Multi-node realtime consistency

## Summary
Make realtime behavior consistent across multiple nodes and processes while preserving user-visible parity.

## Scope
- Ensure message delivery, unread state, presence-like ephemeral state, and moderation-sensitive events behave correctly in multi-node deployments.
- Remove hidden single-node assumptions from realtime and cache paths.
- Define where authoritative state lives and how fanout/replay work across nodes.
- Keep the user-visible behavior stable during failover, rebalance, or reconnect.

## Acceptance Criteria
- [ ] Core chat behavior remains correct when multiple app nodes handle the same guild.
- [ ] Cross-node fanout and recovery do not duplicate or lose user-visible events.
- [ ] Ephemeral state has a clear multi-node coordination story.
- [ ] Operators can reason about failure modes and expected consistency boundaries.

## Operational Notes
- Self-hosted clusters need explicit support boundaries; this issue should not hand-wave distributed correctness.
- User-visible consistency matters more than preserving accidental single-node optimizations.