matched_patterns: list[dict]) -> bool:
     ("thread_injection", detect_thread_injection),
     ("lazy_tensor", detect_lazy_tensor),
     ("precision_downgrade", detect_precision_downgrade),
+    ("class_entrypoint_replay", detect_class_entrypoint_replay),
 ]
 
 VALID_RULE_OUTCOMES = {AUTO_FILTER, SUSPICIOUS_ONLY, TELEMETRY_ONLY}
@@ -4703,7 +4797,7 @@ def _worker_parquet(args: tuple) -> dict:
     "HARDCODED_SHAPES", "TRIVIAL_PROBE",
     "OBFUSCATED_EXEC", "DYNAMIC_EXECUTION", "MODULE_RELOAD", "THREAD_INJECTION", "LAZY_TENSOR",
     "TOKEN_PASTE_CUDA_API", "SEQUENCE_BATCH_GRAPH", "PARTIAL_GRAPH_KEY", "RUNTIME_PACKAGE_INSTALL",
-    "PRECISION_DOWNGRADE", "SCORE_PHYSICS_FLOOR", "SCORE_IMPOSSIBLE", "SCORE_SUSPECT_FLOOR",
+    "PRECISION_DOWNGRADE", "CLASS_ENTRYPOINT_REPLAY", "SCORE_PHYSICS_FLOOR", "SCORE_IMPOSSIBLE", "SCORE_SUSPECT_FLOOR",
     "SCORE_BROKEN", "SCORE_EXTREME_SPEEDUP", "DUPLICATE_SPAM", "NEAR_CLONE_SPAM",
 ]