fix: add Windows-specific LMDB cleanup handling

[jaelliot]

• Add explicit env.sync(force=True) before close to flush writes
• Add 100ms delay on Windows after close when clear=True
• Ensures file locks are released before directory cleanup
• Fixes test_mailboxing PermissionError on Windows CI

Resolves: Pre-existing Windows CI test failure
Related: PR #1143

[jaelliot]

Local Windows Testing Results

I tested the Windows LMDB fix locally and got unexpected results that warrant discussion before merging.

Test Environment

• OS: Windows 10 via WSL2 (Ubuntu 22.04)
• Python: 3.14.2
• Test Method: Ran pytest from Windows PowerShell against WSL filesystem (\\wsl$\Ubuntu-22.04\...)

Results

All LMDB-related tests PASSED ✅:

pytest tests/app/test_storing.py::test_mailboxing -v
# 1 passed in 1.57s

pytest tests/db/test_dbing.py -v  
# 4 passed in 0.89s

pytest tests/app/test_habbing.py -v
# 10 passed in 6.90s

Important Caveat ⚠️

This may not reflect true native Windows behavior because:

1. WSL filesystem layer: Files accessed through \\wsl$\... go through WSL's translation layer
2. Different locking semantics: WSL may handle file locks differently than native Windows NTFS
3. CI uses native Windows: GitHub Actions runs on native Windows filesystem, not WSL

Why This Matters

The original CI failures were:

• PermissionError: [WinError 32] The process cannot access the file because it is being used by another process
• On native Windows filesystem in GitHub Actions

My local tests may be passing because WSL provides different file locking behavior.

Recommendation

We should push this fix and verify it works in the actual GitHub Actions CI environment (native Windows) before considering it resolved. The local WSL tests are encouraging but not definitive.

Code Changes Summary

Added to LMDBer.close():

• 10ms delay after env.close() on Windows
• Exponential backoff retry (50ms → 25s total)
• Explicit garbage collection between retries
• Better error classification (only retries on WinError 32/errno 13)

[jaelliot]

Converting to draft - shifting focus per team guidance

After multiple iterations attempting to fix Windows CI failures (LMDB cleanup, port binding issues), tests are still failing. Per team discussion, switching focus to Sphinx documentation work which has no platform-specific issues.

What was attempted:

Windows LMDB retry logic with exponential backoff
Windows port mapping (5000-5999 → 8000+)
Python version pinning to 3.14.2
libsodium CI installation
Current status:

Tests still fail with similar errors
Local Windows testing via WSL was not representative of CI environment
Will revisit if maintainers have guidance, but not blocking other work
Moving to Sphinx documentation PRs now.