1. Most of the big problems with IRC as a protocol (nicknames, ghosting, missed messages) come from the assumption that the core of the network should be stateless. If the core of the network is stateless, you can't have nickname reservation, you can't transparently attach two clients to the same nickname, and you can't replay missed messages to people who were disconnected (hence the need for clients to maintain a TCP connection to the server at all times, hence netsplits being disruptive, etc.).
2. Conventional IRC setups solve some of these problems by adding a single privileged node, the "services framework" (Anope and Atheme are examples), that stores some persistent state (typically, nickname and channel reservations). This actually abandons the high-availability properties of the original IRC design: it is theoretically possible to design a highly available services framework (using a clustered database or whatever) but AFAICT none of the frameworks that currently exist are HA.
3. Oragono is a single instance that provides an integrated IRC server, services framework, and "bouncer" (history retention and playback). Client connectivity problems are solved by allowing transparent reattach to the original nickname after authentication with SASL, then automatically replaying history. (Client support for this is still patchy [2], but anything that supports znc.in/playback [3], like Textual, will work well; you can also configure Oragono to try and track what you missed and replay it on reconnection.)
4. To make Oragono highly available, it can be deployed in Kubernetes (virtualizing the embedded database file and the external IP, spinning up a replacement instance on failure).
[1] https://github.com/oragono/oragono
I had looked at IRCv3 briefly when starting the project, but it looked like IRCv3 didn’t solve netsplits.
Have I overlooked anything (or has something changed), or did you just ask out of curiosity?
Keep me posted regarding how RobustIRC is going for your network!
Darned CAP theorem.
I've been looking at this system recently and I'm going to be including robust IRC access to my network (shameless plug[1]).
The code is clean, I think I have a good understanding of how it works and from my testing it does seem to work really well!
That said, have you looked at how much of your original problem space is solved by IRCv3?
[0]: https://github.com/robustirc/bridge/blob/master/AUTHORS#L10
[1]: ircs://irc.darkscience.net:6697/darkscience