stablesort: introduce a small debugstablesortcache command This give a small insight on the cache content.

stablesort: add some documentation for stablesort This should help people to understand what is going on here.

fold: check allowdivergence before folding obsolete changesets (issue5817)

tests: demonstrate that fold doesn't care about allowdivergence config option

stack: remove unnecessary copying of rdependencies rdependencies is not modified in any way in this method, so no need to copy it.

stack: make a deep copy of `dependencies` before modifying its items The algorithm later on in this method uses .remove() to remove individual elements from items in dependencies, which before this patch modified the cached property contents. So for further use that dictionary was in the form of {1: set([])}, i.e. all sets were empty. This deep copy block could be way simpler, but the problem is that sometimes we get lists of _succs() from evolvebits.builddependencies(). Note: this happens only in topic's stack version of builddependencies() and it looks like a suboptimal way to handle multiple successors (see evolve's counterpart function). stack.builddependencies method is removed, it has served its purpose (see the previous patch).

stack: demonstrate that not reusing cached property gives different results Instead of using self._dependencies, which is a cached property, let's get fresh deps and rdeps in stack.revs method using a newly introduced stack.builddependencies method (will be removed in the next patch). This makes it easier to explain why the next patch is correct.

branching: merge with stable

changelog: add missing entry for 9af212b8565a

evolve: test that target is not orig in _solveunstable() (issue6097) `newer` is the result of obsutil.successorssets() and can be [[orig.node()]], in which case later on in this function evolve will try to rebase orig onto orig, which is not correct. So let's just check this particular case. This fix doesn't cover cases when successorssets() result contains orig.node() not at [0][0]. Such cases need tests.

tests: demonstrate an orphan changeset cause "relocate node on top of itself" See issue6097.

obslog: only indent the first chunk and chunks just after newlines (issue6175)

tests: demonstrate too many spaces in olog -p output with word-diff

rewind: add --keep flag that "doesn't modify working directory" The actual logic is more complicated than the flag description, but it's sufficiently similar to other --keep flags in action. Unlike strip (or prune), rewind always needs to modify the working directory to commit new revisions that "revive" old ones [1], see _revive_revision() (and rewriteutil.rewrite()). Because of that we don't prevent rewind from modifying wdir, but instead use hg.updaterepo() to update to the old changeset after the "revival" process is complete. Then we rebuild the dirstate based on the commit that rewind would update to without --keep. Since dirstate.rebuild() doesn't restore status of some files (added, removed, also copies and renames), we rely on cmdutil.revert(). It's a fairly crude solution and needs to be removed when implementing the missing copy tracing between oldctx and newctx (which are related only by obsolescence). [1] IOW this means that --keep doesn't allow rewinding if wdir is dirty (unlike e.g. strip).

tests: separate rewinding of merge commits, temporarily drop an error case The error case will come back in the following commit, because it organically fits there without the need of a separate working clone.