lgtm

What _is_ the difference between "conclude" and "commit"?

btw, @quark mentioned there could be a perf hit (@sid0 mentioned this too a long time ago iirc)

Strong +1

Nice.

I added a comment and also filed T24918295 for the wider task around allowing lazy filectx writes (but as Durham mentioned that's a strict improvement on the status quo).

Should we even be calling _makebackup in the case of an inmemory merge? Like, maybe the makebackup should be conditional based on if the source file context is actually a workingctx?

I'll add a comment.

lgtm

Based on Messenger discussion, this version just skips the manifest verification and future steps which require it, but keeps changeset verification. Can probably get by without upstream refactor.

@yuja is that just because of the _manifest property cache, or are there others as well?

(Need to look at https://phabricator.intern.facebook.com/D6558110 too which is touching verify)

OK, extracting just the manifest verification will take a bit of upstream refactoring. I'll send an upstream change to do that + a followup to scope this to just changesets. But people are still running verifies that take 4 days and make tons of packfiles, so I'm still +1 on landing this while that's happening.

I can do that, we just wanted to get a MVP out ASAP since people are hurtin'.

Added a test, description and moved to config.

A better way might just be to use None for self.wtcx when rebasing on disk, which will cause fresh wctxs to be used every time and feel less icky. We could use self.inmemory instead of self.wctx.isinmemory().