fix: discard dead pool connections instead of reconnecting to stale leader

When a pooled connection was dead (is_connected=False), the pool
called conn.connect() which reconnected to the original address.
After a leader election, this reconnects to a follower. Now discards
the dead connection and creates a fresh one via leader discovery.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: antoineleclair

src/dqliteclient/pool.py

@@ -90,11 +90,13 @@ async def acquire(self) -> AsyncIterator[DqliteConnection]:
                     f"(max_size={self._max_size}, timeout={self._timeout}s)"
                 ) from None
 
+        # If connection is dead, discard and create a fresh one with leader discovery
+        if not conn.is_connected:
+            self._size -= 1
+            conn = await self._create_connection()
+
         self._in_use.add(conn)
         try:
-            # Verify connection is still good
-            if not conn.is_connected:
-                await conn.connect()
 
             yield conn
         except BaseException:

tests/test_pool.py

@@ -109,6 +109,33 @@ async def test_acquire_timeout_when_pool_exhausted(self) -> None:
                     pass
 
 
+    async def test_dead_connection_triggers_leader_rediscovery(self) -> None:
+        """A dead connection should be replaced via leader discovery, not reconnected."""
+        pool = ConnectionPool(["localhost:9001"], max_size=1)
+
+        dead_conn = MagicMock()
+        dead_conn.is_connected = False  # Connection is dead
+        dead_conn.connect = AsyncMock()
+        dead_conn.close = AsyncMock()
+
+        new_conn = MagicMock()
+        new_conn.is_connected = True
+        new_conn.connect = AsyncMock()
+        new_conn.close = AsyncMock()
+        new_conn.execute = AsyncMock(return_value=(1, 1))
+
+        # Initialize with the connection that will go dead
+        with patch.object(pool._cluster, "connect", return_value=dead_conn):
+            await pool.initialize()
+
+        # When acquiring, the dead conn should be discarded and a new one created
+        with patch.object(pool._cluster, "connect", return_value=new_conn):
+            async with pool.acquire() as conn:
+                assert conn is new_conn  # Got a fresh connection, not the dead one
+
+        # Dead connection should NOT have had connect() called (no stale reconnect)
+        dead_conn.connect.assert_not_called()
+
     async def test_close_handles_checked_out_connections(self) -> None:
         """close() should close in-flight connections, not just idle ones."""
         pool = ConnectionPool(["localhost:9001"], max_size=2)