Identifying merge conflicts is unreliable

[sivel]

I haven't looked into it in much depth, but the mergeable and mergeable_status attributes seem unreliable. One run may identify 100 PRs with merge conflicts, the next could return 4.