fix: serialize graph persistence and preserve dirty state for mutations occurring during a save

[DivyamTalwar]

@haolpku

Summary

NetworkX graph persistence now captures immutable revisions, serializes writes per store, preserves dirty state when a mutation happens during a save, and uses a unique temporary file for each atomic replacement.

Root cause

The old worker iterated the live graph and unconditionally set _dirty = False after the write. A mutation could land after the worker captured the old graph but before that reset, leaving newer in-memory data absent from disk while the next persist returned early. Concurrent writes also shared one PID-derived temporary filename.

Change

• Protect graph capture and mutation bookkeeping with an in-process state lock.
• Serialize persist calls with an async lock.
• Track a mutation revision and clear dirty state only when the saved revision is still current.
• Write detached documents through unique same-directory temporary files and clean them up on failure.
• Keep profile/source/lineage reconciliation removals under the same state lock.

Regression coverage

Barrier/event-driven tests cover mutation during save, overlapping saves, unique latest-file results, write failures, retry behavior, and existing graph traversal/reload behavior.

Verification: python -m pytest -q — 197 passing, 5 optional SDK tests skipped.

The patch is intentionally in-process. It does not claim cross-process locking, fsync durability, or a storage-format migration.

Please review the concurrency model and merge if CI is green.

[DivyamTalwar]

@haolpku — Could you please review the deterministic graph persistence-race fix and merge it if CI confirms the full suite? The PR explicitly does not claim cross-process locking or fsync durability.