#27894 ci(nightly-next): the Ghostty helper takes an idle owned Mac while the app build holds the dedicated builder

NEXT nightly fell 3h25m behind feat-cmux-next on 2026-10-30 (the rule is at most 30 minutes). Two causes, the second fixed here:

1. A Release compile red from about 23:40Z to 01:23Z: JoinConsent+App.swift:222:29: error: 'async' call in a function that does not support concurrency failed build-nightly-app for every promoted tip from 43303f771ee3 to 00c873b578db. #27758 (cec0a36b8ac6) fixed it.
2. The first green tip after that, 373ec0cfaa2f (run 40990112293), built its app on the dedicated builder by 02:30Z. Its Ghostty CLI helper then sat queued for 39 minutes on blacksmith-6vcpu-macos-26, and the run was cancelled unpublished at 02:51Z. The helper reuse lookup had missed (the repo GITHUB_TOKEN got a 403 from the contents API), and the helper never takes the dedicated builder, whose hook admits only build-nightly-app. So the helper's runs-on fell to blacksmith, even though owned side Macs were idle.

The change: pick-nightly-build-runner now also outputs helper.

• helper is owned when owned is not the dedicated builder.
• When the app build has the dedicated builder, helper comes from a fresh owned_route() read: an idle side or nightly-slot Mac. The verify jobs already pick their Mac this way.
• With no idle Mac, helper is empty and the helper takes the moving Blacksmith label, as before.

build-nightly-ghostty-cli-helper now reads helper instead of owned, on the same rule: the pick must be for this attempt, and it is never the dedicated builder. The job output stands on the same terms as owned (marker uploaded outside attempt 1 of a push).

Risk: low. The change only affects routing of the helper job on nightly-next. The fallback is still Blacksmith macOS 26, and the dedicated builder is still excluded.

Tests (run on a Linux lane VM):

• tests/test_nightly_build_runner.py: 33 tests OK, including a new test, the helper takes an idle owned Mac while the app build holds the builder.
• tests/test_nightly_trusted_routing.py: 22 tests OK.
• tests/test_ci_nightly_off_blacksmith.py: 15 tests OK.
• tests/test_nightly_next_train.py: 38 tests OK.
• tests/test_nightly_sign_runner.py: OK.
• tests/test_nightly_universal_build.sh: PASS.
• test_nightly_next_guard: PASS.
• test_nightly_workflow_size: PASS.
• test_ci_workflow_guards_are_wired: PASS.

tests/test_ci_signing_owned_mac.py fails one test (test_every_listed_hook_is_one_glaeda_released_with_the_check) on base too.

Agent session: agent session

🤖 Generated with Claude Code

---

^{Need help on this PR? Tag @codesmith-bot with what you need.}

---

Summary by cubic

Fixes a nightly-next CI routing gap where the Ghostty CLI helper queued on Blacksmith while idle owned Macs were available, leaving nightly-next 3h25m behind feat-cmux-next after a Release compile red.

Because the dedicated builder admits only build-nightly-app, the helper previously fell back to the Blacksmith label whenever the app build held the builder, even with idle side Macs. pick-nightly-build-runner now also outputs helper: owned when that isn't the dedicated builder, otherwise a fresh owned-route read for an idle side or nightly-slot Mac, or empty so the helper keeps taking Blacksmith when no Mac is idle. build-nightly-ghostty-cli-helper reads the new output on the same marker terms as owned.