fix(detector): detect object attribute replay

kernelguard.py

@@ -2462,6 +2462,111 @@ def _has_ver(expr: ast.AST | None) -> bool:
     return []
 
 
+def detect_object_attribute_replay(code: str | SubmissionFacts) -> list[dict]:
+    """Pattern: object attribute stores and replays a prior input result."""
+    facts = ensure_submission_facts(code)
+    tree = facts.ast_tree
+    if tree is None:
+        return []
+    entrypoint_name = entrypoint_label(facts.entrypoint_name)
+    mutating_methods = {"append", "extend", "insert", "update", "setdefault", "__setitem__"}
+
+    objects = {
+        stmt.targets[0].id
+        for stmt in tree.body
+        if isinstance(stmt, ast.Assign)
+        and len(stmt.targets) == 1
+        and isinstance(stmt.targets[0], ast.Name)
+        and isinstance(stmt.value, ast.Call)
+    }
+
+    def _param_names(node: ast.FunctionDef | ast.AsyncFunctionDef) -> set[str]:
+        params = {arg.arg for arg in node.args.args}
+        params.update(arg.arg for arg in node.args.posonlyargs)
+        params.update(arg.arg for arg in node.args.kwonlyargs)
+        return params
+
+    def _input_derived(expr: ast.AST | None, params: set[str]) -> bool:
+        return bool(_expr_names(expr) & params)
+
+    def _object_attr(expr: ast.AST | None) -> tuple[str, str] | None:
+        if isinstance(expr, ast.Subscript):
+            return _object_attr(expr.value)
+        if (
+            isinstance(expr, ast.Attribute)
+            and isinstance(expr.value, ast.Name)
+            and expr.value.id in objects
+        ):
+            return expr.value.id, expr.attr
+        return None
+
+    def _mutated_attr(expr: ast.AST, params: set[str]) -> tuple[str, str] | None:
+        if isinstance(expr, ast.Assign) and _input_derived(expr.value, params):
+            for target in expr.targets:
+                attr = _object_attr(target)
+                if attr:
+                    return attr
+        if not isinstance(expr, ast.Call):
+            return None
+        if not (
+            isinstance(expr.func, ast.Attribute)
+            and expr.func.attr in mutating_methods
+        ):
+            return None
+        attr = _object_attr(expr.func.value)
+        if attr and any(_input_derived(arg, params) for arg in expr.args):
+            return attr
+        return None
+
+    for node in ast.walk(tree):
+        if not isinstance(node, (ast.FunctionDef, ast.AsyncFunctionDef)):
+            continue
+        if not is_entrypoint_name(node.name):
+            continue
+
+        params = _param_names(node)
+        mutated = {
+            attr
+            for expr in ast.walk(node)
+            for attr in [_mutated_attr(expr, params)]
+            if attr is not None
+        }
+        if not mutated:
+            continue
+
+        for child in ast.walk(node):
+            if not isinstance(child, ast.If):
+                continue
+            if _body_has_calls(child.body):
+                continue
+            if _expr_names(child.test) & params:
+                continue
+            tested = {
+                attr
+                for expr in ast.walk(child.test)
+                for attr in [_object_attr(expr)]
+                if attr is not None
+            }
+            returned = {
+                attr
+                for stmt in child.body
+                if isinstance(stmt, ast.Return)
+                for attr in [_object_attr(stmt.value)]
+                if attr is not None
+            }
+            if tested & returned & mutated:
+                return [{
+                    "pattern": "OBJECT_ATTRIBUTE_REPLAY",
+                    "severity": "critical",
+                    "evidence": (
+                        f"{entrypoint_name} returns object attribute state "
+                        "populated from a prior input"
+                    ),
+                }]
+
+    return []
+
+
 RE_OBJECT_ID_DATA = re.compile(r"\bdata_id\s*=\s*id\s*\(\s*data\s*\)")
 RE_RESULT_BANK_SET = re.compile(r"_superbatch_results\s*\[\s*(?:did|data_id)\s*\]\s*=")
 RE_RESULT_BANK_RETURN = re.compile(r"return\s+_superbatch_results\s*\[\s*data_id\s*\]")
@@ -3479,6 +3584,10 @@ class RulePolicy:
         "LAST_CALL_REPLAY", "result_reuse", "hard", AUTO_FILTER, (),
         AMD_MANUAL_LAST_CALL_REPLAY_FIXTURES, "keep",
     ),
+    "OBJECT_ATTRIBUTE_REPLAY": RulePolicy(
+        "OBJECT_ATTRIBUTE_REPLAY", "result_reuse", "hard", AUTO_FILTER, (),
+        (), "keep",
+    ),
     "SHAPE_OUTPUT_REPLAY": RulePolicy(
         "SHAPE_OUTPUT_REPLAY", "result_reuse", "hard", AUTO_FILTER, (),
         AMD_PACKAGE_SHAPE_REPLAY_FIXTURES, "keep",
@@ -3762,6 +3871,7 @@ def support_only_patterns(matched_patterns: list[dict]) -> bool:
     detect_decode_mm_ref,
     detect_result_caching,
     detect_last_call_replay,
+    detect_object_attribute_replay,
     detect_shape_output_replay,
     detect_timed_input_replay,
     detect_cuda_graph_replay,
@@ -3800,6 +3910,7 @@ def support_only_patterns(matched_patterns: list[dict]) -> bool:
     ("decode_mm_ref", detect_decode_mm_ref),
     ("result_caching", detect_result_caching),
     ("last_call_replay", detect_last_call_replay),
+    ("object_attribute_replay", detect_object_attribute_replay),
     ("shape_output_replay", detect_shape_output_replay),
     ("timed_input_replay", detect_timed_input_replay),
     ("cuda_graph_replay", detect_cuda_graph_replay),
@@ -4696,7 +4807,7 @@ def _worker_parquet(args: tuple) -> dict:
     "EVALUATOR_EXPLOIT", "HARNESS_RUNTIME_PATCHING", "MODULE_MUTATION", "GLOBALS_MUTATION", "CODE_REPLACEMENT",
     "FRAME_WALK_ACCESS", "FRAME_WALK_MUTATION", "SYS_MODULES_ACCESS", "GLOBALS_ACCESS", "CODE_ACCESS",
     "TRUSTED_MODULE_IMPORT",
-    "OUTPUT_REPLAY_CACHE", "LAST_CALL_REPLAY", "SHAPE_OUTPUT_REPLAY", "TIMED_INPUT_REPLAY", "CONFIG_CACHE_EXPLOIT", "POINTER_REPLAY", "RESULT_BANK_REPLAY", "PREPROCESS_CACHE", "WORKSPACE_CACHE",
+    "OUTPUT_REPLAY_CACHE", "LAST_CALL_REPLAY", "OBJECT_ATTRIBUTE_REPLAY", "SHAPE_OUTPUT_REPLAY", "TIMED_INPUT_REPLAY", "CONFIG_CACHE_EXPLOIT", "POINTER_REPLAY", "RESULT_BANK_REPLAY", "PREPROCESS_CACHE", "WORKSPACE_CACHE",
     "RUNNER_PLAN_CACHE", "CUDA_GRAPH_PYTHON", "CUDA_GRAPH_REPLAY",
     "TIMER_MONKEYPATCH", "FAKE_BENCHMARK_EMIT", "STDIO_REDIRECT", "UNSYNC_MULTISTREAM", "CUDA_EVENT_DISABLE_TIMING",
     "SCALED_MM_REF", "DECODE_MM_REF", "SILENT_FALLBACK", "REFERENCE_PRECOMPUTE_REPLAY", "TORCH_COMPILE_CACHE",