Flaky test: pipeline job control on macOS free-threaded Python (wrapper-stop-and-interrupt-busy-shell-producer) #1791

Summary

tests/test_pipeline_job_control.py::test_pipeline_stops_with_cmd2_and_returns_terminal[wrapper-stop-and-interrupt-busy-shell-producer]
fails intermittently on the tests (macos-latest, 3.15t) CI job. It has failed on branches that don't
touch pipeline or job-control code, so it looks like a pre-existing race in either the test or cmd2's
job-control handling.

Failing runs

All of these failed on macos-latest, 3.15t with the same test variant:

• https://github.com/python-cmd2/cmd2/actions/runs/40803872074 (fix/window-title-flush-v2, Flush window title updates without waiting for a prompt #1790)
• https://github.com/python-cmd2/cmd2/actions/runs/40802879714 (fix/window-title-flush, Flush window title updates without waiting for a prompt #1783)
• https://github.com/python-cmd2/cmd2/actions/runs/40150857988 (fix/window-title-flush, Flush window title updates without waiting for a prompt #1783)
• https://github.com/python-cmd2/cmd2/actions/runs/39753826343 (fix/history-range-past-end)

A related variant, [orphaned-job-control], failed once on macos-latest, 3.15 (not free-threaded):
https://github.com/python-cmd2/cmd2/actions/runs/40127777165

What fails

This variant uses finish="exit_sigint", stop_job=True, shell_child=True, launcher="sh",
producer="shell" and relay="worker". The test runs busy | bash -c 'python pager.py; :' from cmd2
inside an interactive bash, sends Ctrl-Z, and then waits for the terminal's foreground process group to
return to the outer shell (tests/test_pipeline_job_control.py:335):

send("\x1a")
wait_until(lambda: os.tcgetpgrp(master) == process.pid)

This times out on the first Ctrl-Z. From the failure output of run 40803872074:

• The transcript ends with PAGER_READY, BUSY_READY and then PAGER_RESUMED. The pager prints
  PAGER_RESUMED from its SIGCONT handler, but the test never sent fg. So the pipeline seems to have
  been stopped and then continued, rather than staying stopped.
• SHELL_OWNS_INPUT never appears, so the outer shell never got the terminal back.
• The foreground process group is still the pipeline's group, and every process is running (S+), not
  stopped (T):

foreground process group: 2567
  PID  PPID  PGID STAT  COMMAND
 2559  2089  2559 S<s   bash --noprofile --norc -i
 2560  2559  2560 S<    /bin/bash -c .../python .../application.py; :
 2561  2560  2560 S<    PythonT .../application.py                      (cmd2)
 2567  2561  2567 S<+   /bin/bash -c .../python .../pager.py; :
 2568  2567  2567 S<+   PythonT .../pager.py
 2570  2561  2567 S<+   PythonT .../busy.py
resizes: []
terminal size at timeout: (24, 80)

Notes for investigation

• In this variant (relay="worker"), the test application patches signal.pthread_kill so the
  job-control relay signal goes to the watcher thread that sends it, rather than to the main thread
  blocked in a pipe write or wait. All the free-threaded failures are for this variant, so the race may
  be in how cmd2 relays SIGTSTP/SIGCONT and hands the terminal back when the main thread isn't the one
  handling the signal.
• It may be a timing problem in the test (for example Ctrl-Z arriving while the pipeline is still being
  set up) or a real bug where cmd2 continues the pipeline without stopping itself and returning the
  terminal to its parent shell.
• It's intermittent: most macOS 3.15t runs pass, including recent runs on main.