[reliability] Relay-cost telemetry + closer-peer-first enforcement (DERP model)

[pvg13]

Problem

The relay is correctly payload-free (wake/fallback only) and DCUtR upgrades relayed→direct, but there is no telemetry on relay byte usage and nothing actively prefers/confirms the direct path once it exists. The "closer-peer-first" principle is implicit in libp2p connection racing, not measured or enforced — so relay cost is uncontrolled and invisible.

Proposed approach (Phase 0 telemetry → Phase 1 enforcement)

Adopt the Tailscale DERP discipline:

1. Meter relay vs direct bytes; expose a "% traffic over relay" gauge via NetworkEvent/diagnostics.
2. Once a DCUtR direct path is up, demote/detach the relay path for data so steady-state traffic leaves the relay.
3. Surface per-peer path (direct/relayed) for observability.

"You can't optimize what you don't measure" — the telemetry is the cheap prerequisite.

Files

wavesyncdb/src/diagnostics.rs, push.rs, peer_manager.rs, relay_manager.rs, behaviour.rs.

Ref: docs/research/sync-reliability.md §5 (DERP model) + §6 P1.