Windows dev environment: 56 tests fail on POSIX assumptions — add platform markers and Windows-equivalent invariant checks

[hellices]

Problem

A full local test run on Windows fails 56 tests that all pass in CI (Linux). Verified identical on main — none are product regressions. They are test-environment assumptions, measured and classified:

Class	Representative tests	Mechanism (measured)
POSIX permission bits	test_save_agent_config_preserves_restrictive_file_mode	NTFS has no rwx bits: os.chmod(0o600) only toggles the read-only flag and stat.S_IMODE reads 0o666 for any writable file, so == 0o600 can never pass on Windows
Directory fsync	tests/core/test_audit.py (test_append_fsyncs_before_returning, fail-closed tests)	the audit trail expects file + directory fsync (>= 2 fsync calls); Windows cannot os.open a directory for fsync, so only one fsync is observed
Symlink privileges	tests/core/test_transfer.py::TestPackExtract::test_pack_dereferences_symlink_source	WinError 1314: creating symlinks requires admin rights or Developer Mode; Path.symlink_to() fails outright on a normal account
POSIX path semantics	TestValidateSpec (/tmp/...), tilde-user expansion, Downloads assumptions	/tmp does not exist on Windows (local directory does not exist: \tmp); ~user expansion differs
tty/subprocess semantics	tests/ui/test_node_shell.py, shell-fallback tests in test_app.py	kubectl attach/exec flows assume POSIX terminal suspend/restore

33 other tests already carry platform skip markers; these 56 are the remainder.

Why it matters

• Local development on Windows cannot use pytest / --lf output as a signal: 56 permanent reds bury real regressions.
• Contributors on Windows may misattribute these failures to their changes (this cost an investigation cycle in practice).

Proposal

Two-track, because a blanket skip would weaken security-invariant coverage:

1. Environment-limitation tests → platform markers. Symlink creation, /tmp paths, ~user expansion, tty attach flows: pytest.mark.skipif(sys.platform == "win32", reason=...) (or a shared posix_only marker in conftest.py). These test POSIX behaviors korvid does not promise on Windows.
2. Security-invariant tests → Windows-equivalent assertions, not skips. The audit fail-closed contract (design doc: if the audit entry cannot be written, the write is blocked) and the config file-mode restriction are invariants that must hold on Windows too:
   • audit: assert the file fsync happens and that a failed append still blocks the write on Windows (directory-fsync expectation becomes POSIX-conditional);
   • config mode: assert the POSIX bits on POSIX; on Windows assert what NTFS can express (e.g. the file is not world-readable via ACLs, or at minimum document the limitation explicitly in the test).

Acceptance criteria

• uv run pytest on a clean Windows checkout reports 0 unexpected failures (only pass/skip).
• No security-invariant test is skipped on Windows without a Windows-equivalent assertion or an explicit documented limitation.
• A shared marker/helper exists so future POSIX-assuming tests opt in explicitly instead of failing by surprise.
• CI (Linux) coverage is unchanged.

[hellices]

Implementation design (issue body is the approved spec):

Considered approaches

1. Recommended — shared POSIX marker + Windows full-test CI + OS-equivalent invariant assertions. Makes the failure set observable on every PR, removes only environment assumptions, and keeps Linux coverage unchanged.
2. Inline skipif(win32) on every failure. Fast but duplicates policy/reasons and future tests regress silently — rejected.
3. Rewrite production paths so every POSIX test behaves identically on Windows. Scope creep and often impossible (chmod bits, directory fsync, symlink privilege, terminal suspension) — rejected.

Test infrastructure

• Add tests/platforms.py with WINDOWS, POSIX, and a shared posix_only(reason) marker factory. Reasons must name the unavailable capability (permission bits, privileged symlink creation, POSIX expanduser/path, terminal suspend), not merely say “Windows”.
• Add required windows-test CI job on windows-latest, Python 3.12, uv sync --locked --dev --all-extras, then the full uv run pytest -q. Linux matrix/coverage remains byte-for-byte unchanged.
• First commit only the Windows job/helper and open a draft PR: the failing job is the RED inventory. Use its exact failing node IDs instead of guessing from /tmp string searches (many /tmp values are remote pod paths and are valid on Windows).

Classification/fixes

• Skip with shared marker: tests whose subject is an OS capability korvid does not promise on Windows: creating symlinks without Developer Mode/admin, POSIX ~user lookup, local hard-coded POSIX paths, and real terminal suspend/attach semantics. Mocked remote Linux paths and pure argv builders remain tested on Windows.
• Audit durability (never blanket-skip): Windows must still prove the log file fsync occurs before append returns and any file-fsync failure propagates/fails closed. POSIX additionally proves directory fsync / containing-parent durability. Tests branch only the expected sync count/capability; production _sync_dir already documents NTFS journal semantics.
• Config/token-store atomicity (never blanket-skip): retain fsync-before-replace, interrupted-write preservation, restrictive chmod request, and atomic replacement assertions on Windows. POSIX alone asserts S_IMODE == 0o600; Windows test explicitly documents that Python/NTFS cannot represent POSIX mode bits and therefore asserts the security behavior Windows can express (atomic write, fsync, managed chmod call) rather than a false ACL claim.
• UI terminal flows: skip only tests that rely on Textual suspend/real POSIX tty semantics; keep command construction, approval, audit, cleanup and mocked subprocess outcomes cross-platform.

Documentation and acceptance

• Add docs/windows.md documenting supported dev setup, expected skips, symlink/Developer Mode limitation, and the POSIX-mode limitation; link from README contributor docs.
• Windows job must finish with 0 failures. Review every skip: security-invariant skips require an adjacent Windows-equivalent test or documented limitation.
• Linux full gate/coverage must remain green and unchanged.