Somehow I ended up with the following in my ~/.gitconfig:
[branch "master"]
remote = origin
merge = master
rebase = true

I assume something went crazy while I was running the git.git tests
some time ago, and that I never noticed until now.

This is not a good configuration, but it shouldn't cause problems. But
it does. Specifically, if you have this in your config, and you
perform the following set of actions:

create a remote
fetch from that remote
create a branch off of the remote master branch called "master"
delete the branch
delete the remote

The remote delete fails with the message "Could not find key
'branch.master.rebase' to delete". This is because it's iterating over
the config entries (including the ones in the global config) and
believes that there is a master branch which must therefore have these
config keys.

https://github.com/libgit2/libgit2/issues/3856

When we create a blame origin, we try to look up the blob that is
to be blamed at a certain revision. When this lookup fails, e.g.
because the file did not exist at that certain revision, we fail
to create the blame origin and return `NULL`. The blame origin
that we have just allocated is thereby free'd with
`origin_decref`.

The `origin_decref` function does not only decrement reference
counts for the blame origin, though, but also for its commit and
blob. When this is done in the error case, we will cause an
uneven reference count for these objects. This may result in
hard-to-debug failures at seemingly unrelated code paths, where
we try to access these objects when they in fact have already
been free'd.

Fix the issue by refactoring `make_origin` such that we only
allocate the object after the only function that may fail so that
we do not have to call `origin_decref` at all. Also fix the
`pass_blame` function, which indirectly calls `make_origin`, to
free the commit when `make_origin` failed.

git only checks ceiling directories when its search ascends to a parent
directory.  A ceiling directory matching the starting directory will not
prevent git from finding a repository in the starting directory or a
parent directory.  libgit2 handled the former case correctly, but
differed from git in the latter case: given a ceiling directory matching
the starting directory, but no repository at the starting directory,
libgit2 would stop the search at that point rather than finding a
repository in a parent directory.

Test case using git command-line tools:

/tmp$ git init x
Initialized empty Git repository in /tmp/x/.git/
/tmp$ cd x/
/tmp/x$ mkdir subdir
/tmp/x$ cd subdir/
/tmp/x/subdir$ GIT_CEILING_DIRECTORIES=/tmp/x git rev-parse --git-dir
fatal: Not a git repository (or any of the parent directories): .git
/tmp/x/subdir$ GIT_CEILING_DIRECTORIES=/tmp/x/subdir git rev-parse --git-dir
/tmp/x/.git

Fix the testsuite to test this case (in one case fixing a test that
depended on the current behavior), and then fix find_repo to handle this
case correctly.

In the process, simplify and document the logic in find_repo():
- Separate the concepts of "currently checking a .git directory" and
  "number of iterations left before going further counts as a search"
  into two separate variables, in_dot_git and min_iterations.
- Move the logic to handle in_dot_git and append /.git to the top of the
  loop.
- Only search ceiling_dirs and find ceiling_offset after running out of
  min_iterations; since ceiling_offset only tracks the longest matching
  ceiling directory, if ceiling_dirs contained both the current
  directory and a parent directory, this change makes find_repo stop the
  search at the parent directory.

The MSYS2 build system automatically compiles all code with position-independent
code. When we manually add the -fPIC flag to the compiler flags, MSYS2 will
loudly complain about PIC being the default and thus not required.

Fix the annoyance by stripping -fPIC in MSYS2 enviroments like it is already
done for MinGW.

Indicate that if you make changes to libgit2 that you must distribute
the source _to libgit2_, not the source _of your program_.

The old pthread-file did re-implement the pthreads API with exact symbol
matching. As the thread-abstraction has now been split up between Unix- and
Windows-specific files within the `git_` namespace to avoid symbol-clashes
between libgit2 and pthreads, the rewritten wrappers have nothing to do with
pthreads anymore.

Rename the Windows-specific pthread-files to honor this change.

The function pthread_num_processors_np is currently unused and superseded by the
function `git_online_cpus`. Remove the function.

When no index file exists and a baseline is not explicitly provided, use
an empty baseline instead of trying to load `HEAD`.

`git_checkout_head` is sadly misunderstood as something that can
switch branches.  It cannot.  Update the documentation to reflect this.

The thread local storage is used to hold some global state that
is dynamically allocated and should be freed upon exit. On
Windows, we clean up the C run-time right after execution of
registered shutdown callbacks and before cleaning up the TLS.

When we clean up the CRT, we also cause it to analyze for memory
leaks. As we did not free the TLS yet this will lead to false
positives.

Fix the issue by first freeing the TLS and cleaning up the CRT
only afterwards.

When removing an entry from the index by its position, we first
retrieve the position from the index's entries and then try to
remove the retrieved value from the index map with
`DELETE_IN_MAP`. When `index_remove_entry` returns `NULL` we try
to feed it into the `DELETE_IN_MAP` macro, which will
unconditionally call `idxentry_hash` and then happily dereference
the `NULL` entry pointer.

Fix the issue by not passing a `NULL` entry into `DELETE_IN_MAP`.

When we receive a packet of exactly four bytes encoding its
length as those four bytes it can be treated as an empty line.
While it is not really specified how those empty lines should be
treated, we currently ignore them and do not return an error when
trying to parse it but simply advance the data pointer.

Callers invoking `git_pkt_parse_line` are currently not prepared
to handle this case as they do not explicitly check this case.
While they could always reset the passed out-pointer to `NULL`
before calling `git_pkt_parse_line` and determine if the pointer
has been set afterwards, it makes more sense to update
`git_pkt_parse_line` to set the out-pointer to `NULL` itself when
it encounters such an empty packet. Like this it is guaranteed
that there will be no invalid memory references to free'd
pointers.

As such, the issue has been fixed such that `git_pkt_parse_line`
always sets the packet out pointer to `NULL` when an empty packet
has been received and callers check for this condition, skipping
such packets.

When adding a new entry to an existing index via `git_index_read_index`,
be sure to remove the tree cache entry for that new path.  This will
mark all parent trees as dirty.

Read a tree into an index, write the index, then re-open the index and
ensure that we are treesame to the original.

Update the flags to reset the path_len (to emulate `index_insert`)

Treat index entries with different modes as different, which they
are, at least for the purposes of up-to-date calculations.

Clear any error state upon each iteration.  If one of the iterations
ends (with an error of `GIT_ITEROVER`) we need to reset that error to 0,
lest we stop the whole process prematurely.

Read a tree into an index using `git_index_read_index` (by reading
a tree into a new index, then reading that index into the current
index), then write the index back out, ensuring that our new index
is treesame to the tree that we read.

typo fix

Don't generate conflicts when checking out a modified submodule and the
submodule is dirty or modified in the workdir.

Haiku and Hurd both pass extra bits in struct stat::st_mode.