I should read the specs, but since it's such a foundational issue maybe someone who knows could respond briefly? the problem with a flat addressing space is that it requires every intermediate node to have state about every address, or perform a costly discovery mechanism for those it doesn't know about. is there a clever answer to this?

We have an answer, but it isn't really clever. We do have both built in and pluggable address lookup services.

Our default enabled address lookup service is using DNS in a creative way, but we also have a service that is fully peer to peer and is using the mainline DHT, specifically the bep_0044 extension that allows you to store a tiny bit of arbitrary data for an Ed keypair that you control.

https://www.bittorrent.org/beps/bep_0044.html

https://pkarr.org

Some custom transports such as TOR hidden services have a discovery system built in. In these cases we can just use the existing discovery system.

See for example https://github.com/n0-computer/iroh-tor-transport

The secret is that iroh still uses IPs under the hood :) But with QUIC, your connections aren't bound to your four-tuple, your connection can migrate from e.g. WiFi to Cellular with only a small blip/hiccup. And with QUIC multipath, you can have multiple four-tuples "active" at the same time. iroh uses e.g. a "real" IP path mainly, with a websocket-based HTTPS path via relay servers as the backup (e.g. in case UDP is blocked).