fix(cluster): detect stale sandbox base image in local cluster #335

Bug Report

Agent Diagnostic

During development of #268 (bypass detection), the sandbox base image was updated in the openshell-community repo to include iptables (OpenShell-Community#36). Despite the upstream image being rebuilt and pushed to ghcr.io, the local k3s cluster continued using the cached copy without iptables — causing bypass detection to silently degrade.

The only fix was manually evicting the image: openshell doctor exec -- crictl rmi ghcr.io/nvidia/openshell-community/sandboxes/base:latest

Expected Behavior

The local cluster should detect when base:latest (or any :latest-tagged sandbox image) has a newer digest available upstream, and either:

1. Automatically re-pull on sandbox creation (preferred for :latest tags)
2. Warn the user that the cached image is stale

Current Behavior

k3s caches the image after the first pull and never checks the registry again for :latest tag updates. The imagePullPolicy for sandbox pods is not set to Always, so kubelet uses the cached copy.

Possible Fix

Set imagePullPolicy: Always on sandbox pod specs when the image tag is latest (or untagged). This matches standard Kubernetes convention and ensures :latest always gets a fresh pull. For pinned tags (e.g., :v1.2.3), the current IfNotPresent behavior is correct.

The relevant code is in crates/openshell-server/src/sandbox/mod.rs where the pod spec is constructed.

Reproduction

1. Deploy a local cluster
2. Create a sandbox (pulls base:latest)
3. Update the base image upstream (e.g., add a package)
4. Create another sandbox — still uses the old cached image