Card: shm_broadcast hang triage — pidfd_getfd/seccomp/Yama cause from the offload PR thread; Marlin note scoped to this image

README.md

@@ -124,6 +124,7 @@ the working set, not the table.
   add a third mount:
   `-v $PWD/connector_mrv2.py:/usr/local/lib/python3.12/dist-packages/vllm/v1/ple_offload/connector.py:ro`.
   Verified with the overlays here: normal boot, sanity and accuracy unchanged.
+- Same hang, different cause: the GPU worker and the offload subprocess exchange a CUDA IPC file descriptor with `pidfd_getfd(2)`; where seccomp denies it (a managed container host without `CAP_SYS_PTRACE`) or Yama's default `ptrace_scope=1` blocks it (bare metal), the boot hangs after graph capture with no error naming the cause. `--cap-add=SYS_PTRACE` on the `docker run`, or `sysctl kernel.yama.ptrace_scope=0` on bare metal; reports and a preflight probe: [one](https://github.com/vllm-project/vllm/pull/53899#issuecomment-5448518741), [two](https://github.com/vllm-project/vllm/pull/53899#issuecomment-5553027248). Not needed on our box or on the 64 GB-host field report.
 
 ---