[Tests] Replace flaky sleep with polling in test_background_cancel

[sjhddh]

Summary

Replace the fixed 0.5s sleep (marked with # FIXME: This test can be flaky) with a proper polling loop that waits for the response status to change before attempting to cancel.

Changes

• Poll with 0.1s intervals until status changes from "queued" (max 5s)
• Use proper assertions in the post-cancel verification loop
• Remove FIXME comment as the flakiness is addressed

Motivation

• Fixed sleeps are unreliable: too short on slow machines, wasteful on fast ones
• Polling adapts to actual server response times
• Reduces test flakiness in CI

Test Plan

• Run pytest tests/v1/entrypoints/openai/serving_responses/test_stateful.py::test_background_cancel multiple times
• Verify test passes consistently

Risk

• Low risk: only changes test synchronization logic
• No changes to production code

[gemini-code-assist]

Code Review

This pull request improves the reliability of test_background_cancel by replacing a fixed sleep with a polling mechanism. This is a great change to address test flakiness. I've suggested a further refinement to the polling loop to use a monotonic clock for more accurate timeout handling, which will make the test even more robust.

[gemini-code-assist]

The current polling loop uses elapsed += poll_interval to track time, but this doesn't account for the time spent in the client.responses.retrieve(response.id) call. This can cause the effective timeout to be significantly longer than max_wait_seconds if the API call is slow, potentially slowing down tests in CI.

Using a monotonic clock like asyncio.get_running_loop().time() provides a more accurate and robust timeout mechanism. I've also reordered the loop to check the condition before sleeping, which is slightly more efficient.

Suggested change

	max_wait_seconds = 5.0
	poll_interval = 0.1
	elapsed = 0.0
	while elapsed < max_wait_seconds:
	await asyncio.sleep(poll_interval)
	elapsed += poll_interval
	response = await client.responses.retrieve(response.id)
	if response.status != "queued":
	# Started processing or completed - try to cancel
	break
	loop = asyncio.get_running_loop()
	start_time = loop.time()
	max_wait_seconds = 5.0
	poll_interval = 0.1
	while loop.time() - start_time < max_wait_seconds:
	response = await client.responses.retrieve(response.id)
	if response.status != "queued":
	# Started processing or completed - try to cancel
	break
	await asyncio.sleep(poll_interval)