<!-- agent -->
`Reference.__str__` returns the short name, so `refs/heads/main` becomes
`main` and `refs/remotes/origin/main` becomes `origin/main`. The revision
APIs turned reference objects into revisions with `str(ref)` or
`Git._check_operand(ref, "revision")`, and Git's revision lookup
(gitrevisions(7)) tries `refs/tags/<name>` before `refs/heads/<name>` and
`refs/remotes/<name>`, only warning about the ambiguity on stderr. A tag
that shares a branch's short name therefore took the branch's place.
Whoever controls a remote can supply such tags: `git fetch` auto-follows
tags that point into fetched history, and a clone brings in every tag.

With a branch `main` and fetched tags `main` and `origin/main` on another
commit, `repo.heads.main.commit` still reported the branch, but these APIs
used the tag's commit on both backends:

- `Repo.commit()`, `Repo.tree()`, `Object.new()`, `Repo.iter_commits()`
- `Repo.merge_base()`, `Repo.is_ancestor()`, `Diffable.diff()`
- `Repo.blame()`, `Repo.blame_incremental()`, `Repo.archive()`
- `IndexFile.from_tree()`, `IndexFile.merge_tree()`, `IndexFile.reset()`
- `HEAD.reset()`, `Repo.create_head()`, `Repo.create_tag()`

In the documented merge workflow, `merge_base(new_branch, original_branch)`
followed by `index.merge_tree(original_branch, base=...)`, the merge commit
received the tag's tree while its parents named the branch commits.
`repo.head.reset(repo.heads.main, working_tree=True)` and
`repo.index.reset(ref, head=True)` moved the current branch to the tag's
commit and wrote its files into the working tree.

Add `Git._check_revision()`. For a `SymbolicReference` it reads the exact
reference through `dereference_recursive()`, the `show-ref`/Gix read behind
`ref.object`, and passes on the object ID, which is how
`SymbolicReference.set_object()` already resolves reference arguments.
Strings keep Git's name guessing, as the changelog already documents.
`merge_base()` resolves its top-level arguments before
`_unpack_args()` flattens them. A reference that does not exist now raises
`ValueError`, as `ref.object` does, instead of resolving to whatever Git
guesses for its short name. On the CLI backend each reference argument
costs one or two extra short Git calls; the Gix backend reads references
natively, and `doc/gix-backend.md` is unchanged.

Validation: the new `shadowed` case in `tests/fixtures/exact-refs.sh` backs
`test_reference_arguments_do_not_resolve_same_named_tags` in
`test/test_refs.py`, which checks the APIs above for a branch and a
remote-tracking branch. It fails before the change and passes after it with
`GIT_PYTHON_BACKEND=cli` and `GIT_PYTHON_BACKEND=gix` for SHA-1/SHA-256 and
files/reftable (Git 2.52.0, GixPython 0.3.0, Python 3.11, macOS). Full
suites with the change: CLI 1,915 passed and Gix 2,018 passed. Each run
also had 116 failures, and the same 116 tests fail on the unpatched tree;
they need history and tags such as `0.1.5`, which this shallow clone lacks.
`ruff check`, `ruff format --check`, ShellCheck, `mypy --python-version=3.11`
and `basedpyright --warnings` pass.

Agent-Model: claude-opus-5-5

<!-- agent: Codex -->

Converting reference arguments to object IDs prevents same-named tags from
replacing branches, but loses the reference identity required by several Git
operations. Reflog walks then return no commits, fork-point lookup fails with
`No such ref`, and the default `upload-archive` server rejects the literal ID.
References inside nested `merge_base()` arguments also still became short names
and could select a shadowing tag.

Keep the exact-reference existence and symbolic-chain checks in
`Git._check_revision()`, then pass the validated full ref name. This preserves
Git's ref-dependent behavior without operation-specific conversion switches.
Strings retain Git's revision interpretation and commit objects retain their
IDs. Normalize `os.PathLike` operands through `os.fsdecode()` before validation
so their filesystem paths, including byte paths, are used instead of their
display strings. Flatten nested list/tuple merge-base arguments while preserving
reference objects until each operand is validated.

Extend `tests/fixtures/exact-refs.sh` with an upstream advance and rewind plus
a topic based on its old tip. Regressions compare branch, remote-tracking and
HEAD behavior with Git 2.54.0 across SHA-1/SHA-256 and files/reftable repositories.
They cover shadowing tags, missing refs, nested operands, option spellings,
remote archives and safe/unsafe `PathLike` values. Update the changelog and
document the retained CLI fallbacks: GixPython 0.3.0 exposes neither the required
reflog-walk/fork-point operations nor an upload-archive client.

Assisted-by: GPT 6.1 Sol
Co-authored-by: GPT 6.1 Sol <codex@openai.com>

<!-- aent -->
SUSE and openSUSE package builders disable network access for reproducible
builds. Allow them to exclude network-using tests with
`pytest -m 'not net'` while retaining local filesystem tests.

Register `net` in `pyproject.toml` and document in `CONTRIBUTING.md`
that it includes loopback sockets and fixture-triggered package downloads,
not just internet access. Mark the daemon-based repository tests, stalled
remote timeout cases, daemon restart test, installation test, and the
already-skipped external clone test. Mark only the daemon-launching
cleanup parameter so the other cleanup cases remain selected.

Existing skips and test bodies remain unchanged. Normalize TOML formatting
and imports in the touched files.

Assisted-by: GPT 6.0 Astra
Co-authored-by: GPT 6.0 Astra <codex@openai.com>

…gs-exact-name

fix: read reference arguments by exact name in revision APIs

…s/no-network

no network

When a `patch_env()` body deletes a variable that was initially absent,
cleanup raises `KeyError` and can mask an exception from the body.
Use `os.environ.pop(name, None)` to restore absence idempotently while
preserving restoration of existing values.

Add 18 regression cases for absent, empty and populated original values,
keep/replace/delete body mutations and normal/exceptional exits.
The focused CLI-backend tests changed from 2 failed and 16 passed to
18 passed. Ruff lint/format and `git diff --check` passed. The full suite
was not run because its required `GixPython` Windows wheel was unavailable.

Agent-Model: unknown

…v-deleted-variable

fix(util): tolerate deleted temporary environment variables

<!-- agent -->
Restore Windows CI coverage for `PathLike` revision arguments. The test
introduced in `6de90f1c` uses `Path("refs/heads/main")`, which emits
backslashes on Windows. Those are invalid Git reference names, so resolving
`SymbolicReference` fails during ref-name validation in all four
files/reftable and SHA-1/SHA-256 variants. The same failures were already
present in PR gitpython-developers#2336, before the unrelated `patch_env()` change in PR gitpython-developers#2340.

Use `PurePosixPath` for the test's Git reference names, which use forward
slashes on every platform. It still implements `os.PathLike`, preserving
coverage through `SymbolicReference` and direct and nested `merge_base()`
arguments alongside `PathLikeMock`. Keep native `Path` for filesystem
operations elsewhere in the module and preserve production validation.

Assisted-by: GPT 6.1 Sol
Co-authored-by: GPT 6.1 Sol <codex@openai.com>

Took a quick look, pretty zoomed out, but think it's
a desirable change as finally the test-suite is independent
of the input repository or state.

<!-- agent -->
Make PR and post-merge test results independent of checkout depth, refs,
and history. The shared test repository and dependency clone sources
previously came from the source checkout, requiring historical objects
and a bootstrap script even though setup already used `gix.testtools`.

Generate small repositories from scratch through `gix.testtools`, with
named root, release, merge, conflict, rename, date, and submodule states.
Adapt legacy assertions to those scenarios while retaining merge,
serialization, traversal, index, remote, and nested-submodule coverage.
Make dirty-submodule removal checks explicit for staged, unstaged, and
untracked data, including dry runs.

Remove the source-history environment override, dependency factory,
checkout lock cleanup, and bootstrap script. The main CI suites now use
shallow checkouts and remove source `.git` metadata before pytest. Add a
regression that initializes the shared fixture from a source copy without
Git metadata, and document the isolation requirement. Hash raw parser
samples in cache arguments to avoid archive filename limits on cold runs.

Assisted-by: GPT 6.0 Astra
Co-authored-by: GPT 6.0 Astra <codex@openai.com>

<!-- agent -->
GitDB tests previously opened the source checkout's `.git` directory or
`GITDB_TEST_GIT_REPO_BASE`, making their input depend on clone history.
Give each test class a disposable repository generated by a fixture
script, with packed history and additional loose objects. Isolate Git
configuration and identity during setup and register class cleanup.

Run the script with Bash so the standalone suite retains Python 3.8
support without adding a GixPython dependency. Remove the history override
and full-depth checkout from dependency CI, remove source `.git` before
running the suite, and document the new setup.
Use the actual database size when writing a pack instead of the unrelated
iteration count, which the smaller fixture exposed.

Assisted-by: GPT 6.0 Astra
Co-authored-by: GPT 6.0 Astra <codex@openai.com>

…counts

<!-- agent -->
Address benchmark review feedback on the generated fixtures. Copilot noted
that "performance tests still report their iteration limits of 1,000–10,000
as processed counts" even though the fixture has only 90 packed objects.
Use the existing streaming counters and clamp list-based limits to the
available SHA count so both totals and throughput reflect completed work.

Copilot also requested: "Include the benchmark harness sources in the
fingerprint, and add a harness-only change to the digest regression test."
Hash the Python sources in `test/performance` alongside `VERSION` and `git`.
This prevents comparing different timed workloads that return the same
result. Extend the generated source fixture and document the digest scope.

Assisted-by: GPT 6.0 Astra
Co-authored-by: GPT 6.0 Astra <codex@openai.com>

…s/fix-ci

test: isolate repository fixtures and restore Windows reference coverage