feat: sandbox-runtime integration with e2e Helm tests

[moshloop]

Summary

• Add bubblewrap-based sandbox support via @anthropic-ai/sandbox-runtime for secure template execution
• Configurable via --sandbox CLI flag and Helm chart values (sandbox.enabled, sandbox.settingsPath)
• Add ConfigMap for srt-settings with network/filesystem isolation config
• Add Ginkgo/Gomega e2e tests using commons-test/helm that build Docker image, install chart on Kind, and verify deployment
• Update CI workflow to run Go e2e tests instead of manual curl scripts

Test plan

• All 5 e2e tests pass locally (chart install, ConfigMap, healthz, render, sandbox disable/enable)
• CI builds Docker image and runs tests on Kind cluster
• Helm lint passes

In general, the problem is fixed by adding an explicit permissions block that scopes the GITHUB_TOKEN to the minimum required access. For this workflow, both jobs only need to read repository contents to check out the code; all other operations (Docker build, Kind cluster setup, Go tests, kubectl calls) happen locally or against the ephemeral cluster and do not require repository write permissions.

The best fix, without changing existing functionality, is to define a root-level permissions block alongside name and on, so it applies to both lint and e2e jobs. Setting it to contents: read is sufficient, as none of the steps need to write to the repo or other GitHub resources. Concretely, in .github/workflows/helm-test.yml, insert:

permissions:
  contents: read

between the name: line and the on: block (or just above jobs:; root-level is conventional and clear). No additional imports, actions, or other definitions are required.