Lab: graceful shutdown is a protocol¶

The scripts behind the shutdown measurements: two servers with the same two routes, and a driver that plays Kubernetes against them.

uv venv --python 3.13 .venv
uv pip install --python .venv/bin/python "servicewright[fastapi]" httpx
.venv/bin/python driver.py server_plain.py
.venv/bin/python driver.py server_servicewright.py                  # drain_delay_seconds=0
DRAIN_DELAY=1.5 .venv/bin/python driver.py server_servicewright.py  # the protocol with its window

Script	What it is
`routes.py`	`/work` (20 ms) and `/slow` (2 s), shared by both servers
`server_plain.py`	FastAPI under `uvicorn.run`, uvicorn's own SIGTERM handling
`server_servicewright.py`	the same routes under servicewright's lifecycle; `DRAIN_DELAY` sets `AppSpec.drain_delay_seconds`
`driver.py`	starts a server, waits for readiness, starts one slow request, sends SIGTERM, keeps sending one request every 20 ms for `LAG` seconds (the endpoint propagation lag), polls readiness, and counts what came back