Long shell commands already work. The native shell tool lacks a supported job handle for continuing agent work while a dev server, interactive command, build, or watchdog runs.

## Acceptance

- Start a job with explicit channel/chat/lane ownership; inspect status, readiness, output, and exit outcome. - Supply input where required, poll/wait, and stop the intended job without interrupting unrelated work. - Let the agent continue other work and receive completion/failure information. - Define reconnect, stop/kill, helper restart/update, and disconnected-workspace behavior. Report jobs that must restart; do not imply arbitrary process migration. - Run a real dev server and a build while steering the chat; inspect logs, reconnect, and stop the correct process.

Keep machine-specific process handling in the execution host and durable run/tool state in pi. This is distinct from the existing human terminal UI and from displaying live tool output (#54).

Dogfooding core/QOL: #5. Related: #12, #16, #29. [Review evidence](https://github.com/githubnext/ace2/issues/5#issuecomment-5986485157).