fix(ci): keep PR package checks off display runners
Leo: "WHY ARE TEST RUNS EVEN GETTING F***ING STUCK" and "i dont like doing display tests anymore".
PR package tests and generated-file checks can wait on a console/display runner even though display-only suites are already advisory on PRs. Route those jobs through the existing non-display owned-Mac picker and its bounded, stall-aware Blacksmith overflow. Preserve on-demand and non-PR display coverage, runner trust and workflow permissions.
Keep generated-file evidence fail-closed. A cancelled prerequisite's missing verdict cannot be excused by matching the same missing evidence on base; rerun the prerequisite and its dependent reporter together. The companion autolander repair is manaflow-ai/cmuxterm-hq#5740. Incident: #29683, run 41251420343 attempt 2, Swift job 124694215347 was cancelled with no runner or steps and the dependent generated-files reporter failed without evidence.
Prior art: #28936 made display-only skips advisory; #28605 improved display/session overflow. This removes the remaining PR dependency on a display label while preserving those behaviors.
Validation: fleet run 41261540608 tested final head 49f1147bf0605707678adeba099569aca463e89c: 543 of 544 focused tests passed. The sole failure, RunsOnFallbackTests.test_a_picked_signing_label_has_a_rerun_alternative, failed with the exact same assertion on base b52afcec46e48870db26590033880365ada50bf3 in that run. No PR-owned test failures. The pre-push guard needs complete Git history; its shallow-snapshot run incorrectly treated the boundary commit as a root and is being repeated with full history.
The existing owned-pool switch and fork/owner trust clauses are preserved. Current picker allocations permit bounded overflow; stale placement cannot take the unallocated generic MACOS_RUNNER_26 value.
The jobs retain their existing warm sources in .github/warm-sources.json: shared SwiftPM state, Actions caches and same-tree product reuse. This changes runner selection without adding compiles or changing the shared cache. These package tests require macOS; Linux cannot run their AppKit dependencies.
Performance
The latest fleet daily summary predates this change. The baseline includes all events for these job names; no measured savings are claimed before deployment. Existing warm sources remain intact.
| Metric | Before | After | Delta | Harness or source |
|---|---|---|---|---|
| Blacksmith minutes/day, Swift + daemon test jobs | ā2,183.1 on 2026-10-30 | Not yet observed | Not measured | Fleet daily summary, runtime_by_job: 68 Ć 1,887.4s + 2 Ć 1,321.0s |
Agent session: agent session
š¤ Generated with Claude Code