Label CUDA Graph backward annotations

## Human Note

## Agent note

Preserve the prototype annotation API's `autograd_phase` metadata when generating GPU annotation
boxes. Backward kernels inherited from a forward region now render as labels such as
`attention backward`, while unowned backward plumbing can retain the explicit `backward` label.

## Test Plan

```bash
~/.venvs/nightly/bin/python -m pytest test/test_profiler.py test/test_perfetto.py -q
uvx ruff check transformer_nuggets/utils/perfetto.py test/test_perfetto.py
uvx ruff format --check transformer_nuggets/utils/perfetto.py test/test_perfetto.py
prek
```

Author: drisspg

test/test_perfetto.py

@@ -73,6 +73,29 @@ def test_cuda_graph_annotations_become_contiguous_gpu_boxes():
     assert len(trace["traceEvents"]) == 3
 
 
+def test_cuda_graph_annotation_boxes_label_backward_phase():
+    graph_id = 2
+    trace = {
+        "traceEvents": [
+            {
+                "ph": "X",
+                "cat": "kernel",
+                "name": "kernel",
+                "pid": 0,
+                "tid": 7,
+                "ts": 10,
+                "dur": 3,
+                "args": {"graph id": graph_id, "graph node id": 1},
+            }
+        ]
+    }
+    annotations = {(graph_id << 32) | 1: [{"name": "attention", "autograd_phase": "backward"}]}
+
+    processed = add_cuda_graph_annotation_boxes(trace, annotations)
+
+    assert processed["traceEvents"][-1]["name"] == "attention backward"
+
+
 def test_cuda_graph_annotation_boxes_accept_monitor_embedded_metadata():
     trace = {
         "traceEvents": [

transformer_nuggets/utils/perfetto.py

@@ -66,13 +66,21 @@ def write_trace(path: str | Path, trace: dict[str, Any], *, indent: int | None =
         json.dump(trace, f, indent=indent)
 
 
-def _annotation_name(entries: Sequence[Any] | None) -> str | None:
-    name = None
+def _annotation_label(entries: Sequence[Any] | None) -> str | None:
+    metadata: dict[str, Any] = {}
     for annotation in entries or ():
-        if isinstance(annotation, dict) and "name" in annotation:
-            name = str(annotation["name"])
+        if isinstance(annotation, dict):
+            metadata.update(annotation)
         elif isinstance(annotation, str):
-            name = annotation
+            metadata["name"] = annotation
+
+    name = metadata.get("name")
+    if name is None:
+        return None
+    name = str(name)
+    autograd_phase = metadata.get("autograd_phase")
+    if autograd_phase == "backward" and name != "backward":
+        return f"{name} backward"
     return name
 
 
@@ -133,7 +141,7 @@ def add_cuda_graph_annotation_boxes(
                 embedded = None
             if isinstance(embedded, list):
                 entries = embedded
-        name = _annotation_name(entries)
+        name = _annotation_label(entries)
         if name is not None:
             annotated_work[(event.get("pid"), event.get("tid"))].append((event, name))