test(eval): harden relay descendant cleanup against process-group timing

[liugddx]

Problem

The test job for #3722 failed in the unrelated eval lifecycle suite at packages/eval/harbor/test_relay_lifecycle.py:648:

RelayLifecycleTest.test_settle_kills_descendants_before_verification_boundary
self.assertFalse(late_write.exists())
AssertionError: late_write remained present

The test starts a setsid --wait process whose parent exits while a forked descendant writes late-write after 300 ms. The relay should quiesce the entire process group before the verifier observes the workspace, but this run observed the delayed write.

Failure evidence: https://github.com/apache/maka/actions/runs/32754131081/job/97517633860

Scope

This is pre-existing and unrelated to #3722:

• test(windows): close packaged sandbox lifecycle evidence #3722 does not modify packages/eval/harbor/test_relay_lifecycle.py or relay_agent.py.
• The file blobs are identical between the PR head and upstream/main.
• The same eval suite passed on a later main run: https://github.com/apache/maka/actions/runs/32765873057/job/97559885543
• test(windows): close packaged sandbox lifecycle evidence #3722's package and Windows sandbox W0 checks passed; only the generic test job failed.

Suspected area

packages/eval/harbor/relay_agent.py uses the scope PID file and kill -0 -- -<pgid> to decide whether a process group is active before sending TERM/KILL. The failure suggests a timing or process-group observability gap: the scope was treated as inactive before the descendant's delayed write, so no effective termination was observed. The exact pgid/signal trace is not currently logged.

Suggested follow-up

1. Add raw scope PID, process-group liveness, signal return code, and post-signal verification evidence.
2. Make descendant quiescence robust when the group leader exits before the descendant.
3. Keep the regression test deterministic on the Linux CI runner and verify repeated runs.

This issue should be tracked separately from the Windows sandbox lifecycle work in #3722.