[alex-feel]

Fixes #1397.

_create_ssl_connection() creates TLSUpgradeProto inside loop.create_connection(). When that call fails or is cancelled after the protocol exists, the transport is closed and connection_lost(None) sets ConnectionError('unexpected connection_lost() call') on on_data, which nothing awaits any more, so asyncio logs "Future exception was never retrieved". Since 0.32.0 Pool.close() cancels the floor-maintenance connect, so closing a pool while the server is unresponsive can log it.

on_data is now cancelled, or its exception retrieved, when create_connection() does not return, the way __connect_addr() handles the connected future. test_connection_cancelled_during_transport_setup covers ssl='prefer' as well; it fails on master with the two unretrieved futures and passes with the change.

[janbernloehr]

Reviewed 7d41291 and independently tested its TLS helper against asyncpg 0.32.0 on CPython 3.14.5/macOS.
Across 54 real loopback TCP/TLS comparison runs, the baseline emitted nine unretrieved-Future reports; the patched helper emitted none. Cancellation, deadlines, connection/TLS/certificate failures, verified TLS success and advisory negotiation retained the same primary outcomes. The updated regression test body also failed before the fix and passed afterward in an isolated harness. This was local validation, not a full upstream suite run.
The protocol capture and Future cleanup address the reported ownership gap while preserving the original exception.
One non-blocking coverage suggestion: exercise an ordinary connection exception alongside CancelledError, including failure before the protocol factory is called. This would cover the upgrade_proto is None branch. These additional cases passed in my local probe.