PR 2 of N in the Windows-CI green-up roadmap (after #711 fixed the
path-separator class). Skips 9 tests across 5 files that exercise
features which either don't exist on Windows or take a deliberately
different code path that's already covered by an existing
`test_windows_skips_kill`-style sibling.

Pattern: single uniform `@pytest.mark.skipif(sys.platform == "win32",
reason="…")` decorator. No assertion logic touched, no production
source modified.

POSIX file mode (4 tests): Windows ignores POSIX permission bits
entirely, so `stat.S_IMODE(...)` always returns the NTFS-default 0o666 /
0o644 / 0o777 regardless of what the production code requests via
`os.chmod` or `os.fchmod`. The intent under test (atomic_write defaults
to 0o600 for state files, custom mode is honoured, install lands at
0o644) is meaningful only on POSIX.

- `test_context_atomic.py::test_atomic_write_explicit_mode_0o600`
- `test_context_atomic.py::test_atomic_write_respects_custom_mode`
- `test_context_install.py::test_install_files_written_with_default_mode`
- `test_context_settings.py::TestClaudeSettingsAtomicWrite::test_mode_is_0o600`

`fcntl` (1 test): the Python module doesn't exist on Windows and the
test imports it inside the function body, so on Windows the test
crashes with `ModuleNotFoundError: No module named 'fcntl'` at collect
time. The fix it pins (LOCK_SH + LOCK_NB coexistence on the same file)
is meaningful only on POSIX.

- `test_server_sigterm.py::test_legacy_lock_sh_allows_multiple_holders`

Signal / SIGKILL escalation (4 tests): the production path in
`upgrade_cmd` already takes a Windows-specific branch ("Detected
Windows; skipping process termination") that is fully covered by
`test_windows_skips_kill` (which we leave running on Windows). The
SIGTERM/SIGKILL escalation tests assert on POSIX-only output strings
and signal sequences and fail otherwise. Skipping them does NOT lose
Windows coverage of the upgrade flow — `test_windows_skips_kill` keeps
running.

- `test_upgrade_cmd.py::test_no_running_server_just_reinstalls`
- `test_upgrade_cmd.py::test_running_server_sigterm_path`
- `test_upgrade_cmd.py::test_running_server_escalates_to_sigkill`
- `test_upgrade_cmd.py::test_pid_file_unlink_skipped_if_respawned`

Out of scope (other Windows CI failure classes — separate PRs):

- `test_uninstall_cmd.py` failures: NOT POSIX-only mode bits — they
  surface NTFS file-locking semantics (.db / .db-shm / .db-wal aren't
  deleted while open, lock detection, inventory text differences).
  Multi-factor; deferred to PR 4/5.
- HOME monkeypatch reach (`Path.home()` vs `USERPROFILE`) — issue #644
- Path canonicalization in indexing — issue #647
- Symlink + git-status detection (flaky class)
- `'home'` fs alias / tilde expansion — issue #646
- `config_signature` NTFS mtime resolution — issue #645
- Wizard auto-detect Claude/Codex memory dirs — issue #316

Verification:

- `uv run pytest` on the 5 touched files (macOS): 91 passed.
- `uv run ruff check + format --check` clean.
- Windows CI before/after counts to be appended to the PR body once the
  branch run finishes; expected reduction is ~9 (all 9 targeted tests
  show up as SKIPPED on Windows, FAIL count drops accordingly).

Refs #637 (file-mode class umbrella; this PR doesn't touch the
XDG_RUNTIME_DIR tests #637 was originally written for, so leaving it
open).

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <[email protected]>