diff --git a/rlix/pipeline/miles_pipeline.py b/rlix/pipeline/miles_pipeline.py
index 18e7195..474fe52 100644
--- a/rlix/pipeline/miles_pipeline.py
+++ b/rlix/pipeline/miles_pipeline.py
@@ -674,17 +674,20 @@ def _after_training(self, step: int) -> None:
         # Coordinator drives sync_base_weights_to_active under its
         # resize lock; pipeline NEVER calls service / finalize /
         # set_weight_version directly (F05).
-        ray.get(
-            self._coordinator_handle.sync_base_weights_to_active.remote(int(step))
-        )
-        # Release the actor_train allocation back to the scheduler so
-        # other pipelines can step. R11-F1: only flip the ledger flag
-        # after a SUCCESSFUL release.
-        released = self._notify_release_cluster_gpus(
-            cluster_id=self._actor_train_cluster_id, global_step=int(step)
-        )
-        if released:
-            self._actor_train_allocated = False
+        try:
+            ray.get(
+                self._coordinator_handle.sync_base_weights_to_active.remote(int(step))
+            )
+        finally:
+            # Release the actor_train allocation back to the scheduler so
+            # other pipelines can step even if weight sync fails after the
+            # train actor has offloaded. R11-F1: only flip the ledger flag
+            # after a SUCCESSFUL release.
+            released = self._notify_release_cluster_gpus(
+                cluster_id=self._actor_train_cluster_id, global_step=int(step)
+            )
+            if released:
+                self._actor_train_allocated = False
 
     # ------------------------------------------------------------------
     # M4 minimal hard cleanup
diff --git a/tests/test_miles_pipeline_after_training_cleanup.py b/tests/test_miles_pipeline_after_training_cleanup.py
new file mode 100644
index 0000000..c0b2839
--- /dev/null
+++ b/tests/test_miles_pipeline_after_training_cleanup.py
@@ -0,0 +1,39 @@
+from __future__ import annotations
+
+import ast
+from pathlib import Path
+
+
+REPO_ROOT = Path(__file__).resolve().parents[1]
+
+
+def _function_def(tree: ast.AST, name: str) -> ast.FunctionDef:
+    for node in ast.walk(tree):
+        if isinstance(node, ast.FunctionDef) and node.name == name:
+            return node
+    raise AssertionError(f"missing function {name}")
+
+
+def _calls_attr(node: ast.AST, attr: str) -> bool:
+    return any(isinstance(child, ast.Attribute) and child.attr == attr for child in ast.walk(node))
+
+
+def test_after_training_releases_train_gpus_when_weight_sync_fails() -> None:
+    source = (REPO_ROOT / "rlix" / "pipeline" / "miles_pipeline.py").read_text(
+        encoding="utf-8"
+    )
+    tree = ast.parse(source)
+    after_training = _function_def(tree, "_after_training")
+
+    guarded_sync = [
+        node
+        for node in ast.walk(after_training)
+        if isinstance(node, ast.Try)
+        and any(_calls_attr(stmt, "sync_base_weights_to_active") for stmt in node.body)
+    ]
+
+    assert guarded_sync, "sync_base_weights_to_active must be guarded by try/finally"
+    assert any(
+        any(_calls_attr(stmt, "_notify_release_cluster_gpus") for stmt in node.finalbody)
+        for node in guarded_sync
+    ), "actor_train release must run from the sync finally block"