test: fix WS test port race; narrow to single smoke test covering both transport ends

[maxisbey]

Problem

tests/shared/test_ws.py had 4 tests sharing a subprocess+uvicorn fixture with a TOCTOU port race: bind port 0 → release → subprocess rebinds. Under pytest-xdist another worker claims the port in the gap, and the WS client connects to the wrong server's HTTP endpoint, yielding a 403 Forbidden on upgrade (job 63414508256).

Disposition of the 4 tests

Only one of the four was actually testing WebSocket. The other three tested MCP session semantics with WS as an incidental pipe — swap in any transport and they pass identically.

Test	What it tested	Outcome
test_ws_client_basic_connection	Init + ping over WS — subprotocol handshake, JSON-RPC serialization round-trip	Kept
test_ws_client_happy_request_and_response	read_resource() returns the right content	Duplicate of test_client.py::test_read_resource → deleted
test_ws_client_exception_handling	Server-raised error surfaces as MCPError with correct code	Moved to test_client.py::test_read_resource_error_propagates (in-memory)
test_ws_client_timeout	Session survives a client-side timeout, can issue requests after	Duplicate of test_88_random_error.py → deleted

~150 of the old 211 lines were subprocess launcher + port-polling + fixture plumbing shared by all four. With one test remaining on a shared helper, the file is 51 lines.

The smoke test

test_ws_client_basic_connection now uses run_uvicorn_in_thread (tests/test_helpers.py), which pre-binds the socket and calls listen() before the server thread starts. The port is known immediately — the kernel's listen queue buffers any early connection until uvicorn's loop reaches accept(). No polling, no sleep, no gap for another xdist worker to steal the port.

Knock-on effects of in-thread

Two subprocess-era assumptions stopped holding once uvicorn runs in the test process:

• Coverage now reaches server/websocket.py. The function-level # pragma: no cover existed because the subprocess couldn't be instrumented; strict-no-cover (ci: run strict-no-cover in scripts/test to catch stale pragmas locally #2305) correctly flags it as stale. Narrowed to the three error paths the smoke test doesn't exercise. The one test now covers both client/websocket.py and server/websocket.py at 100%.

• pytest's warning filters now apply to uvicorn. uvicorn.Config.load() calls asyncio.iscoroutinefunction() during ASGI autodetection; Python 3.14 deprecates it; filterwarnings=error promotes it; the server thread crashes silently before binding. Fixed by passing interface="asgi3" to skip the autodetect — Starlette is asgi3, the detection was always going to land there.

Test plan

• ./scripts/test (3.13) — 100% coverage, strict-no-cover clean ✓
• uv run --python 3.14 pytest tests/shared/test_ws.py — 0.46s ✓
• pytest tests/shared/test_ws.py -n 4 --flake-finder --flake-runs=5 — 20 parallel runs, 0 flakes ✓

_{AI Disclaimer}