sessions: 🐛 Let the root stop a done dispatch in its own Orca Run

[yasyf]

Context: At 2:07 AM on October 6, 2026, the release-v3 drive root ran orca orchestration worker-stop --dispatch ctx_a0f447e7e42c to end hsbc-routing-check, a lane it had already judged done. orca_ends_session blocked the call with "worker-stop ends the worker or run … leave ending them to the owner." The owner then ruled in cc-notes answer edfa70e that the root stops a done lane itself instead of messaging it to send worker_done, and asked for this hook to be fixed.

Summary: orca_ends_session allows one literal orca orchestration worker-stop --dispatch <ctx> when the caller is the main session whose ORCA_TERMINAL_HANDLE is the coordinator of that dispatch's Run. The guard reads the Run from orca orchestration worker-show and the coordinator from run-show, using the coordinates() check the settled-close lift already relies on. Every other case stays blocked: lane terminals, in-process subagents, dispatches in another Run, an unreadable dispatch, loop-respelled or conditional calls, run-stop, coordinator-stop, and worker-release. The worker-stop block message names the allowed route.

Motivation: The root decides when a lane is done: stood down, superseded, or its deliverable on disk. Making it signal the lane to report completion first adds a round trip and gains no safety, because the root already owns the Run. pkill, signals, and terminal closes are untouched.

Details: closed_terminal and the new stopped_dispatch share lone_target, which accepts exactly one literal, non-expandable target. spelled() rejects calls the scan respelled from a loop, so for d in …; do worker-stop --dispatch $d still blocks. The inline tests pin the 2:07 AM refusal as an Allow for the coordinating root, plus Blocks for a lane terminal, a subagent, another Run, a missing worker-show, a loop, a && chain, run-stop, and worker-release. The packs guide and the changelog record the narrowed rule.