The problem is that hg merge/update/etc work the following way:
1. figure out what files to update
2. apply the update to disk
3. apply the update to in-memory dirstate
4. write dirstate
where step3 looks at the filesystem and assumes it sees the result of
step2. If a file is changed between step2 and step3, step3 will record
incorrect information in the dirstate.
I avoid this by passing the size step3 needs directly from step2, for
the common path (not implemented for change/delete conflicts for
instance). The code is a bit awkward, as there doesn't seem to be a
standard way in python to return a value at the end of a generator.
I didn't fix the same race for the exec bit for now, because it's less
likely to be problematic and I had trouble due to the fact that the
dirstate stores the permissions differently from the manifest (st_mode
vs '' 'l' 'x'), in combination with tests that pretend that symlinks
are not supported.
However, I moved the lstat from step3 to step2, which should tighten
the race window markedly, both for the exec bit and for the mtime.