- For albedo_openai_transport, we race both v4+v6 Happy Eyeballs style, picking the first to connect
- For albedo_http, we tell httpc to use ipv6fb4 mode, which tries 6 first and falls back to 4.
The best approach would probably for the streaming transport (which really isn't OpenAI specific) to move into albedo_http and for everything to use one HTTP transport. That's a bigger lift than I wish to make without consultation