fix: harden retention cleanup safety

backend/service/storage_metrics.py

@@ -1,8 +1,8 @@
 #!/usr/bin/env python3
-"""Inspect and optionally clean stale job directories.
+"""Inspect retention candidates and optionally clean verified managed jobs.
 
-Default mode is a safe dry-run that reports reclaimable bytes. Pass --apply to
-actually remove unreferenced job directories.
+Default mode is a safe dry-run. Metadata-less workspaces are reported for
+manual investigation but are never removed by ``--apply``.
 """
 
 from __future__ import annotations
@@ -46,7 +46,11 @@ def referenced_job_ids() -> set[str]:
 def main() -> None:
     parser = argparse.ArgumentParser(description=__doc__)
     parser.add_argument("--max-age-days", type=float, default=30)
-    parser.add_argument("--apply", action="store_true", help="Actually delete stale job directories")
+    parser.add_argument(
+        "--apply",
+        action="store_true",
+        help="Delete only CleanupService-verified candidates; orphan workspaces remain report-only",
+    )
     parser.add_argument("--json", type=Path, default=None, help="Write JSON report")
     args = parser.parse_args()
 
@@ -89,10 +93,11 @@ def main() -> None:
         if os.environ.get("CLEANUP_APPLY_ENABLED", "false").strip().lower() != "true":
             parser.error("--apply requires CLEANUP_APPLY_ENABLED=true")
         applied = cleanup.apply(datetime.now(timezone.utc))
-        freed = 0
-        for item in orphaned:
-            freed += storage.remove_job_dir(item["job_id"])
-            store.delete_job(item["job_id"])
+        freed = sum(
+            item.size_bytes
+            for item in applied.temporary_jobs
+            if item.job_id in applied.deleted_job_ids
+        )
         print(
             "freed_bytes="
             f"{freed} deleted_job_ids={applied.deleted_job_ids} "