src/specify_cli/shared_infra.py:265

• missing_ok=True suppresses only FileNotFoundError; another cleanup failure (for example, PermissionError) will still replace the exception raised by the write/replace operation. Since this finally block is intended not to mask that failure, suppress OSError around the best-effort cleanup.

        temp_path.unlink(missing_ok=True)

src/specify_cli/integrations/manifest.py:442

• missing_ok=True handles a vanished temporary file, but it still lets every other OSError from cleanup mask the exception that entered this finally block. Make removal best-effort by suppressing OSError so callers receive the original manifest-save failure.

            temp_path.unlink(missing_ok=True)

src/specify_cli/_utils.py:183

• If unlinking fails for any reason other than the file being absent, this cleanup still replaces the original exception despite the PR's stated goal. Catch OSError from this best-effort removal before re-raising the original failure.

            if temp_path:
                temp_path.unlink(missing_ok=True)

tests/extensions/test_update_agent_context_python_parity.py:393

• This fixture symlinks a directory, but Path.rglob() does not recurse through directory symlinks by default. The test therefore passes even without _resolved_rel() and does not cover the newly added containment check; symlink the discovered plan.md file itself to the outside file instead.

        specs = repo / "specs"
        specs.mkdir(parents=True, exist_ok=True)
        (specs / "linked").symlink_to(outside.parent, target_is_directory=True)
        # Sanity: the plan really is reachable through the symlink.
        assert (specs / "linked" / "001-x" / "plan.md").is_file()

tests/extensions/test_update_agent_context_python_parity.py:556

• The removed Python code already used recursive rglob("plan.md"), so the claim that this PR fixes a one-level Python scan is inaccurate. Reword this as an ongoing parity check rather than attributing behavior to the old implementation.

    With no feature.json, discovery falls back to scanning under specs/. A plan
    at specs/<scope>/<feature>/plan.md sits more than one level deep; the old
    Python one-level glob missed it while PowerShell already recurses (#3024).

extensions/agent-context/scripts/python/update_agent_context.py:186

• This resolved-path filtering and its parity tests are outside the PR's stated temporary-file-cleanup scope. Please either move the agent-context behavior change to a dedicated PR or update the PR title/description to explain and justify the additional behavior.

        def _resolved_rel(p: Path) -> Path | None:
            # Resolve symlinks before checking containment: relative_to() is
            # lexical and would otherwise accept a plan reached through a specs/
            # symlink that points outside the project, emitting an
            # in-project-looking path for an out-of-project file (or picking it
            # as "most recent").
            try:
                return p.resolve().relative_to(root)
            except (OSError, ValueError):

tests/extensions/test_update_agent_context_python_parity.py:355

• This describes a one-level Python glob, but the removed implementation already used (root / "specs").rglob("plan.md"). Please avoid presenting this as a regression fixed by this PR; the implementation change here adds resolved-path containment filtering instead.

This issue also appears on line 554 of the same file.

    A plan created under specs/<scope>/<feature>/plan.md (as produced via
    SPECIFY_FEATURE_DIRECTORY) is more than one level below specs/. The old
    Python port used a one-level specs/*/plan.md glob and missed it, while the
    bash/PowerShell twins recurse (#3024). This locks in the parity.

src/specify_cli/commands/bundle/init.py:883

• The PR description limits the change to temporary-file cleanup in three files, but this adds a new 50 MiB behavior limit to remote bundle downloads. Either split this functional/security change into its own PR with focused bundle-level coverage, or update the title and description so reviewers and release notes capture it.

            raw = read_response_limited(resp, max_bytes=MAX_DOWNLOAD_BYTES, error_type=BundlerError)