Windows. Heavy runs go in the owner's window. Keep a numbered night plan in memory: order, pass criterion, who acts on the result, and which step drops first when time runs short. Plan it from gh pr list --label waits:<lane>.
Wake-ups. Session crons don't fire while the session is idle. Run orch-wakeat "YYYY-MM-DD HH:MM" in the background; its exit wakes you. A dispatch that must happen even if the session is gone goes in a systemd user timer that refuses outside the window.
One dispatch at a time. A concurrency group without cancel-in-progress keeps one pending run; a second dispatch replaces it. Chain orch-dispatch <workflow> <ref>, which exits once its run has started. orch-idlewait exits when no RESOURCE_JOBS job is queued or running; orch-runwait <run-id>… when the first listed run completes.
Greenlight. Before saying go to a gh pr ready, ask which lanes the head will run: one that needs the resource waits for its window unless a waiver is in both the PR body and the head commit message.
Before a window, run orch-conflicts: a PR that conflicts with main gets no workflow run, so a ready or a dispatch on it does nothing.
A runner on the owner's machine runs as the owner and reaches their real data. Check their backups are fresh before a night of runs there.

Gotchas

A dispatch runs the workflow file of the ref, so an old branch runs the old workflow.
A merge commit drops a waiver that lived only in an earlier commit message; put the marker in the merge commit or the PR body.
Times in UTC with the owner's offset; date -u for stored timestamps.

../orchestrate-sessions/reference/worked-example.md shows one self-hosted Mac runner scheduled this way; the evidence is in ../orchestrate-sessions/reference/incidents.md.

scarce-resource

Scarce resource