Add FUTURE.md: vision, architecture, and roadmap

FUTURE.md

@@ -0,0 +1,78 @@
+# Future Vision
+
+WASTE's design philosophy is still sound: small trusted groups, no central server, encrypted everything, equal nodes. That's essentially what Signal's sealed sender and private groups do today, just without a self-hosted option. This gap is worth filling.
+
+---
+
+## Architecture
+
+Two clean layers, connected by the IPC port.
+
+### Daemon
+The real application. A long-running background process that handles everything:
+
+- Peer mesh and connection management
+- Cryptography and handshake
+- NAT traversal and relay fallback
+- File transfer
+
+Exposes a local JSON API over TCP (`127.0.0.1:17337`). Can run headlessly — SSH into a box and the mesh stays alive even with no UI attached.
+
+### UI Layer
+Talks to the daemon over the IPC port. The separation means the UI is replaceable without touching the core.
+
+Target: a web frontend (React or similar) wrapped in a native binary using a Tauri-style approach — native packaging, OS webview, no Electron weight. Avoids the wxWidgets ugliness of the old wxWASTE fork and the Qt licensing headaches of the VIA fork.
+
+---
+
+## Protocol Modernization
+
+### NAT Traversal
+The main unsolved problem from the original WASTE — one party always needed an open port.
+
+- Try UDP hole punching (STUN) first
+- Fall back to an encrypted relay (DERP-style) when hole punching fails
+- Relay sees only opaque encrypted blobs — end-to-end encryption holds
+- Run the relay on a Hetzner VPS; `waste-relay` already implements the blind-forward model
+
+### Bootstrapping & Rendezvous
+No DHT needed at small group scale (10–50 nodes). Keep it simple:
+
+- Each peer generates an Ed25519 keypair on first run — the public key **is** their identity
+- Share a small signed invite file (`.waste-invite`) out of band: email, Signal, whatever
+- The invite contains: current IP:port hint + public key + short-lived signature
+- Once two peers connect, they gossip each other's addresses to mutual friends
+- A local known-peers list in the data directory is sufficient at this scale
+
+### Identity
+- Persistent Ed25519 keypair, generated once
+- Public key = stable identity, not a mutable nickname
+- No phone number, no central registry — closer to Signal's model than WASTE's original unregistered aliases
+
+### Transport (Long-term)
+Current transport is TCP with custom framing. QUIC is worth revisiting once the core is solid — it gives multiplexing and better NAT traversal behavior essentially for free.
+
+---
+
+## Roadmap
+
+| Priority | Item |
+|---|---|
+| 1 | Deploy `waste-relay` to Hetzner; verify cross-internet NAT traversal |
+| 2 | Invite file format (`.waste-invite`) — solve bootstrapping without manual IP sharing |
+| 3 | Peer gossip — auto-connect to friends-of-friends after initial invite |
+| 4 | File transfer — chunked, encrypted, resumable |
+| 5 | Message persistence — SQLite via `modernc.org/sqlite` |
+| 6 | UI — web frontend consuming the IPC port; native packaging |
+| 7 | UDP hole punching — full STUN implementation in `internal/nat` |
+| 8 | QUIC transport — replace TCP framing for better NAT behavior |
+
+---
+
+## What to Keep from WASTE
+
+- **Small group** — not a public network, not federated, not discoverable
+- **No registration** — no phone number, no email, no central service
+- **Encrypted everything** — at rest and in transit, end to end
+- **Equal nodes** — no peer is "the server"; the relay is dumb infrastructure only
+- **The soul** — a private overlay for people you actually trust