Direct UDP as an upgrade target

[lann]

The route-table upgrade model (#11, 27f0ea7) moves relay-dialed connections onto WebRTC channels; UDP stayed a dial-time-only path. Upgrading relay→UDP needs what the channel got for free from signaling: attributing inbound datagrams on the shared socket to a peer before QUIC decrypts (disco-style addressing), plus a reachability probe before flipping the route. Related: #2's QAD question (address discovery is the other half of learning a peer's ip:port), and the flip-back gating noted for #4 (path health applies to any non-relay route). Until then, ip entries dial direct or not at all — a recorded narrowing at connect.

[lann]

Post-#21 (the noq-proto migration), two seams in the endpoint are the wiring points for this issue, recorded here so the latitude is findable:

• The event arms. endpoint_impl.rs (and the spike pump) match connection events exhaustively with explicit no-op arms: Event::Path(_) | Event::NatTraversal(_) => {}. noq-proto delivers the NAT-traversal and path machinery through exactly those events — this issue's implementation replaces the no-ops rather than hand-building the extension (the inheritance Upstream iroh crate as endpoint core: feasibility and the ecosystem-compatibility ruling #14's step-3 comment records).
• The path-zero latitude. connection.path reads network_path(PathId::ZERO) and keys the route table on that one remote: the current model treats every connection as single-path, with wire upgrades as route flips under a stable standin address. Wiring QUIC-native multipath/NT means real second paths with real addresses coexisting with synthetic standins — how the route table, the standin allocator, and noq's path selection divide that responsibility needs a ruling before code, since it decides whether path keeps reporting a single kind or becomes a per-path report.

[lannbot]

Routing two #13 findings here, since they are this issue's design evolution rather than tail cleanup:

• C11 (single-path model): connection.path — and the path-changes watch connection.path-changes: a coalescing watch replaces upgrade polling #57 added — assume exactly one wire at a time. iroh 1.0 connections hold a set of paths with a selected one (paths(), path_events(), per-path RTT), and the current model cannot even express "relay + webrtc simultaneously during upgrade," which is what the implementation actually does mid-flip. Direct-UDP-as-upgrade is the point where this stops being tolerable: the surface likely becomes list<path> plus events, and path-changes' element type evolves with it (its coalescing-watch contract carries over).
• C13's self-address line: there is still no way to learn observed addresses — direct-addr is the local bind — so assembling a dialable endpoint-addr record for publishing is impossible until this issue's QAD-equivalent lands. Discovery work is sequenced behind that.