Summary

• test_aborts_on_mtime_change is the last residual Windows CI failure from Windows: config_signature unchanged after write — NTFS mtime resolution / cache #645. The test simulates a concurrent writer by patching _read_with_mtime to write to the target between the read and the production code's mtime re-check. On Windows the two write_text calls can land inside the same system-clock tick, so st_mtime_ns reports the same value and the abort branch never fires.
• Force a distinguishable mtime via os.utime(..., ns=(atime, mtime + 1ms)) after the simulated write so the test deterministically exercises the production guard regardless of OS timer granularity. A real-world second writer (separate mm context sync run, user editing in editor) is naturally well above the threshold — this only covers the test-induced race.
• Test-only change, no production impact.

Why mtime collides on Windows

NTFS native FILETIME resolution is 100ns, but WriteFile-induced metadata updates use the system clock, which advances at ~15.6ms on typical Windows hosts (default timer interrupt). Two writes within that tick get the same mtime even though the second strictly happened after the first. The fix bumps mtime by 1ms (1_000_000 ns) — well above the synthetic resolution, well below any real-world inter-process gap.

Status of the broader #645 cluster

The issue body listed 17 test_web_hot_reload.py failures. The latest Windows CI run on main (run 25264999636, 2026-05-02) reports 0 hot-reload failures — those have been cleared by previous Windows CI work (PRs #711, #713, #714, #716, #732). Only this single test_context_settings.py::test_aborts_on_mtime_change failure remained, and this PR closes it.

Test plan

• uv run pytest packages/memtomem/tests/test_context_settings.py -v — all 35 tests pass on macOS
• uv run ruff check + ruff format --check — clean
• Windows CI matrix — should drop from 1 failure to 0

References

• Closes Windows: config_signature unchanged after write — NTFS mtime resolution / cache #645
• Sibling Windows fixes: test(windows): set_home helper + sweep ~133 HOME monkeypatch sites (refs #644) #714 (HOME → USERPROFILE), fix(config): support Windows backslash paths in categorize_memory_dir (#316) #716 (Windows: wizard fails to auto-detect Claude/Codex memory dirs (backslash path regex) #316 backslash regex), test: skipif POSIX-only assertions on Windows (mode bits, fcntl, signal) — Windows CI PR 2/N #713 (POSIX skipifs), test: normalize path separators in 7 cross-platform assertions (#643) #711 (path separators), Windows: capture installed_at from FS mtime with us ceiling, not wall clock #732 (installed_at us ceiling), fix(runtime-paths): make Windows runtime dir resolution NTFS-aware (closes #637) #738 (Windows: replace POSIX file-mode XDG_RUNTIME_DIR tests with NTFS-aware equivalents #637 NTFS runtime dir)

🤖 Generated with Claude Code