Remove systemd from Backplane: agents and work_run run as plain processes, restarts are picked back up

[i2cjak]

Backplane no longer uses systemd for its own work. Keeping work going across restarts is Backplane's job: when a process dies or the hub restarts, the thread picks back up.

• Agents (claude, codex, grok) start as plain child processes again. The bp-agent-* scopes, the memory-limit setting, the stale-scope cleanup at startup and the systemctl show diagnosis are all removed (src/core/contain.bend deleted). An agent killed by SIGKILL now reads "agent was killed (signal 9, likely out of memory)" (Exit.why, law exit_why_killed). Lost work is told and resumed as before.
• work_run jobs run as detached setsid processes, never as bp-<id> units. If a restart cuts one off, it reads as lost, the agent is told and can start it again. work_watch follows a pid or a file; the systemd unit kind is gone. Stop and check still verify the process identity first.
• The agents' prompt and the tool descriptions no longer mention systemd or a memory limit.
• The contain_* laws are removed with the code they covered, at the user's request. No other law is touched.

Tests: scripts/check.sh (All terms check), scripts/test.sh 89/89, native build. e2e pass: watch, watch_scripts, restart_work, restart, spawn_gen, exit, wake, retire, durable, prep, live, stall. far_e2e and agents_far_e2e fail at "they pair by themselves", and fail the same way on a build of unmodified main, so the failure predates this change.