[tonghuaroot]

BaseProactorEventLoop._start_serving() ran protocol_factory() and transport creation in the accept callback guarded only by except OSError / except CancelledError. A per-connection failure therefore leaked the accepted socket and either stopped the accept loop (non-OSError) or closed the listening socket (OSError).

Wrap the per-connection setup so the accepted socket is closed and the error is reported to the loop exception handler, while accept-stage errors keep their existing handling. This is the proactor counterpart of gh-155934 (selector) and gh-153133 (create_connection).

The added regression tests fail without the fix and pass with it; the full test_asyncio suite passes.

• Issue: asyncio proactor server leaks the accepted socket and stops accepting when transport creation fails #156365

[OsmnvAslan]

Ran this locally: it merges cleanly into current main, test_asyncio passes, and both new tests fail without the fix. It matches the selector change from gh-155934, so LGTM.

One small thing: the new context = {...} inside loop() shadows the context parameter of _start_serving(). It's unused in the proactor path right now, so nothing breaks, but renaming it (e.g. exc_context) would avoid a surprise if context=context is ever passed to the transport here like in the selector version.