Arcanist should be a tool not a process mandate. We self merge changes from time to time now and its fine. Mostly this comes up for deploys but is also a useful (expert) tool for small, new projects - especially when you are working alone and many weeks away from a first release. These are the realities as they stand now. If we want to change this then lets do that but I don't want us to change it simply because a tool forces it on us. I think it'd be more useful to switch to arcanist and _then_ have the discussion about whether or not we should turn off self merges.
AFAIK differential doesn't actually do the merge. Instead once a patch has been uploaded with arc and is ready for merge the reviewer downloads the patch locally via arc, merges it into their local branch and then pushes the reviewed change back to the canonical git repo (which can be hosted anywhere and is not tied to the review process directly).
Yes, it seems that you push a commit to review, and anyone can specify a reviewer for it (you can also do this from the commit message). Then when one of those reviewers accepts it, a command is shown for someone with edit rights to land the commit.
That's my understanding, yes. See the comment by @mmodell on D1#30. This makes sense actually because differential itself is completely agnostic of the the backing VCS in use. It is really just a structured discussion system tuned to the general process of code review. This is a stark contrast to Gerrit which is a highly opinionated git based review and control system.
For both OCG and the Collection extension it is very hard for me to find reviewers. And for Parsoid we often have a number of "jointly authored" patches where the self-review requirement gets very weak. Usually we go with "someone other than the last person who pushed a new version of the patch" but that's often the original author of the patch.
In both cases a *mandatory* ban on self-review would be very disruptive. I have no problem with making self-review a big red button that makes you click through "I know what I'm doing", or allowing certain projects to disable self-review, or to making self-review a "privilege" that only certain accounts were able to use. But blanket ban would be very disruptive.
I also don't like the fact that direct pushes can bypass review entirely. That option is present in gerrit but we *very rarely* use it (only when gerrit is hugely broken in some way, or to create new branches). I prefer to have all code changes documented in a consistent way. Even if it's a self-review, having the change in gerrit and +2'ed means that there's a paper trail and a mechanism for folks to be notified of the change and comment on it after the fact. I think that falling back to a cowboy-style direct-push-to-git would be a huge step back.
this is not really a valid task. there is nothing preventing self-merge inherent to arcanist. It's perfectly possible to merge and push, regardless of whether we use arcanist/differential.