fix(tests): suppress uvicorn's 3.14 DeprecationWarning in thread, rejoin

Two CI failures, same root cause of running uvicorn in-process instead
of a subprocess:

1. Python 3.14: asyncio.iscoroutinefunction is deprecated. uvicorn's
   config.load() calls it, emitting a DeprecationWarning. Pytest's
   filterwarnings=error applies to the uvicorn thread (same process),
   so the warning becomes an exception that kills the thread before
   server.started is set. The old subprocess fixtures masked this —
   subprocesses don't inherit pytest's warning filters. Fix: wrap
   server.run in a catch_warnings block that ignores DeprecationWarning.
   Also detect a dead thread in the startup poll so we fail fast rather
   than spinning for 20s on future thread crashes.

2. Windows 3.12/3.13: Proactor pipe transports leak as unclosed sockets
   when the daemon thread is reaped before its event loop finishes
   shutdown. The prior perf commit removed thread.join() to save ~200ms
   per fixture, but that doesn't hold on Windows. Bring back the join.
   Module-scoped fixtures absorb most of the cost anyway (one server
   per module instead of per-test).

Author: maxisbey

tests/test_helpers.py

@@ -67,12 +67,28 @@ def run_server_in_thread(app: ASGIApp, lifespan: Literal["auto", "on", "off"] =
     config = uvicorn.Config(app=app, host="127.0.0.1", port=0, log_level="error", lifespan=lifespan)
     server = uvicorn.Server(config=config)
 
-    thread = threading.Thread(target=server.run, daemon=True)
+    def _run() -> None:
+        # On Python 3.14, asyncio.iscoroutinefunction raises a DeprecationWarning
+        # when uvicorn's config.load() calls it. Pytest's `filterwarnings = error`
+        # applies to this thread (same process), so the warning becomes an
+        # exception that kills the thread before server.started is set. The old
+        # subprocess fixtures masked this because subprocess doesn't inherit
+        # pytest's warning filters. catch_warnings is thread-local since 3.12;
+        # on 3.10/3.11 it's process-global but the warning doesn't fire there.
+        with warnings.catch_warnings():
+            warnings.filterwarnings("ignore", category=DeprecationWarning)
+            server.run()
+
+    thread = threading.Thread(target=_run, daemon=True)
     thread.start()
 
-    # Wait for uvicorn to bind and start accepting connections
+    # Wait for uvicorn to bind and start accepting connections. If the thread
+    # dies (e.g., config.load() raised), bail early with the thread's exception
+    # rather than spinning for 20s.
     start_time = time.time()
     while not server.started:
+        if not thread.is_alive():  # pragma: no cover
+            raise RuntimeError("uvicorn thread exited before server started")
         if time.time() - start_time > 20.0:  # pragma: no cover
             raise TimeoutError("uvicorn server did not start within 20 seconds")
         time.sleep(0.01)
@@ -82,20 +98,18 @@ def run_server_in_thread(app: ASGIApp, lifespan: Literal["auto", "on", "off"] =
     try:
         yield f"http://127.0.0.1:{port}"
     finally:
+        # force_exit skips uvicorn's graceful connection drain. We still join
+        # the thread: on Windows, Proactor pipe transports leak as unclosed
+        # sockets if the daemon thread is reaped before its event loop finishes
+        # shutdown. Module-scoped fixtures make the ~200ms join cost acceptable
+        # (one server per module instead of one per test).
         server.should_exit = True
         server.force_exit = True
-        # Don't block on thread.join() — uvicorn polls should_exit every 0.1s and
-        # its shutdown() adds another 0.1s, totaling ~200ms teardown latency per
-        # fixture. The thread is a daemon so it will be reaped by the interpreter;
-        # we just signal exit and move on. The socket is closed by uvicorn's
-        # shutdown regardless, so the port is released for the next test.
-        # When uvicorn shuts down with in-flight SSE connections, the server
-        # cancels request handlers mid-operation. SseServerTransport's internal
-        # memory streams may not get their `finally` cleanup run before GC,
-        # causing ResourceWarnings. These are artifacts of test abrupt-disconnect
-        # patterns (open SSE stream → check status → exit without consuming),
-        # not bugs. Force GC here and suppress the warnings so they don't leak
-        # into the next test's PytestUnraisableExceptionWarning collector.
+        thread.join(timeout=5)
+        # SseServerTransport's internal memory streams may be GC'd without
+        # finalizers when uvicorn cancels in-flight SSE handlers. Force GC here
+        # and suppress the ResourceWarnings so they don't leak into the next
+        # test's PytestUnraisableExceptionWarning collector.
         with warnings.catch_warnings():
             warnings.simplefilter("ignore", ResourceWarning)
             gc.collect()