[linuxnow]

Based on #173, by Sami Jawhar (sjawhar). His two commits are cherry-picked onto the current main with their authorship, so the checks run against the contract this main renders with; the branch of #173 was cut before it.

Closes #113.

Two outcomes of the test job depended on timing in its fedora:44 container on a loaded host.

• a_worker_line_too_long_to_relay_is_dropped_whole, run alone 60 times, missed the sabotage arm "the end of a worker line too long to relay is sent as a line" 9 times. The relay's buffer stops doubling at 2 MiB, the length of the test's longest line. When one read fills it with that line, the read stops just before the line's NUL. With the arm, the line's end arrives as an empty line, which the check did not count. The check now counts it. It catches the arm in 100 of 100 runs and passes 100 of 100 without it.
• The two "left in its own session" probe tests failed in 10 of 60 runs (30 of each). The fork fault returned before its leftover process had written its pid. The runner killed the leftover before that write, so the pidfile was missing or empty. No leftover outlived a run, so the runner is not at fault. Writing the pid by rename, as test-probe flake: the fork fixture's pidfile can be read empty #113 proposes, would leave the file missing, which the test cannot tell from a fault that never forked. The fault now waits on a pipe that the leftover closes once its pid is written. After the change, 120 of 120 runs pass (60 of each). With a temporary 200 ms sleep after the leftover's setsid(), both tests fail every run before the change and pass every run after.

The steps of ci.yml's private-refs and test jobs and of fuzz.yml pass 3 of 3 runs there.