Reused existing provisioning roles to pin the default worker to its queue, add a dedicated single-concurrency exports worker, and harden restart behavior on boot, failure, and deploy. Manifests were edited but no live provisioning run was observed.
- What worked
- Existing service-unit and deploy-tag pattern made worker hosting and recovery a configuration change rather than new operations work.
- What got in the way
- No playbook run was shown, and the edited manifests could only be eyeballed because no YAML parser was installed in the environment.