Sweep leaked ROS processes from dead agent worktrees

[MJohnson459]

Adds pixi run sweep to reap ROS processes leaked by dead agent jobs, fixes the
pixi run kill foot-gun that motivated it, and closes the leak at source.

WHAT WAS THERE: 44 orphaned ROS processes, 299 MB, up to 4 days old, from five
ended jobs -- 19 twist_mux, 7 icp_odom_gate, 5 foxglove_bridge, 5 ros2 wrappers,
3 mosquitto, 3 twist_relay, task_server, a benchmark recorder.

THE SWEEP (mote_bringup/sweep_orphans.py, stdlib-only, pixi run sweep, reports
by default, --kill to act, --json). Matching is on identity in /proc: a ROS
environment, a path under an agent root, absence from the sweeper's own
ancestry, orphaned by ancestry, and older than --min-age (30 min). Two findings
changed the design from the task's sketch:

• /proc//exe alone misses a third of them. A worktree running against the
  primary checkout's pixi env has its interpreter and binaries outside the job
  dir; only the ament/colcon/pixi prefixes implicate it. Provenance is exe +
  cwd + argv paths + those prefixes.
• ancestry alone would kill live work. run_sim_smoke.sh setsids its launch,
  so a healthy smoke test has ppid 1 and looks orphaned. --min-age separates
  them, and reporting is the default.

pixi run kill HAD THE FOOT-GUN THE TASK WARNED ABOUT. Its pkill -9 -f 'robot_state_publisher|...' matched the shell running the task -- those names
are in its own command line -- SIGKILLed it (reproduced: exit 137), so the
ros2 daemon stop/start; rm -f /dev/shm/fast* half NEVER RAN. It also matched
every other checkout and worktree on the box, not just the current checkout as
the task assumed. It now uses the same /proc path scoped to the checkout
(--scope .), without the orphan/age tests since a wedged stack is live.

HOW THEY ESCAPED -- not the guessed cause. Not interactive pixi run robot/
sim: it is the test fixtures, on the SUCCESS path. ros2 run Popens the real
executable and installs no SIGTERM handler (only KeyboardInterrupt, assuming a
terminal Ctrl-C reached the whole group), so proc.terminate() on the wrapper
kills the wrapper and hands the node to init -- one leaked node per run.
Confirmed in ros2run/api/init.py and reproduced live. spawn_reapable/
reap_group put the child in its own session and signal the group; applied to
test_twist_mux_arbitration.py, test_foxglove_teleop.py and
tools/icp_gate_replay.py (bench.py/replay.py/camera_layer_decay.py already did
this by hand). A job killed outright, taking pytest with it, no fixture can fix
-- that is the sweep's half.

VERIFIED

• 28 new tests pass (mote_bringup/test/test_sweep_orphans.py). Mostly about
  what the sweep DECLINES to kill: a run owned by a shell, a young detached
  run, a file manager holding a job path, the sweeper's own ancestry, another
  checkout. Selection is a pure function of a process table so those are
  fabricated; two cases spawn a real orphan via double-fork and reap it.
• Leak fix measured on the real test: pre-fix fixture, 6 pass and 1 twist_mux
  survives; post-fix, 6 pass and 0 survive.
• Dry run on the live box finds all 44 leaks and correctly skips claude bg-spare, cosmic-files and the shells that a pkill -f would have hit.
• pre-commit clean; pixi arg passthrough confirmed for the task shape.
• test_twist_mux.py::test_the_base_launches_the_mux fails identically on
  unmodified main under this ad-hoc pytest invocation -- pre-existing, an
  install-space path, not from this change.

NOT DONE: the 44 processes are still running. The permission classifier blocked
the actual kill, which is the right call for a destructive machine-wide action.
Run pixi run sweep to review, then pixi run sweep -- --kill to reap. Note
that pixi run kill cannot be exercised until this branch is in the checkout
the task runs from.