What's up with Revlog/Mercurial repository?

* Directories aren't tracked. Required, and in fact easy to add on the
  Manifest level.

* Current implementation in Mercurial expects that it can store all
  files in memory? That's ... not very optimal with large files.

* Large files (> 4GiB) don't fit into the Revlog index.

* Not suitable for mixed-revision working copies, *unless* we invent
  an additional cache/index for the WC.

* Partial clone / history trimming is impossiblish -- but required for
  a not-quite-centralized VCS.

* Cherry-picking: merge history isn't complete enough.

* SHA-1 collisions are scary. Should use SHA-256, or better.

* Object identity across renames? (Manifest again, rather than Revlog proper.)

* Additional indexes (e.g., all files touched by user). Seems to be
  orthogonal to the basic repository backing-store issues.

Summary: Very nice idea, but needs work if we want to use it as a backing
store for a Subversion-like VCS.