[paskal]

Previously, the shutdown tests slept for a fixed period and then asserted that the injected exit function had been called. That fails whenever the scheduler runs the timer later than the sleep, which is what happens on a loaded CI runner. The injected exit now reports through a channel, so each test waits only as long as it actually needs and fails on a deadline instead of on a guess.

Three cases needed more than a mechanical replacement:

• The second-signal case ran with a 500ms shutdown timeout while asserting on an exit that the second signal was supposed to cause; the timeout could have produced that exit instead, so the test could pass for the wrong reason. Its timeout is now long enough that only the second signal can trigger the exit.
• The concurrent-signals case keeps a short timeout deliberately. os/signal drops a send to a full channel, so a burst of signals may well be delivered once and the exit has to come from the timeout; the deterministic claim, that a second signal forces the exit, is what the case above covers. It now runs on a signal no other case sends and joins its senders, so nothing it sends can reach a later case.
• The timeout accuracy case measured from before the signal that starts the timer and asserted a 150ms window around it. It now measures from the shutdown callback, runs on its own signal, and keeps the meaningful bound, that the forced exit never happens early; the upper bound stays loose enough to survive a delayed timer while still catching a timeout off by an order of magnitude.

The HTTP tests no longer sleep before the first request either: the listener is created before the server starts and already accepts connections, so the request blocks until Serve picks it up. Those requests now carry a client timeout, so a stuck server fails the test instead of hanging it. The sleeps in the example tests stay, as they pace the printed output and removing the one in Example_httpServerWithContext would race the server start against the cancellation.