[Piergiuseppe]

Two tests that have nothing wrong with the code they test fail now and then, and with ci-passed required each failure blocks a merge.

• test_the_cli_exits_within_its_shutdown_timeout_while_a_job_ignores_cancellation failed on pull request Add tinyjobs move, and delete in the backend protocol #18 (Python 3.13, Postgres 14): a worker with a 2 s shutdown timeout took 2.27 s to exit on the runner, against a bound of 2.15. The tests that time a worker process gave it 0.1, 0.15, 0.3 or 0.6 s to exit, depending on the test. They now share PROCESS_EXIT_SECONDS = 0.6, which is still under the 1.25 s of the next phase a regression would add.
• test_the_cli_exits_within_the_shutdown_timeout_with_outcomes_stuck_on_a_lock failed about half the time against a local Postgres 17, on main too: the cleanup dropped the test's tables one by one while a statement the exited worker had left running on the server still held one of them, and Postgres reported a deadlock. The cleanup now drops them in one statement and tries again after a deadlock. Six runs in a row pass.

Only tests change.

Still open, not in this pull request: test_one_stop_signal_ends_a_process_whose_thread_writes_to_a_stalled_stdout timed out three times in about ten full local runs under load and passes alone; not understood yet.