Fixes #1405

The bug, and one symptom the issue does not mention

SubmoduleCollection.get() promises:

Unlike __getitem__, this returns None if the submodule is not found.

It only catches KeyError. But git_submodule_lookup reports GIT_EEXISTS, not GIT_ENOTFOUND, when a repository exists at the path yet was never registered as a submodule — libgit2's message is submodule 'x' has not been added yet. check_error turns that into AlreadyExistsError, which escapes get().

Because __contains__ is implemented as self.get(name) is not None, the same input makes a containment test raise:

outer = pygit2.init_repository(outer)
pygit2.init_repository(outer/nested)

outer.submodules.get(nested)      # ValueError: submodule nested has not been added yet
nested in outer.submodules        # same ValueError
outer.submodules.get(absent)      # None  (correct)

Reproduced on 1.20.0 from PyPI and on this branch's parent built against libgit2 1.9.7.

The reporter later wondered whether their confusion about nested repositories was the real issue. It was not the whole of it: whatever one thinks libgit2 should report for a nested repository, get() and in are documented to answer "absent" rather than raise, and today they raise.

The change

get() also catches AlreadyExistsError. __getitem__ is untouched, so callers who need to distinguish "a repository is there but unregistered" from "nothing is there" still can — the new test pins that too.

I considered instead translating GIT_EEXISTS to KeyError inside __getitem__, which would arguably make its own docstring more accurate. I did not, because it changes an exception type users may already catch, and it is your call rather than mine. Happy to switch if you prefer it.

Verification

• pytest test/test_submodule.py — 28 passed
• Full suite minus network/ssh — 595 passed, 7 skipped, 2 xfailed, 1 xpassed, 1 failed
• That one failure is pre-existing: test_status_file_unicode_normalization[café.txt] fails identically with my changes stashed on a clean checkout, which is the usual macOS filesystem normalisation issue
• Mutation check: reverting only pygit2/submodules.py fails exactly the new test, test_lookup_nested_repo_that_is_not_a_submodule
• ruff check and ruff format --check clean on both files
• Built locally with LIBGIT2=$(brew --prefix libgit2) pip install -e .

Per CONTRIBUTING.md, the commit carries Assisted-by: Claude Code (Claude Opus 5). I can explain the change and stand behind it.