inhibit: remove unused bookmark operation wrapping Before this patch, inhibit was wrapping bookmarks.write and bookmarks.recordchange. Since all the usage of bookmarks.write are not replaced by bookmarks.recordchange all bookmarks operation happen in a transaction. Inhibit already wraps the transaction mechanism to make sure that no revision can end up being obsolete and visible. This makes the wrapping of bookmarks.write superfluous. Wrapping bookmarks.recordchange was wrong in the first place and redundant with wrapping transactions.

inhibit: update test with new behavior from core The recent changes in core (to make sure that all bookmarks write are done at the end of a transaction) changed the behavior of inhibit. When one rebases a set of hidden revs somewhere, the hidden revs become visible at the end of the rebase. Both the previous behavior and new behavior seems acceptable but the tests have to be changed subsequently. This patch modifies one of the test for inhibit when the case mentioned above happens. We prune the revs inhibited by the rebase to make the change minimal.

inhibit: remove unused import The error module from mercurial was imported but unused in the inhibit code.

inhibit: improve handling of error cases for bookmark -D Before this patch bookmark -D was crashing in some cases like: hg book -Dmaster-fix. This was because -m is a valid option for hg bookmark. This patch disallows using -m with -D (moving and pruning does not make much sense) and throws an error instead of crashing if the command is malformed.

directaccess: fix case of shortened hash containing only digits For directaccess, there are four cases for what looks like short hashes made of digits only: 1 | the hash is a revision number and not a short hash for another revision: we don't change the visibility 2 | the hash is a revision number and a short hash for another revision: we don't change the visibility 3 | the hash is not a revision number and not a short hash for another revision: we don't change the visibility 4 | the hash is not a revision number but is a short hash for another revision: we make it visible Before the patch we were not lifting visibility in case number 4. This patch fixes the issue.

evolve: evolve --divergent failed when merge had nothing to do (issue4950) Before this patch, if the divergence resolution lead to an empty commit, the transaction would be rolled back and the divergence not resolved. This patch allows empty commits in that case. It is legitimate to end up with an empty commit in that case, for example when divergent commits differ over their commit messages.

inhibit: use dirty revset trick to speedup the inhibit search space. See documentation for details. I'm a bit sad about this changesets.

inhibit: use 'repo.revs' instead of 'repo.set' This skip the creation of a changectx object.

_inhibitmarkers: turn the result from generator into a list A generator object is always True, we have to turn it into a list in order to be able to use it in the tests. Before this patch we were rewriting the 'obsinhibit' file all the time, this had a performance impact especially on repository where transaction are expensive (hooks, cache invalidation, etc). Writing this file less often have a minor effect on test where a stripped changeset does not get removed from the obsinhibit file is still inhibited when unbundled back. I think this is fine and I'm moving forward.

split: drop the experimental flag I've been using it without issue for some time. The only blocker in my mind was the lack of commit message preservation and this is now fixed.