Skip to content

CI: macOS harness failures and timeouts block dependency PR 1125 #1126

Description

@InauguralPhysicist

PR #1125 remains unready: both macOS jobs report shell/test-output portability failures before timing out, while the extensions job reaches its limit during a third variant suite. The dependency diff only updates pinned Docker/Pages action versions.

Read-only follow-up for PR #1125, head 2ce7802ed07138efe0771351173bd95f07c79fbb, CI run 34329218345 attempt 2.

The direct Actions job-log API recovered all three logs. Each job annotation confirms the 30-minute execution limit, but the logs show more than a timeout-only problem.

Observed failures before timeout on both macOS jobs:

  • Replay selftest returned 1, with 5/6 controls passing. Its clean-boundary positive returned 0, but the count line contained padded 0 ledgered instead of the expected 0 ledgered.
  • Strict differential failed with line 197: BASHPID: unbound variable. Both the matcher selftest and subsequent run report an untrustworthy section verdict.

These failures do not establish a runtime semantic regression or an action-version cause.

Where the jobs stopped:

  • macOS latest, job 103459661817: [99d] passed at 00:00:53Z; [99e] SIGUSR1 observer dump began immediately afterward, with no further suite output until cancellation 00:24:41Z. Log lines 2278–2279 show that interval. Cleanup line 2314 names an orphan eigenscript PID 73248.
  • macOS Intel, job 103459661756: [99d] passed 00:04:07Z; [99e] then remained the last observed section until 00:24:37Z cancellation (lines 2345–2346). Cleanup line 2381 names eigenscript PID 88750. Replay/strict errors are visible around lines 2316 and 2331.
  • Extensions, job 103459661842: http+model and gfx full-suite steps succeeded. The zlib suite began 00:20:18Z and reached [69] Leak Guard immediately after [71] teardown passed. Lines 5242–5247 show the last PASS, section header and cancellation 00:24:23Z. Net suite, LSP/DAP checks and fuzz smoke were skipped.

Times are UTC 2026-09-12. The macOS logs reveal process names/PIDs, not actual child argv or stacks. The extensions log supplies no active child identity. Thus [99e] and [69] are last observed work, not fully attributed hang diagnoses. Extensions also had already consumed most of its shared 30-minute budget before zlib began.

No rerun, workflow change, source change or timeout increase was made in this investigation. Raising the deadline alone is not an established fix: macOS already has hard failures, and its silent [99e] interval differs from the extensions job's cumulative budget. PR #1125 remains unready.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions