F: publish a finished task's pull request in production

Why

After #91, the runner can execute a task, but nothing opens the task's pull request. PullRequestPublisher (runner/publish.ts) and GitBranchPusher (runner/branch-push.ts, #98, PR #101) are built and tested, but no production code constructs them or calls publish(). #91 part 1 (PR #102) builds the execution deps in runner/production.ts. #91 part 2 adds start and resume. Neither mentions publishing. So a task that finishes its last item stays running and has no PR.

Needed

In runner/production.ts, build PullRequestPublisher with the real gateways: GhAlreadyFixedGateway, the gh pull request gateway, and a GitBranchPusher.
- The pusher pushes from the runner-owned repository and uses the URL from pushUrl (config.github.repository, GH_HOST). Never pass url in production (F2d: push the task head to its branch with a real BranchPusher (#98) #101 review, finding 6).
- ownedCommits(identity) gives the task ledger's origin: 'owned' SHAs. Before Add pre-merge rebase and head-bound command checks #22 adds rebasing, decide whether this set must also include heads that codeboost pushed but the ledger marks as foreign (F2d: push the task head to its branch with a real BranchPusher (#98) #101 review, finding 4).
Publish when a task's run ends:
- All items done, task running, no active attempt: publish(identity) (ready).
- The task moved to needs human: publish(identity, { problems }) (draft).
- Run the publish after the run settles, with no attempt active, because assertPublishable refuses while an attempt is active.
Shutdown: track each in-flight publish like an ItemExecutor run. Abort it and await it before the Store closes (AGENTS.md: track an in-flight irreversible subprocess as part of shutdown). PullRequestPublisher.close() already aborts its runs.
Add an /api/runner action publish to retry a publish that failed or was refused (for example OpeningUnsettled, BranchPushRefused, or a GitHub error). It takes expectedStateVersion and an action ID, like the other runner actions. Show the last publish outcome in GET /api/runner.
Keep publishing off in demo mode, even when a github block exists (AGENTS.md: keep irreversible integrations disabled in demo mode).

Tests

A real Git remote (local bare repository) and fake gh gateways: a task that completes opens one PR, and the branch is at the task head.
A needs-human task opens a draft PR with its problems.
Shutdown during the push or during the PR opening: the Store closes only after the publish has settled, and the next start recovers the opening by its marker.
A refused publish (BranchPushRefused) leaves the task running. The publish action retries it once a person has fixed the branch.

Depends on #91 (PR #102) and #98 (PR #101).

🤖 Generated with Claude Code

Decisions for this issue, made 2026-10-03. Implemented in PR #110.

Base branch: a new github.baseBranch, required when a runner block exists. The CLI checks it before the lock, and setUpRunner checks it before anything is created.
Trigger: the task's status, once the run has settled:
- a running task with every plan item done gets a ready publish;
- a task in needs human, by any path (safety finding, time limit, budget expiry at admission), gets a draft publish. Its problems say what sent it there.
Last publish outcome: durable. Schema v10 adds publish_outcomes, and GET /api/runner shows it as publish.last.
Startup: one background publish if one is owed (a lost opening, or a finished run whose publish did not complete). The CLI starts it after it re-verifies the runner lock.

Defaults I chose:

ownedCommits: the ledger's origin: 'owned' SHAs only. Before Add pre-merge rebase and head-bound command checks #22, foreign entries come only from recordRebase, which production never calls. Add pre-merge rebase and head-bound command checks #22 must revisit this (finding 4 of the F2d: push the task head to its branch with a real BranchPusher (#98) #101 review).
The publish action: runs in the background, and its replay reports the settled outcome.
Upgrade: tasks already running or in needs human at the upgrade are recorded as not published, so the first start opens no PR that nobody asked for.

Follow-ups, deliberately not done in #110:

Cancel during a draft opening: it can leave a cancelled task with an open draft PR. The PR is recorded, and closing it belongs to "close the draft on cancel".
Untracked push on a crash: the production push registers no process group, so a push running when the process crashes is not tracked. A normal shutdown aborts and awaits it.
Commits outside the ledger: the pushed head is snapshot.head, which may contain commits the ledger does not record. This belongs with Add pre-merge rebase and head-bound command checks #22.
Outcome version stamp (low confidence): the state version saved with the outcome is read at record time, not at the publish's last guarded write. No failing interleaving was found.

Follow-up issues for the deferred findings: #111 (close the draft on cancel), #112 (track the push's process group across a crash), #113 (publish only ledger-recorded commits; with #22), #114 (outcome version stamp, low confidence).