Don't exclude imports in the else branch of TYPE_CHECKING guards

[pylaterreur]

Fixes #323.

With exclude_type_checking_imports=True, the imports in the elif/else branches of an if TYPE_CHECKING: statement were excluded along with the guarded ones, even though those branches run at runtime. Contracts that such imports break were therefore kept. For example, build_graph("filelock", exclude_type_checking_imports=True) has no import from filelock to filelock._read_write or filelock._async_read_write, which filelock/__init__.py imports in such an else branch.

The import parser now only marks the branches guarded by TYPE_CHECKING as type checking only:

• else branches, and elif branches with other conditions, are runtime code.
• elif TYPE_CHECKING: branches are type checking only, as they were with the Python import scanner.
• A nested if TYPE_CHECKING: restores the enclosing state when it ends, rather than resetting it to "not type checking".

Graphs built without exclude_type_checking_imports are unchanged.

On the installed versions of filelock, numpy, aiohttp, propcache, docutils, anyio and packaging (with include_external_packages=True), this adds 14 runtime imports that were missing, such as filelock -> filelock._read_write, numpy._typing._array_like -> numpy._core.multiarray and aiohttp.connector -> ssl. It also removes 2 type checking only ones that were kept: packaging.version -> typing_extensions and packaging.specifiers -> typing_extensions, which are imported under elif TYPE_CHECKING:.

There's no measurable performance change. Building the graph of nine installed packages (1,920 modules, release builds, one thread, with exclude_type_checking_imports=True) takes 3,399M instructions before and 3,401M after, within the run-to-run noise of about 2M. The instructions spent in the import visitor go down by 0.4M.

• Add tests for the change. In general, aim for full test coverage at the Python level. Rust tests are optional.
• Add any appropriate documentation.
• Add a summary of changes to the latest section at the top of CHANGELOG.rst. (If it's not there, add it.)
• Add your name to AUTHORS.rst.
• Run just full-check. (I ran the lint (Python, plus Rust with the pinned 1.97.0 toolchain), the docs build, cargo test, and the Python tests on 3.10 to 3.14. I didn't run them on 3.14t, 3.15 or 3.15t.)

🤖 Generated with Claude Code

https://claude.ai/code/session_01GwUvhmzgfkiZ6vo4uGyesN

[codspeed]

Merging this PR will not alter performance

✅ 26 untouched benchmarks
⏩ 23 skipped benchmarks¹

---

_{Comparing pylaterreur:fix-type-checking-else (dc2c02b) with main (676e490)}

Footnotes

1. 23 benchmarks were skipped, so the baseline results were used instead. If they were deleted from the codebase, click here and archive them to remove them from the performance reports. ↩