generating-followups

RLM-isolated follow-up generation: orchestrator + per-user subagent architecture where cross-user contamination is architecturally impossible.