youknowone · Oct 8, 2026 · Oct 8, 2026
diff --git a/crates/vm/src/object/core.rs b/crates/vm/src/object/core.rs
@@ -173,17 +173,14 @@ pub(super) unsafe fn default_dealloc<T: PyPayload>(obj: *mut PyObject) {
         return; // resurrected by __del__
     }
 
-    // Only tracked objects take the trashcan recursion guard and untrack path.
-    // Untracked objects either own no children (int, float, str, ...) or, like
-    // non-escaped frames, are released at interpreter depth with at most one
-    // unguarded link before their tracked children (dicts, functions, code)
-    // re-enter guarded deallocation, so recursion stays bounded. A frame stored
-    // in an object graph is forced to escape, becoming tracked and guarded here.
-    // Read once and reuse for both gates below.
+    // Untracked payloads can still own recursive chains of references, such as
+    // detached frames returned by sys._current_frames(). Guard traversable
+    // payloads regardless of tracking, while keeping scalar deallocation cheap.
     let tracked = obj_ref.is_gc_tracked();
+    let guarded = tracked || T::HAS_TRAVERSE;
 
     // Trashcan: limit recursive deallocation depth to prevent stack overflow
-    if tracked && !unsafe { trashcan::begin(obj, default_dealloc::<T>) } {
+    if guarded && !unsafe { trashcan::begin(obj, default_dealloc::<T>) } {
         return; // deferred to queue
     }
 
@@ -262,7 +259,7 @@ pub(super) unsafe fn default_dealloc<T: PyPayload>(obj: *mut PyObject) {
     }
 
     // Trashcan: decrement depth and process deferred objects at outermost level
-    if tracked {
+    if guarded {
         unsafe { trashcan::end() };
     }
 }

diff --git a/extra_tests/snippets/stdlib_threading_current_frames.py b/extra_tests/snippets/stdlib_threading_current_frames.py
@@ -97,4 +97,48 @@ def f123():
     leave.set()
     t.join()
 
+
+def release_deep_snapshot():
+    # Detached snapshots can own a deep chain of untracked frames. Releasing
+    # one on a small native stack must not recurse through the entire chain.
+    depth = 5000
+    ready = threading.Event()
+    release = threading.Event()
+    disposed = threading.Event()
+
+    def recurse(n):
+        if n:
+            recurse(n - 1)
+        else:
+            ready.set()
+            release.wait()
+
+    def inspect():
+        frame = sys._current_frames()[worker.ident]
+        del frame
+        disposed.set()
+
+    old_limit = sys.getrecursionlimit()
+    old_stack = threading.stack_size()
+    worker = threading.Thread(target=recurse, args=(depth,), daemon=True)
+    try:
+        sys.setrecursionlimit(depth + 100)
+        worker.start()
+        assert ready.wait(30), "worker did not reach the bottom of its stack"
+        threading.stack_size(256 * 1024)
+        observer = threading.Thread(target=inspect, daemon=True)
+        observer.start()
+        observer.join(30)
+        assert not observer.is_alive(), "snapshot disposal did not finish"
+        assert disposed.is_set(), "snapshot disposal failed"
+    finally:
+        threading.stack_size(old_stack)
+        release.set()
+        worker.join(30)
+        sys.setrecursionlimit(old_limit)
+    assert not worker.is_alive(), "worker did not finish"
+
+
+release_deep_snapshot()
+
 print("ok")