fix(core): shield remaining asyncpg dispose seams + observe nest_asyncio skip

Address claude-review feedback and the Windows SQLite Unit CI failure on PR #902.

- db.py engine_session_factory finally: shield created_engine.dispose() with
  suppress(CancelledError) to match the other dispose seams (review item 2).
- db.py run_migrations finally: shield temp_engine.dispose() the same way; it is
  on the same asyncpg startup path as the crash (review item 1).
- alembic/env.py: log the nest_asyncio ImportError/ValueError at DEBUG instead of
  silently passing, so the now-expected uvloop ValueError on every Postgres
  startup is observable without being noisy (review item 3).
- tests/cli/test_cli_exit.py: loosen test_bm_version_exits_cleanly subprocess
  timeout 10s -> 20s to match its sibling --help test. The new uvloop code never
  runs on the 'bm --version' path (eager version_callback exits before
  app_callback's body, and db.py is not even imported), so the Windows-only
  failure is uv-run startup contention under full-suite load, not a regression.

Verified: ruff, ty, the affected unit tests, and the Postgres dispose-cycle
integration test (testcontainers) all pass locally.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Signed-off-by: phernandez <paul@basicmachines.co>

Author: phernandez

src/basic_memory/alembic/env.py

@@ -5,6 +5,8 @@
 from contextlib import suppress
 from logging.config import fileConfig
 
+from loguru import logger
+
 # Allow nested event loops (needed for pytest-asyncio and other async contexts)
 # Note: nest_asyncio doesn't work with uvloop or Python 3.14+, so we handle those cases separately
 import sys
@@ -14,9 +16,15 @@
         import nest_asyncio
 
         nest_asyncio.apply()
-    except (ImportError, ValueError):
-        # nest_asyncio not available or can't patch this loop type (e.g., uvloop)
-        pass
+    except (ImportError, ValueError) as exc:
+        # Trigger: nest_asyncio is absent (ImportError) or refuses to patch the
+        # running loop (ValueError, e.g. the uvloop policy installed for the
+        # Postgres backend - #831/#877).
+        # Why: the uvloop ValueError is now an *expected* path on every Postgres
+        # startup, so swallowing it silently hides a routine branch.
+        # Outcome: log at DEBUG (observable, not noisy) and fall through to the
+        # thread-based migration fallback.
+        logger.debug(f"nest_asyncio not applied ({exc!r}); using thread-based migration fallback")
 # For Python 3.14+, we rely on the thread-based fallback in run_migrations_online()
 
 from sqlalchemy import engine_from_config, pool

src/basic_memory/db.py

@@ -410,7 +410,14 @@ async def engine_session_factory(
 
         yield created_engine, created_session_maker
     finally:
-        await created_engine.dispose()
+        # Trigger: context-manager teardown can run while the surrounding task is
+        # being cancelled (e.g. a test aborting mid-fixture).
+        # Why: on the asyncpg backend a cancellation landing mid-dispose surfaces
+        # the "IndexError: pop from an empty deque" race (#831/#877); shield the
+        # dispose and suppress CancelledError to match the other dispose seams.
+        # Outcome: the per-context engine always disposes cleanly under cancellation.
+        with suppress(asyncio.CancelledError):
+            await asyncio.shield(created_engine.dispose())
 
         # Only clear module-level globals if they still point to this context's
         # engine/session. This avoids clobbering newer globals from other callers.
@@ -479,7 +486,11 @@ async def run_migrations(
         # Trigger: run_migrations() created a temporary engine while module-level
         # session maker was not initialized.
         # Why: temporary aiosqlite worker threads can outlive CLI command execution
-        # and block process shutdown if the engine is not disposed.
-        # Outcome: always dispose temporary engines after migration work completes.
+        # and block process shutdown if the engine is not disposed. On the asyncpg
+        # backend a cancellation landing mid-dispose surfaces the same "IndexError:
+        # pop from an empty deque" race as the other dispose seams (#831/#877), so
+        # shield the dispose and suppress CancelledError to match them.
+        # Outcome: always dispose temporary engines cleanly, even under cancellation.
         if temp_engine is not None:
-            await temp_engine.dispose()
+            with suppress(asyncio.CancelledError):
+                await asyncio.shield(temp_engine.dispose())

tests/cli/test_cli_exit.py

@@ -15,7 +15,11 @@ def test_bm_version_exits_cleanly():
         ["uv", "run", "bm", "--version"],
         capture_output=True,
         text=True,
-        timeout=10,
+        # `uv run` startup under full-suite load (especially on Windows runners)
+        # can exceed a tight 10s budget even though --version short-circuits
+        # before any heavy work. This test guards against hangs, not a startup
+        # performance budget, so match the looser --help timeout below.
+        timeout=20,
         cwd=Path(__file__).parent.parent.parent,  # Project root
     )
     assert result.returncode == 0