fix(storage): retry os.replace in atomic_write on Windows NTFS lock c…

[zhiwu215]

Description

Fixes transient PermissionError: [WinError 5] Access is denied during atomic_write on native Windows (NTFS) environments.

On Windows, os.replace (MoveFileExW with MOVEFILE_REPLACE_EXISTING) fails with PermissionError if the target file (such as kv_store_doc_status.json) is momentarily held open by another async worker, an IDE file watcher, or Windows Defender. Previously, atomic_write lacked a retry mechanism, causing a single millisecond-level lock collision during concurrent document ingestion to crash the entire process with Pipeline halted on internal storage error.

This PR introduces a lightweight exponential backoff retry specifically for PermissionError on win32, while preserving zero-overhead single-pass behavior on POSIX platforms.

Related Issues

Fixed #3879

Changes Made

• Added _replace_file helper in lightrag/file_atomic.py with exponential backoff retry (up to 10 attempts, starting at 50ms) on PermissionError.
• Scoped multi-attempt retry specifically to sys.platform == "win32" (max_attempts = 10), preserving single-attempt zero-overhead behavior on POSIX platforms (max_attempts = 1).
• Replaced direct os.replace call in atomic_write with _replace_file.
• Maintained pure standard library dependencies (stdlib only).

Checklist

• Changes tested locally
• Code reviewed
• Documentation updated (if necessary)
• Unit tests added (if applicable)

Additional Notes

Local Testing & Verification

• Environment: Native Windows (NTFS), Python 3.12, LightRAG.
• Test Case: Batch ingestion and graph building for a 143-chapter novel (500k+ words).
• Result: Prior to this change, the pipeline consistently halted around document 80~90 due to concurrent doc-status flush lock contention. With this retry logic applied, all 143 chapters completed smoothly with zero pipeline aborts.