Multiple issues/bugs doing dual hermes to one shared LLM

### Investigation Summary

Running into port propogation issues and bugs trying to run 2 hermes to one shared LLM on a dgx spark.
I've had probably 10-15 painful rebuilds, onboards and starting fresh. All in an architecture that is supposed to be by design.

### Description

I have a dgx/hermes troubleshooter project that searches through new features, bugs, and forum posts. It wanted me to make posts through all these repos but seems pain staking. Consolidating here.

Report Repo
Port propagation (#1+#8) [NVIDIA/NemoClaw](https://github.com/NVIDIA/NemoClaw/issues/new/choose)
Rebuild regression (cite #2273/#2366) NVIDIA/NemoClaw
snapshot create hardlink rejection NVIDIA/NemoClaw
policy set h2 crash [NVIDIA/OpenShell](https://github.com/NVIDIA/OpenShell/issues) — different repo

### Reproduction Steps

try to onboard to hermes to one LLM docker. You'll break on one, then have to go to the other. Then switch ports and the other breaks. Then stop one to start the other. Then starting the other it breaks... painful...

### Environment

DGX Spark, nemohermes

### Debug Output

[dgx-spark-forum-post.md](https://github.com/user-attachments/files/31355018/dgx-spark-forum-post.md)

### Logs

```shell

```

### Checklist

- [x] I confirmed this bug is reproducible
- [x] I searched existing issues and this is not a duplicate

Start with the attached dgx-spark-forum-post.md and reproduce the reported two-Hermes/shared-LLM setup on a DGX Spark. Separate the port propagation, rebuild regression, snapshot hardlink, and policy crash reports, then document a reproducible failure and identify which repository owns each fix.

Written by the indexing model from the issue text.