Add direct messages between friends after the social MVP

## Goal

Let accepted friends coordinate outside a room after the friendship, blocking, notification, retention, and abuse controls have proven stable.

## Required contract before implementation

- Only accepted friends may start or continue a conversation.
- Blocking immediately prevents new delivery and hides actionable contact paths.
- Define retention, deletion, export, and account-removal behavior before storing messages.
- Add reporting/moderation and rate limits before broad enablement.
- Treat message content as sensitive: do not put it in analytics, routine logs, notification previews, or PostgreSQL metrics.
- Reconnect and multi-device delivery must reconcile from durable paginated history; Socket.IO is delivery, not the source of truth.

## Acceptance criteria

- [ ] Complete a threat/privacy review covering minors, harassment, impersonation, spam, and moderation access.
- [ ] Define conversation/message schemas, cursor pagination, delivery/read semantics, and retention.
- [ ] Add authenticated APIs and realtime delivery with per-message authorization and idempotency.
- [ ] Add block/report/delete flows before enabling message composition.
- [ ] Add limits for message size, frequency, links, and duplicate/replayed sends.
- [ ] Add tests for unfriending/blocking races, ID tampering, reconnect, multiple tabs, retention, deletion, and rate limits.
- [ ] Roll out behind a separate feature flag with privacy-safe reliability metrics.

## Dependencies

- #75
- #105
- #188

## Non-goals

- Group chat.
- File/image uploads.
- End-to-end encryption claims without a separately reviewed design.
- Shipping as part of the Friend System MVP.

Inizia esaminando le dipendenze #75, #105 e #188, quindi completa la revisione richiesta delle minacce e della privacy prima dell’implementazione. Il lavoro sarà considerato completato quando saranno stati affrontati gli schemi, le API, i flussi di moderazione e sicurezza, i test e il rollout controllato tramite feature flag indicati; la composizione dei messaggi deve rimanere disabilitata finché le misure di protezione non saranno in atto.

Scritto dal modello di indicizzazione a partire dal testo della issue.