[ioquatix]

Correct three figures from #66 that an independent review found overstated or misattributed, checked against samples.csv, and make the screen's layout.py unable to leave a server running.

• Layout A in raw-closed was below two workers in every round, but by 5.3% and 5.8% at the median and 3.5–6.7% in single rounds, not 5–6% in every round.
• "B-u was within 1.1% of B" was the gap between the two "vs C" columns. By the median of each round's ratio, B-u was within 0.3% of B except for hyper-tokio in http-closed, at +2.1% (+1.6% between medians).
• hana's host document: the worst single round was 8.9% below its cohort's median (base-native, A, server CPUs 2,3), so up to 9%, not 8%.

layout.py sample now gives up if the server prints no address within 60 seconds, which previously held the host lock indefinitely. It also turns SIGHUP, SIGINT and SIGTERM into an exit through its cleanup, because the server and client run in their own sessions and a terminal's signals never reached them. The campaign lets the sample's stderr through so that a failure's reason shows. A successful sample measures and prints the same as before, and no measurements were rerun. The results readme notes that these changes came after the campaigns.