[unohee]

Why

AGT-4664 published failed, rejected, infra_error and rate_limited runs as drafts so their work was visible. On 2026-10-20 that produced 25 PRs on cgf-portal in one night (AGT-4786): 19 said "published because this run ended as infra_error", several were +3/-0 against multi-defect issues, every one was titled with the whole issue and closed it with Closes, and the owner closed all 25 unmerged. An infrastructure failure is not a deliverable.

Cycle 3 of the factory-stability loop; check C5.

What changes

• publishFinishedRun takes only the reviewed path and the operator-park path. publishUnfinishedWork and shouldPublishUnfinishedWork are removed (a constant-false function would only leave a switch someone can flip back on).
• opts.lifecycleFailed / opts.failureSummary are no longer read. The signature is kept because the caller sits at runnerExecution.ts's 1500-line cap; the docblock says so.

Why this does not lose work

I checked this before removing anything, because the old docblock says PR creation is what cleaned the tree:

• preserveWorktree already WIP-commits dirty work onto the swarm branch and keeps the directory (INT-2729), and createWorktree reuses that branch (worktreeManager.ts:143, :571). Observed this morning: three preserved swarm/AX-* branches with wip: preserved partial work commits.
• The failure itself is commented on the issue by the worker audit path (the Worker actions — Halted comments seen on AX-2144), not by this function.

What this does NOT do (stated up front)

The agreed behaviour was push the branch + one comment, no PR. The remote push is not done here. The secret-leak guard that must precede any push (assertNoSensitiveDataOnBranch, AGT-4188) lives inside commitAndCreatePRWithHead (worktreeManager.ts: 1484 of 1500 lines, five callers destructure prUrl). Pushing through a second code path would bypass that guard; reusing it needs an extraction, which is its own change. Until then a failed run's work stays on a local branch.

Verified

• Existing tests changed: only those that asserted "an unfinished run opens a draft" (the shouldPublishUnfinishedWork / publishUnfinishedWork blocks, the draft-body test, and the failed-run publication test, now it.each over all four statuses asserting no PR and an untouched result). The tests for approved, scope-mismatch park, worker_no_changes park, lifecycle failure and non-publishing statuses are unmodified.
• publishOnPark.test.ts: 46 pass.
• Mutations, each reverted:
  • restore the draft for all four statuses → 9 fail;
  • restore it for infra_error only → 3 fail;
  • break the approved publish → 3 fail (control: this path is untouched by the change).
• Harness C5: observes the real consequence (commitAndCreatePRWithHead never called) rather than a helper a refactor could rename. On origin/main without this change: 4 fail (the unfinished cases) / 2 pass (approved and operator-park controls); with it: 6/6. An earlier version of that check crashed on a vi.mock hoisting error; I caught it because a check that errors cannot be told apart from one that fails, and fixed it before relying on it.
• src/automation + worktreeManager: 1272 pass. One timing flake, worktreeManager.baseFreshness.test.ts › does not promote an existing draft PR…, hit the 30 s limit once under parallel load; it never references this code and passes alone in 6.4–9.7 s on 3 of 3 runs.
• typecheck, lint (12 warnings = baseline).