feat(sqlserver): add SQLServerSaver LangGraph checkpoint saver

[DevNinja (0xDevNinja)]

Summary

Adds SQLServerSaver, an implementation of langgraph.checkpoint.base.BaseCheckpointSaver backed by a SQL Server / Azure SQL database. This is the LangGraph-persistence integration tracked in #15.

Modeled after langgraph.checkpoint.sqlite.SqliteSaver so behavior is consistent with the other first-party savers, adapted for SQL Server:

• SQLAlchemy + pyodbc transport (sync), same connection-string conventions as SQLServer_VectorStore including Entra ID authentication.
• MERGE INTO ... WHEN MATCHED THEN UPDATE ... WHEN NOT MATCHED THEN INSERT for the put upsert.
• IF NOT EXISTS (...) INSERT for the put_writes insert-or-ignore path used for non-WRITES_IDX_MAP channels.
• TOP N instead of LIMIT N for list(..., limit=...).
• JSON_VALUE(metadata, '$.key') for the metadata filter= argument.
• An internal threading.Lock guards every session so concurrent graph runs in the same process cannot interleave statements on the underlying engine — mirrors SqliteSaver's safety semantics.

Closes #15

Public API

from langchain_sqlserver import SQLServerSaver

saver = SQLServerSaver(
    connection_string="Driver={ODBC Driver 18 for SQL Server};Server=tcp:host,1433;Database=mydb;Uid=user;Pwd=pwd;TrustServerCertificate=yes;",
)
graph = builder.compile(checkpointer=saver)
config = {"configurable": {"thread_id": "user-1"}}
graph.invoke({...}, config)

Implements get_tuple / list / put / put_writes / delete_thread, plus a lazy setup() for table creation. Tables (checkpoints, checkpoint_writes) are configurable via checkpoints_table / writes_table and db_schema.

Files

• langchain_sqlserver/checkpoint.py — new module with SQLServerSaver.
• langchain_sqlserver/__init__.py — export the new class.
• pyproject.toml + uv.lock — add langgraph-checkpoint>=2.0,<4.0 to project dependencies.
• tests/unit_tests/test_checkpoint.py — 8 unit tests covering construction validation, qualified-name building (schema / no-schema / overrides), Entra-ID-vs-uid/pwd routing, the empty-input fast path on put_writes, and SQL-injection guard on the metadata filter. No live DB required.
• tests/unit_tests/test_imports.py — extend the __all__ snapshot.
• tests/integration_tests/test_checkpoint.py — 6 integration tests covering: put + get_tuple round-trip, latest-checkpoint lookup, descending list with limit, before= slicing, put_writes retrievability through get_tuple.pending_writes, and scoped delete_thread. Uses the existing TEST_AZURESQLSERVER_TRUSTED_CONNECTION / TEST_PYODBC_CONNECTION_STRING env vars.

Test plan

• ruff check langchain_sqlserver tests — clean
• mypy langchain_sqlserver tests — clean (14 source files)
• pytest tests/unit_tests/ — 9 passed (8 new + import snapshot), 1 module skipped (pre-existing)
• pytest tests/integration_tests/test_checkpoint.py against a live SQL Server / Azure SQL instance.

Notes

• Connection-string parsing and Entra-ID token-provider helpers mirror those on SQLServer_VectorStore so users get identical auth semantics across vector store, chat history, and checkpoint saver. A follow-up to factor these into a shared internal module is reasonable, but kept out of this PR so it stays additive.
• poetry.lock is intentionally left in tree as-is — uv.lock is the canonical lockfile updated for the new dependency; happy to refresh poetry.lock if maintainers prefer.