fix: configure uvicorn properly for 3.14 and Windows instead of abandoning shutdown

Two root-cause fixes, no warning suppression:

Python 3.14:
  uvicorn's interface='auto' autodetect calls asyncio.iscoroutinefunction
  on the app's __call__ to distinguish ASGI2 from ASGI3. That function is
  deprecated on 3.14, and pyproject.toml's filterwarnings=error promotes
  the warning to an exception in the server thread — which dies silently
  while our pre-bound socket keeps accepting connections into the kernel
  queue, turning a loud startup crash into a 5-minute ReadTimeout.

  Starlette apps are ASGI3. Setting interface='asgi3' skips the autodetect
  entirely — we're just telling uvicorn something we already know.

Windows Proactor:
  force_exit=True short-circuits _wait_tasks_to_complete's connection-drain
  loops immediately. connection.shutdown() was called, but we never waited
  for the transports to actually close; they still have pending Overlapped
  Recv operations when the event loop is torn down, and GC later finds
  them during an unrelated test.

  timeout_graceful_shutdown=1 gives uvicorn a bounded window to drain
  connections naturally, then on timeout calls t.cancel() on remaining
  tasks — proper asyncio cancellation that unwinds the transports through
  their CancelledError paths instead of abandoning them mid-I/O. Tests
  that close their httpx clients cleanly (all of them, in practice) never
  hit the timeout; the drain loop finds connections already empty.

Author: maxisbey

tests/test_helpers.py

@@ -65,6 +65,18 @@ def run_uvicorn_in_thread(app: Any, **config_kwargs: Any) -> Generator[str, None
     """
     host = config_kwargs.setdefault("host", "127.0.0.1")
     config_kwargs.setdefault("log_level", "error")
+    # Starlette apps are ASGI3. Being explicit skips uvicorn's autodetect,
+    # which calls asyncio.iscoroutinefunction — deprecated on 3.14 and
+    # promoted to an error by pyproject.toml's filterwarnings, killing the
+    # thread before it starts serving.
+    config_kwargs.setdefault("interface", "asgi3")
+    # Without a bound, shutdown() waits indefinitely for open connections
+    # (SSE streams left open at test exit) to drain. force_exit short-circuits
+    # that wait but abandons connection transports mid-I/O — on Windows
+    # Proactor those are Overlapped Recv ops that GC later finds still
+    # pending. timeout_graceful_shutdown waits briefly, then cancels the
+    # connection tasks via asyncio cancellation so transports unwind cleanly.
+    config_kwargs.setdefault("timeout_graceful_shutdown", 1)
 
     sock = socket.socket(socket.AF_INET, socket.SOCK_STREAM)
     sock.setsockopt(socket.SOL_SOCKET, socket.SO_REUSEADDR, 1)
@@ -81,6 +93,5 @@ def run_uvicorn_in_thread(app: Any, **config_kwargs: Any) -> Generator[str, None
         yield f"http://{host}:{port}"
     finally:
         server.should_exit = True
-        server.force_exit = True
         thread.join(timeout=5)
         sock.close()