test: remove scheduling sensitivity from the prestart timeout test

The second hook now blocks indefinitely, so the phase deadline lands in it
regardless of scheduler load. Also notes in the docs that threads started
directly do not inherit the capture context.

Author: jasonwang-runpod

docs/serverless/worker.md

@@ -74,7 +74,8 @@ to the failure it reports, as `logs`. Opt in using `RUNPOD_LOG_CAPTURE`:
 
 Capture replaces `sys.stdout`/`sys.stderr` at worker startup, so it sees `print`
 and direct stream writes made while a hook or handler runs. It does not see
-child processes and log handlers created before startup.
+child processes, log handlers created before startup, or threads your code
+starts directly (`asyncio.to_thread` is captured).
 
 ## Worker Refresh
 

tests/test_serverless/test_prestart.py

@@ -81,15 +81,20 @@ def warm_cache():
         assert payload["error_message"] == "CUDA OOM"
 
     def test_timeout_bounds_the_whole_phase_and_names_current_hook(self):
+        started = []
+
         async def first():
-            await asyncio.sleep(0.03)
+            started.append("first")
 
         async def second():
-            await asyncio.sleep(0.03)
+            started.append("second")
+            # Blocks forever, so the deadline lands here on any scheduler.
+            await asyncio.Event().wait()
 
         with self.assertRaises(PrestartTimeout) as ctx:
-            _run(run_prestart_hooks_async((first, second), timeout=0.04))
+            _run(run_prestart_hooks_async((first, second), timeout=0.05))
 
+        assert started == ["first", "second"]
         assert ctx.exception.hook == "second"
 
     def test_hook_raised_cancelled_error_is_a_prestart_failure(self):