A separate workflow on the linux_host runner deploys main to llm-node with its own checkout, /etc/shinkan/prod.env and a compose override that mounts the NAS media without creating host paths. Media retention stays off there until it is explicitly enabled, because both installations share the media pool. The deploy only passes when /api/health/ready reports a ready database and complete schema. The Pi workflow keeps its name and label and now aborts on any other host.
Also documents both deployments and the recovery contract with the backup orchestrator.
version: 0.8.246 (backend)
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>