One-GPU serialization: the chain waits for the kld driver log to close
rc=0, sanity-checks kld-ref, boots the calibration server, pushes the
corpus at parallel 16 (4 engine slots; queue absorbs the rest), then
kills the server so its trap restores production.
