Persistence for Graph runs Provides snapshotting, timing, and history utilities without changing core Graph behavior.

[shaheerzaman]

**Summary

Adds an in-memory persistence module for GraphAI, modeled after pydantic-ai. Provides snapshotting, timing, and history utilities without changing core Graph behavior.**

Key Changes

graphai/persistence/in_memory.py: Simple and full persistence implementations with:

• Node/end snapshots: NodeSnapshot, EndSnapshot
• Status lifecycle: created → pending → running → success|error
• record_run(snapshot_id): async context manager to track timing/status
• load_next(): promotes first created node snapshot to pending
• dump_json()/load_json() (Full): JSON history round-trip (excludes node objects)
• rebind_nodes(name_lookup) (Full): hydrate nodes from names
• deep_copy option (Full) for state/result payloads
• peek_last_id() convenience
• graphai/persistence/init.py: Exposes public API
• Typing/compat:
• NodeSnapshot.node is Any | None to avoid Pydantic forward-ref issues
• BaseStatePersistence.record_run returns AsyncContextManager[None] (mypy clean)

Tests

• tests/unit/test_persistence_in_memory.py:
• Validates SimpleStatePersistence happy/error paths, idempotency, load_next, peek_last_id
• Validates FullStatePersistence deep-copy behavior, history, JSON round-trip, rebind_nodes, error cases
• All tests pass locally

Notes

• Non-breaking: no modifications to graphai.Graph
• Ready for optional follow-up to wire persistence hooks into Graph execution if desired