More developer notes:

• UX is built around rivers filtered to a single "Post" content type. Other stuff can be created (and you can get to listings of, say, files or whatever) but it's off the main path. Groups have almost the same UI to encourage jumping into the discussion.
• In hindsight trying to reuse page handlers by making river queries hookable made development fast, but less readable and more chaotic to manage.
• Uses a hook-based capabilities API to handle more complex permissions decisions
• Users can have multiple roles within different agencies and those roles tie into capabilities. This is a lot more expressive than, say, Arck Roles, but at a huge price since the API has to be laboriously baked into actions/menus. There's no UI for anything but the code is pretty readable.
• The Discussions feature uses groups with a custom subtype, and core modifications are needed to keep these out of the /groups listings, which is a bummer. Allowing most queries to be hook alterable would help (if it's worth the big growth of API surface area).
• The core mods stay on our custom production git branch. Upgrades are merges into that.
• We're using Plupload for async, chunked uploads and transcoding video to WebM/H.264.
• The Add New User form auto-generates a username from display name/e-mail, offers a random password, and allows bypassing the notification.
• We have Etherpad lite integrated (as pages), but are still on Apache 2.2 where websockets are way buggy. Etherpad UX just isn't that great and UI customizations is a big pain.
• The code changes for upgrades to 1.10, 1.11, and 1.12 have been completed, sitting in git branches awaiting a server upgrade. The site began life on an early 1.9 beta snapshot, and I won't do that again!
• Way too many interdependencies among plugins. We have a plugin for each big logical feature, but these aren't really reusable anyway so it arguably just makes code harder to find. Git history is a lot better for finding out how features were added and where code lives.