I don't know why this happens exactly, but given an upstream and fork repository
from a customer, both of which required GC, resolving conflicts would corrupt
the fork so badly that it couldn't be cloned.
This isn't a perfect fix for that case, because the MR may still need to be
merged manually, but it does ensure that the repository is at least usable.
My best guess is that when we generate the index for the conflict
resolution (which we previously did in the target project), we obtain a
reference to an OID that doesn't exist in the source, even though we already
fetch the refs from the target into the source.
Explicitly setting the source project as the place to get the merge index from
seems to prevent repository corruption in this way.
When reading conflicts:
1. Add a `type` field. `text` works as before, and has `sections`;
`text-editor` is a file with ambiguous conflict markers that can only
be resolved in an editor.
2. Add a `content_path` field pointing to a JSON representation of the
file's content for a single file.
3. Hitting `content_path` returns a similar datastructure to the `file`,
but without the `content_path` and `sections` fields, and with a
`content` field containing the full contents of the file (with
conflict markers).
When writing conflicts:
1. Instead of `sections` being at the top level, they are now in a
`files` array. This matches the read format better.
2. The `files` array contains file hashes, each of which must contain:
a. `new_path`
b. `old_path`
c. EITHER `sections` (which works as before) or `content` (with the
full content of the resolved file).
- Add match line header to expected result for `File#sections`.
- Lowercase CSS colours.
- Remove unused `diff_refs` keyword argument.
- Rename `parent` -> `parent_file`, to be more explicit.
- Skip an iteration when highlighting.
