Agent startup services: systemd units for autonomous operation

Summary

Add systemd service units so agents can run autonomously on boot, each executing Claude with their PROMPT.md.

Requirements

1. Per-agent systemd service — e.g. agent-roster-bossypants.service that starts a Claude session with the agent's PROMPT.md
2. Environment setup — each service should have:
   • ANTHROPIC_API_KEY from secrets (depends on #7)
   • GITLAB_HOST and auth tokens
   • Correct HOME and USER
3. Lifecycle management — services should:
   • Start on boot (or be timer-activated)
   • Restart on failure with backoff
   • Be individually controllable (systemctl start/stop agent-roster-bossypants)

Design Considerations

• claude --headless or equivalent CLI mode for non-interactive operation
• Consider whether all agents run continuously or some are on-demand (timer-activated)
• Resource limits per agent (memory, CPU) to prevent runaway usage
• Logging per agent to separate journal units

Dependencies

• Depends on #7 (secret management) for API keys
• Depends on #8 (GitLab provisioning) for git access

Acceptance Criteria

• Systemd service template for agents
• Per-agent service instantiation from roster config
• Environment variables injected from secrets
• Restart policy with backoff
• Tested on earthrealm