fix: keep work stopped under another supervisor token from returning after a Gateway crash (#201186) * fix(gateway): scope crash and child-completion restart recovery to the supervisor token OPENCLAW_GATEWAY_RESTART_RESUME_SCOPE scoped graceful-shutdown markers only. Work without a marker still came back under another token: a run orphaned by a crash, a parked yield, a legacy upgrade cycle, the queued input replayed after a crash, and a child completion that wakes its parent. Each accepted run now records the token it started under (at chat.send admission and at lifecycle start). Startup compares the marker's token, or the recorded one without a marker, before crash, parked and legacy recovery and before taking back completion custody. Out-of-scope work is cleared without resuming; its queued input ends as restart-unrecoverable, and with a token set a session with no recorded token counts as out of scope. Recovered markers carry the token, so a second restart under it still resumes them. Without a token nothing changes. * test: run the restart resume scope test in the database worker pool * test: set the restart resume scope before the first database write * style: format the restart resume scope test





