A pinned routine's commitless branch is never released on a repo where the name never had a PR

[suleimansh]

🤖 automated

Residue of #1643, found while dogfooding #1649 (write-up: #1649 (comment)).

What #1649 fixed. A triage's Run now now goes through the sweep, so the stale-branch release (releaseStalePinnedBranch, #1293) runs before the start. On the rig that worked: the leftover tf-triage-quick was deleted and the agent proceeded.

What it cannot fix. The release is deliberately conservative (src/stale-branch.ts): it deletes the branch only when its name has PR history and no PR is open. A branch with no PR history at all comes back unproven and is kept — the seam cannot tell "an agent still working toward its handoff" from "a leftover that never had a PR".

The triage routines make that leftover every time they finish:

• triage-quick / triage-consensual write only TODO_AGENTS.md, which lives on tf-data, so they push there directly and commit nothing on their own branch.
• A commitless branch skips the handoff (handoff skipped: the session committed nothing) — no push, no PR.
• The run's worktree is reclaimed, but the branch tf-triage-quick stays, at main, with zero commits.
• The next firing (scheduled or clicked) finds it, the prompt's guard says "branch already exists — triage already pending", and the agent aborts in seconds (~$0.18 each time).

So on any repo where the name has ever had a PR (the rig: PR #6), the loop is self-healing since #1649. On a repo where it never had one — a fresh project, or any project after the triage prompt changed to queue-only (#1644), which is what took the PR away — it jams forever on the second triage, exactly as in #1643, and the only way out is still git branch -D tf-triage-quick by hand. Seen on gemstack's origin on 2026-08-23 (an empty tf-triage-quick pushed by a misfired click, deleted by hand).

Options — this is a data-safety policy, so it wants a human decision rather than a fix-by-default:

1. Count "no commits past main" as stale. Extend the seam: a branch with no PR history and no commits unique to its base holds nothing, so release it. Risk: an agent that started seconds ago and has not committed yet looks identical — mitigated in practice by git branch -D refusing while that agent's worktree has the branch checked out (the in-flight guard triage-quick jams after its first run: a closed PR's pinned branch blocks every later firing #1293 already relies on), but not on the remote copy.
2. Have the triage delete its own branch when it ends commitless. The routine knows it never needed the branch; the daemon's settle path already knows handoffSkip === 'no-commits'. No seam policy change; the leftover never exists. Does not help with branches other tools leave behind.
3. Drop the pinned-branch guard for queue-only triages. Since A triage queues work and never does it (#1641) #1644 the triage changes one file on tf-data; two overlapping triages cannot corrupt each other the way two queue rewrites on one branch could. Then there is nothing to release. Biggest change to the prompts' contract (triage-quick jams after its first run: a closed PR's pinned branch blocks every later firing #1293's reason for the pin was duplicate triage, which A triage queues work and never does it (#1641) #1644's queue-only rule makes cheap rather than harmful).

I lean to 2 as the smallest correct change, with 3 as the question worth asking Rom: whether the pin still buys anything after #1644.