Summary:
Previously, this check was subtly wrong, causing it to fail for empty
commits (which have the same manifest as their parents). Update the check to
ensure the actual thing we want to ensure -- that the caller isn't creating
incorrect commits.
The failure would happen in this case:
```
B empty commit (manifest Am)
|
A draft commit (manifest Am)
/
M master (manifest Mm)
```
When metaediting `A`, `metaedit` would then "rebase" `B` onto `A'` to create
`B'`, with both `A'` and `B'` reusing `Am` as their manifest. The old check
would fail here because the parent of `Am` is `Mm`, but the manifest of `A'`
(the parent of `B'`) is `Am`, so when checking the manifest of the parent
against the parent of the manifest, we would find that `Am != Mm` and fail.
The actual check we wanted was that the manifest of the parent of `B'` is the
same as the manifest of the parent of `B` -- which the old check approximated
as long as no manifests were reused in the stack. As this is not sufficient
with empty commits, we instead want to check that the manifest of the old
parent matches the manifest of the new parent, which we now do.
By avoid reading `manifest.parents`, this would also avoid some lazy tree
fetches, therefore speed up metaedit in certain cases.
Reviewed By: quark-zju
Differential Revision: D9013995
fbshipit-source-id: 4d0b05fc9bb81d115cd87ba2bf98aa253ae6f88b
This was originally fixed by Mateusz Kwapich for the `metaedit` command in
fb-hgext with a test for the `metaedit` command. It didn't get upstreamed
because `metaedit` was not in core.
This patch fixes the crash and adds a test about `metadataonlyctx` to
avoid future regressions.
Differential Revision: https://phab.mercurial-scm.org/D550
