Keep a stop the user pressed when a restart cuts the turn off
A turn you stopped just before a restart ends as stopped, not as a failure with a Retry button.
If you pressed stop and the executor died right after, the turn used to end as failed with interrupted_by_restart, and the app offered Retry for a turn you had chosen to stop. Nothing had saved your stop.
Pressing stop now saves a stop: "requested" mark on the turn first. After a restart, a marked turn is recorded as interrupted with no error, and if the agent still reports it running, the new executor sends stop again once. Real failures, such as an agent that stopped answering or a missing workspace, still show their own reason.
Evidence
No screenshot: Cannot be shown in a screenshot; the stop mark and stop request run inside one executor process, so tests reproduce the crash between them.