test: check database ownership against the private temp namespace

requirements-optional.txt

@@ -1,7 +1,7 @@
 # Optional dependencies — install only if you use the corresponding feature.
 # Local OCR for screenshots, scans, labels, and coordinate-grounded text extraction.
-rapidocr==3.10.0
-onnxruntime>=1.31.0,<2
+rapidocr==3.9.2
+onnxruntime>=1.20,<2
 # The app handles their absence gracefully (clear error message on first use).
 #
 # Note: chromadb-client + fastembed moved to requirements.txt — RAG, semantic
@@ -46,7 +46,7 @@ PyMuPDF
 # magika (onnxruntime), already a core dep via fastembed. We avoid the
 # [all]/Azure/audio extras (cloud + heavy). Pinned to a release >30 days old per
 # the dependency-age discussion in issue #485.
-markitdown[docx,pptx,xlsx,xls]==0.1.9
+markitdown[docx,pptx,xlsx,xls]==0.1.6
 
 # Photoshop PSD opening / flattened previews / layer inspection.
 psd-tools

requirements.txt

@@ -5,9 +5,9 @@ python-dotenv
 httpx
 # HTTPS clients using certifi must distrust the compromised eMudhra CA paths.
 certifi>=2026.10.23
-httpcore>=1.0.9,<2.0
-pydantic>=2.14.1
-pydantic-settings>=2.16.1
+httpcore>=1.0,<2.0
+pydantic>=2.13.4
+pydantic-settings>=2.14.1
 SQLAlchemy
 pypdf
 pypdfium2
@@ -25,7 +25,7 @@ numpy
 # ChromaDB service); fastembed runs local ONNX embeddings.
 chromadb-client
 # gRPC vendors independent roots; 1.84.1 removes the incident trust anchors.
-grpcio>=1.85.0
+grpcio>=1.84.1
 fastembed>=0.9.1
 youtube-transcript-api
 yt-dlp
@@ -51,7 +51,7 @@ bcrypt
 # Built-in servers use the v1 low-level Server decorator API. MCP SDK v2 is a
 # breaking rewrite, so keep fresh installs on the maintained v1 line until the
 # servers are migrated together.
-mcp<3
+mcp<2
 pyotp
 qrcode[pil]
 croniter

tests/TESTING_STANDARD.md

@@ -99,8 +99,10 @@ cleanup. Specifically:
   `os.environ[...] = ...` that outlives the test.
 - **Current working directory** - never `chdir` without restoring; never assert
   against cwd-relative paths like `./data`. Use a temp workspace helper instead.
-- **Database** - the root `conftest.py` defaults `DATABASE_URL` to an in-memory
-  SQLite for collection safety. A test that needs a real file-backed DB must opt
+- **Database** - the root `conftest.py` defaults `DATABASE_URL` to a private,
+  worker-owned SQLite file for collection safety and consistent schema across
+  TestClient threads. This fallback is not a per-test data fixture. A test that
+  writes or depends on particular rows must opt
   in explicitly via `tests.helpers.sqlite_db.make_temp_sqlite` and bind its
   `SessionLocal` onto the module under test. Do not rely on a persistent
   on-disk DB existing.

tests/conftest.py

@@ -12,9 +12,17 @@ sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
 # File-backed databases remain fixture-owned. Cleanup restores the caller.
 # Keep the existing generated environment-reference locations stable.
 _database_environment = pytest.MonkeyPatch()
-_database_environment.setenv("DATABASE_URL", "sqlite:///:memory:")
 from tests.helpers.worker_runtime import bootstrap_runtime, configure_runtime
 _runtime_environment = bootstrap_runtime()
+# A file keeps the collection schema consistent across TestClient threads;
+# SingletonThreadPool can evict the populated in-memory main-thread database.
+# One patcher owns this variable, including restoration of an inherited URL.
+_database_environment.setenv(
+    "DATABASE_URL",
+    f"sqlite:///{_runtime_environment.root / 'collection.db'}"
+    if _runtime_environment is not None else "sqlite:///:memory:",
+)
+_collection_database_engine = None
 
 # Pre-import real heavy modules BEFORE any test file's module-level stubs can
 # replace them with MagicMock. Some test files (e.g. test_llm_core_sanitize_*)
@@ -26,6 +34,7 @@ try:
     import sqlalchemy.orm  # noqa: F401
     import core.database  # noqa: F401
     import src.database
+    _collection_database_engine = core.database.engine
 except ImportError:
     pass  # not installed - the stubs below will handle it
 
@@ -100,6 +109,10 @@ def pytest_configure(config):
     only registers marker names; it imports no production module.
     """
     config.add_cleanup(_database_environment.undo)
+    if _collection_database_engine is not None:
+        # Registered after runtime cleanup, so connections close before the
+        # owned file is removed (also required on Windows).
+        config.add_cleanup(_collection_database_engine.dispose)
 
     import pathlib
     from tests._taxonomy import discover_markers

tests/test_database_test_isolation.py

@@ -85,6 +85,9 @@ def test_actual_tests_restore_process_state_and_ignore_inherited_database(tmp_pa
         import os
         import sys
         import pytest
+        import tempfile
+        from pathlib import Path
+        from sqlalchemy.engine import make_url
 
         def snapshot():
             import core
@@ -107,7 +110,11 @@ def test_actual_tests_restore_process_state_and_ignore_inherited_database(tmp_pa
         class StateGuard:
             def pytest_sessionstart(self):
                 self.initial = snapshot()
-                assert self.initial[0] == "sqlite:///:memory:"
+                assert self.initial[0] != inherited_url
+                self.default_path = Path(make_url(self.initial[0]).database)
+                assert self.default_path.is_file()
+                assert self.default_path.name == "collection.db"
+                assert self.default_path.parent == Path(tempfile.gettempdir()).parent
 
             def pytest_collection_finish(self):
                 assert snapshot() == self.initial, "collection changed database bindings"
@@ -118,9 +125,11 @@ def test_actual_tests_restore_process_state_and_ignore_inherited_database(tmp_pa
                 assert snapshot() == self.initial, "test leaked database or module state"
 
         inherited_url = os.environ["DATABASE_URL"]
+        guard = StateGuard()
         result = pytest.main(["-q", "-p", "no:cacheprovider", *sys.argv[1:]],
-                             plugins=[StateGuard()])
+                             plugins=[guard])
         assert os.environ["DATABASE_URL"] == inherited_url
+        assert not guard.default_path.exists(), "collection database outlived its runtime"
         raise SystemExit(result)
     ''')
     result = subprocess.run(

tests/test_worker_runtime.py

@@ -79,3 +79,32 @@ def test_tmp_path_under_private_runtime_fits_unix_socket(tmp_path):
     # witness. A nested pytest-of-<user> basetemp made it 110 bytes under xdist.
     with socket.socket(socket.AF_UNIX) as sock:
         sock.bind(str(tmp_path / "tmux.sock"))
+
+
+def test_collection_database_survives_thread_connection_churn():
+    """TestClient threads must not lose the collection database or its schema.
+
+    SingletonThreadPool's default five-connection limit used to close the main
+    thread's in-memory database after repeated route clients. Each new thread
+    also saw an empty database. Exercise more threads than that limit and then
+    verify both compatibility factories still see all registered tables.
+    """
+    from concurrent.futures import ThreadPoolExecutor
+    from sqlalchemy import inspect
+    import core.database as database
+    import src.database as compatibility
+
+    expected = set(database.Base.metadata.tables)
+
+    def table_names(factory):
+        with factory() as session:
+            return set(inspect(session.connection()).get_table_names())
+
+    assert {"documents", "model_endpoints", "mcp_servers", "calendars"} <= expected
+    assert expected <= table_names(database.SessionLocal)
+    for _ in range(8):
+        # A fresh executor creates a fresh thread rather than reusing one.
+        with ThreadPoolExecutor(max_workers=1) as executor:
+            assert expected <= executor.submit(table_names, database.SessionLocal).result()
+    assert expected <= table_names(database.SessionLocal)
+    assert expected <= table_names(compatibility.SessionLocal)