CI: run the full e2e tier nightly, not per push (#1414)

[Mikola Lysenko (mikolalysenko)]

LLM Description written by Claude Code:claude-opus-5-5

Fixes #1414

Root cause

e2e-full, yarn-berry-full and cargo-vex-matrix-full were gated on event != pull_request && event != merge_group. That also matches push, so every landing on main ran all 41 full-tier legs, even though the workflow header calls this tier nightly. On push run 38197830089 (1779317), 41 of the 50 non-skipped jobs were these legs: ~21 wall-minutes, ~41 billed job-minutes.

Change

The three jobs now use the same gate as e2e-docker: schedule, workflow_dispatch (so release-qa(full) still runs and still finds e2e-full*, per docs/release-train/DESIGN.md), the release/v5-prerelease PR, plus push && vars.CI_SCOPE == 'full', so flipping CI_SCOPE=full still brings the tier back on landings. The trade-off: a regression that only this tier catches shows up at the next nightly instead of on the landing.

e2e-build also stops bundling, compressing and uploading the ~94 MB ubuntu e2e binaries on a push that reuses the merge-queue result, unless `CI_SCOPE` is `full`. Those binaries were kept on reused pushes only for the full tier, and nothing else downloads them there.

Also updated: the header and job comments, the `ci-reuse-merge-group.py` log line, docs/testing/yarn-berry-compatibility.md, and the two scheduling tests that pinned the old condition (test_full_jobs_run_nightly_only_and_share_steps, test_unchanged_pr_merge_skips_only_what_the_queue_re_tests). test_reused_push_keeps_cache_writers_and_unqueued_jobs now also pins the new upload guard.

Credit to Austen Stone (@austenstone), who reported this and proposed the same gate in a fork. This is an independent implementation, not a cherry-pick of that commit.

Runner usage

• Lean PR run: unchanged. These jobs never ran on PRs (~28 Linux jobs, ~8 min).
• Merge queue: unchanged.
• Push to main: 50 → 9 non-skipped jobs, about −41 billed job-minutes per landing, plus the e2e-build bundle and upload step time (45 pushes on Oct 10-11 would have saved ~1,845 jobs).
• Nightly / dispatch: unchanged.
• ci-ok needs list and required check names are unchanged; ci-ok already treats skipped jobs as passing. Depot mapping and DISABLE_DEPOT_RUNNERS are untouched.

Validation

• python3 -m pytest scripts/tests -q: 334 passed, 6 skipped.
• Red→green: with origin/main's ci.yml and the updated tests, 4 subtests fail (all three jobs in test_full_jobs_run_nightly_only_and_share_steps, plus yarn-berry-full in the scheduling test). With this ci.yml they pass.
• actionlint 1.7.7: 36 findings before and after, all pre-existing "alias node" complaints about steps: *anchor.
• zizmor --offline: identical (59 findings, 0 medium/high) before and after.
• ci.yml parses with PyYAML. No fromJSON matrix was touched.
• No Rust, cfg(windows) or platform code was touched, so no full-scope dispatch is needed.

/code-review high

The review found no correctness bugs. Fixed: the redundant ubuntu artifact upload on reused pushes, and the stale comments, log line, test comment and doc. Not changed:

• Writing the gate once as a YAML scalar anchor. The repo's workflow tests read `if:` lines textually and couldn't see an alias, and actionlint already rejects aliases here. The gate stays written out on each of the three jobs, and the tests pin it.
• docs/release-train/DESIGN.md primary-candidate selection (a green `push` run is accepted). A reused push already skipped every test job and `e2e-docker`, so a green push run was never full-tier evidence. `release-qa(full)` (dispatch) still runs the full tier and is the gate. The release train is release tooling, which this lane doesn't edit.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Cxz4F7nBdZPMSx1tya1sBt

---

Generated by Claude Code

[Mikola Lysenko (mikolalysenko)]

BugBot review

---

Generated by Claude Code

[cursor]

✅ Bugbot reviewed your changes and found no new issues!

Comment @cursor review or bugbot run to trigger another review on this PR

^{Reviewed by Cursor Bugbot for commit fec483f. Configure here.}

[Mikola Lysenko (mikolalysenko)]

Ready for review (burn-down agent)

• Head: fec483f
• CI: 37/37 non-skipped checks green (50 skipped by path filters); mergeable, no conflicts.
• Bugbot: reviewed fec483f, no findings; no open review threads.
• Reviewer focus: the three full-tier e2e jobs now run on schedule / workflow_dispatch / CI_SCOPE=full pushes only, so a regression only that tier catches surfaces at the next nightly rather than on landing (CI perf: full tier (41 legs) runs on every push to main, not just nightly #1414).

---

Generated by Claude Code

[Mikola Lysenko (mikolalysenko)]

Final review brief

What it does. The three full-tier jobs (e2e-full, yarn-berry-full, cargo-vex-matrix-full, 41 legs) stop running on every push to main and run only nightly, on workflow_dispatch (so release-qa(full) still gets them), on the v5 integration PR, and on main pushes when vars.CI_SCOPE == 'full'. e2e-build also stops uploading the ~94 MB ubuntu e2e bundle on reused pushes unless CI_SCOPE is full, since only the full tier consumed it there.

Risk: low. CI config only; no Rust or runtime code. PR and merge-queue tiers are unchanged and ci-ok already treats skipped jobs as passing. The trade-off is intentional: a regression only the full tier catches surfaces at the next nightly instead of on the landing.

Look here

• ci.yml:2037, :2291, :2431: the new gate on the three jobs
• ci.yml:1019-1037: bundle/compress/upload now skipped on reused pushes unless CI_SCOPE=full
• test_ci_scheduling.py:123 and test_ci_e2e_tiers.py:62: tests pinning both

Verified

• Read the full diff. Checked that nothing else downloads e2e-bin-ubuntu-latest on a reused push (e2e, e2e-extended and cargo-vex-matrix already skip on reuse; hosted-e2e builds its own), and that schedule/dispatch never reuse, so they still bundle.
• python3 -m unittest discover -s scripts/tests: 339 run, OK (7 skipped). ci.yml parses. CHANGELOG.md untouched.
• CI on fec483f: ci-ok and clippy green, no failing checks; Bugbot clean; no open review threads.

Changes I made: none.

Open questions: none blocking. Optional: decide whether a nightly-only full tier is acceptable for release cadence (the PR author's stated trade-off).

Auto-merge is armed, so approving sends it straight to the merge queue.

---

Generated by Claude Code