[igerber]

Summary

• GitHub rolled ubuntu-latest to Ubuntu 26.04 on 2026-10-24 (the docs(v4): PR 4b of the 4.0 cut - rebuild tutorials 02/04/05/07/23 on the TWFE event-study mode #867 review ran on ubuntu-24.04 on 10-23; the refactor(v4): PR 4c-i of the 4.0 cut - decouple the TWFE event-study mode from the MultiPeriodDiD containers and migrate the MPD-fixture suites #868 review ran on ubuntu-26.04 on 10-24 with the same workflow and action pin). Two jobs break on the new image: openai/codex-action's drop-sudo safety strategy fails before codex starts (runner is not a member of the sudo group / Could not restrict access to /run/dbus/system_bus_socket; upstream drop-sudo fails on Ubuntu 26.04 openai/codex-action#188, open since 2026-10-20), so the AI PR review job on refactor(v4): PR 4c-i of the 4.0 cut - decouple the TWFE event-study mode from the MultiPeriodDiD containers and migrate the MPD-fixture suites #868 died in under a second and posted nothing; and actions/python-versions has no Python 3.9 build for 26.04, so the 3.9 floor leg of the test matrix and the 3.9 wheel-test leg cannot set up.
• Replace ubuntu-latest with ubuntu-24.04 in all ten active workflows (ai_pr_review, build-wheels, ci-gate, docs-tests, lint, lwdid-data-canary, mmm-interop, notebooks, publish, rust-test) - the image every run used until yesterday. ubuntu-24.04-arm entries are untouched; claude-code-review.yml.disabled is left alone.
• docs/dev-status.md: a section recording the pin, the two breakages, and the un-pin conditions.

Methodology references (required if estimator / math changes)

• Method name(s): N/A - no estimator or math changes.
• Paper / source link(s): N/A
• Any intentional deviations from the source (and why): None

Validation

• Tests added/updated: No test changes. All ten workflow files parse as YAML; tests/test_doc_deps_integrity.py, tests/test_notebook_md_extract.py, tests/test_tracking_files.py, tests/test_openai_review.py (incl. TestLintWorkflowPinSync) and tests/test_changelog_fragments.py pass (677). The pull_request run of this PR itself exercises the pinned review job.
• Backtest / simulation / notebook evidence (if applicable): N/A

Security / privacy

• Confirm no secrets/PII in this PR: Yes

Changelog

• changelog.d/ fragment added (or N/A - no user-visible change): N/A - CI configuration only, no diff_diff/ change.

Follow-ups

• Re-trigger the review on refactor(v4): PR 4c-i of the 4.0 cut - decouple the TWFE event-study mode from the MultiPeriodDiD containers and migrate the MPD-fixture suites #868 once this lands on main (its pull_request review run already failed).
• Un-pin to ubuntu-latest only after codex-action#188 is fixed and a 3.9 build exists for 26.04, verified on a throwaway branch.

[github-actions]

✅ Looks good

No P0–P3 findings identified in the supplied diff.

Executive summary

• All active Linux runner references consistently use ubuntu-24.04 or ubuntu-24.04-arm.
• The Python 3.9 matrix entry is updated correctly; ARM entries remain unchanged.
• Changes affect CI configuration and developer documentation, with no estimator or inference changes.

Methodology

No findings. The Methodology Registry was consulted; no affected methods, equations, assumptions, weighting, or variance calculations require paper-level verification.

Code Quality

No findings. Direct runner settings and matrix entries are updated consistently. No conditions still depend on ubuntu-latest.

Performance

No findings. The manylinux build containers remain unchanged.

Maintainability

No findings. docs/dev-status.md:L159-L175 records the compatibility issues and concrete conditions for removing the pin.

Tech Debt

No findings. The temporary compatibility constraint and its resolution criteria are documented.

Security

No findings. The diff introduces no changes to permissions, action revisions, secrets handling, or sandbox settings.

Documentation/Tests

No findings. CI-only changes do not require a user-facing changelog fragment.

Validation confirmed that all 11 active workflows parse as YAML, all 23 Linux runner references use the intended labels, and both lint-pin consistency checks pass when executed directly. The full pytest suites could not run because pytest and scientific dependencies are absent from this environment.