obslog: add a --origin flag to show predecessors instead of successors This flag will be turned on by default in the next changeset.

obslog: add preliminary support for showing predecessors This functionality is not yet used right now, but will be in the next patch.

obshistory: add _originmarkers() function to be used for obslog --origin

evolve: rename `evolve_source_*` to `divergence_source_*` This makes it clearer that this comes from divergence resolution.

evolve: mark new hash-salting extras keys as uninteresting for effect flag The user doesn't care if the `evolve_source_local`/`evolve_source_other` values in the extras changed, so this patch adds them to the blacklist that already exists for that purpose. This should also fix the flakiness we've seen in some tests. The flakiness seemed to stem from `mercurial.obsutil.geteffectflag()` not sorting extras keys. That means we ran into it relatively often for the `evolve_source_local`/`evolve_source_other` keys because there were two of them. That sorting bug should be fixed upstream.

evolve: handle relocation during divergence resolution producing no changes When resolving divergence and the two divergent commits have different parents, we start by rebasing one of them to have the same parent as the other. That step can result in no changes to commit. When it does, we would crash with a TypeError before this patch. This patch fixes it by instead creating an empty commit in that scenario. The existing code then continues to attempt to merge it, which produces no changes, and no commit is created on top. The other side of the divergence is marked as successor as usual, so orphans from the side that became empty will be evolved to the right place (see test).

tests: show crash from divergence resolution resulting in empty commit When relocating a commit results in an empty commit (because the changes are already in the destination), we get a TypeError from evolve.

evolve: always create commit when resolving divergence When resolving content-divergence, the final commit we create may end up empty (which means that Mercurial won't even create it). We've had code for handling that in evolve ever since 41bf6c27a122 (evolve: stabilize now handle conflicting changeset, 2012-08-23). However, that resolved the issue by marking on the divergent commits as successor. As Pierre-Yves has pointed out (in other code reviews), we should instead be creating a new successor. So that's what this patch does. It does that by setting `ui.allowemptycommit` while creating the final commit. However, that is not enough, because we may end up creating the same nodeid as already existed (we'd then end up trying to mark the "new" commit a successor of itself). To solve that, we add some salt to the commit extras. That salt affects lots of tests.

tests: replace an X+X revset with something meaningful The change in aab2afc2bd82 seem to have been make too hastily.

tests: update section headers and comments in obslog-related tests

tests: use common obshistory_setup.sh script instead of manual setup This is now consistent with other test-evolve-obshistory*.t tests.

tests: remove "second diverged revision" from a phase divergence test It was basically testing the same thing twice with a different comment.

obshistory: omit duplicate notes of obsolescence operations When folding changesets, it's possible to provide a --note flag to e.g. fold or prune, which will store a note for that operation. But since there would be more than one marker for the fold operation and the provided note would be stored in all of them, _markersnotes() previously would return the same text message many times. We don't really want to see the same thing over and over, so let's make notes a set(). This will also make unrelated operations with the same note show the message only once, which is probably an improvement too.

changelog: adjust The changes only apply to newly created sqlite database. There are code pathes that drop existing ones, but that's a more complicated condition.

changelog: mention the sqlite fix

evolve: extract function for relocating divergent commit There was some duplication here, so let's extract to a function. I'm going to add a little bit more shared code next.

caches: preserve permissions of top-level .hg When using hg on a shared filesystem on UNIX, cache files normally inherit the permissions of the .hg directory. This is most commonly used to ensure everything is writable by all users. The sqlite3 cache files don't have a way to directly set the permission, so check if a special mode is necessary and try to apply them to newly created database files.

changelog: mention obsorigin template keyword

templatekw: add obsorigin keyword

obshistory: add predecessorsandmarkers()

obshistory: add a function to group predsets by fold-id

obshistory: add _predecessorsverb() similar to _successorsetverb()

obshistory: add a docstring to patchavailable()

tests: avoid referring to commits by hash in a few places This patch means that the next patch can be written without manual editing of the test cases.

tests: add more verification of output to divergence tests It's not obvious what the obsmarkers should look like in these tests and adding them highlights a bug that happens when commit we create to resolve the divergence doesn't get created because it would be empty. I also added a few graphlog calls to check that the divergence-resolution commit is in the right place (has right parentage).

tests: add debugobsolete calls before many obslog calls As requested by Pierre-Yves. The main reason is to catch superfluous creation of obsmarkers, as well as any future changes to how the obsmarkers get created.

stablerangecache: sanity check subranges Try to detect invalid conditions on insert as would result in deep recursions and final aborts much later. This has been observed on two different machines and the check makes it hopefully possible to find the origin of the problem.

obslog: use unfiltered repo for --no-graph implementation too repo.unfiltered() is used for graph version of obslog, because if we want to filter non-local changesets, we only have to provide --filternonlocal flag on the command line and _nodesandmarkers() will handle filtering. But that means for obslog without -f to work correctly we need to provide unfiltered repo to that function. With this patch both implementations of obslog (--graph and --no-graph) see the same repo.

tests: don't assume that echo will unescape \\

tests: avoid non-portable sed command sed -i is a GNU extension and the i command without newline as well. Just create the file from scratch with printf to side-step the issue.