diagnose-runlisted

The tracking comment looks identical in success and death, and `gh run list --limit 1` happily hands you the wrong run. Resolve the run from the task (`gh run list` matched on the task ref), then read **jobs → failing step → that step's log**. ⚠️ **Read runs over REST**: `gh api repos/{o}/{r}/actions/runs/{id}` and `.../runs?per_page=N` bill the `core` pool, while `gh run view --json` bills **GraphQL** — the pool a driving session drains first, and the one that stops a drive when it empties (#93). ## The fingerprints Match before theorizing — each of these was measured, and each points at a different fix: - **Dies in ~3s, no result payload, before the model** — a setup failure. The two commonest: the actor guard (*"Workflow initiated by non-human actor"* — a bot actor the stub's `allowed_bots` does not admit; on a rename, a stale slug) and a missing base branch (404 in `setupBranch` — the story's named ref does not exist yet). - **Run reports success, task did nothing** — the dead-run shape. `num_turns` small, seconds of model time: check `trigger_phrase` extraction (a handle swallowed as a slash command) and the backgrounding stall (a tidy checklist, boxes unticked). - **Model step failed at env validation in seconds** — the OAuth token is missing or lapsed. Everything scripted still ran; only the model half is dark. - **Every step green, next task never started** — read the dispatch step's notice: dark cascade (declared off — the driver labels by hand) is