CI perf: Gradle patch compatibility — PR runs repeat ci.yml's Gradle e2e tier and run nightly-only extras (~17,000 Linux job-min/day) #1177

Measurement

Window: 2026-10-01 → 2026-10-08. Runs: 453 Gradle patch compatibility pull_request runs in 7d (431 that weren't skipped, ~62/day; 185 in the last 24h across 50 PRs). Jobs sampled: 67 PR runs (40 successful).

• Successful PR run: 536 job-min p50 (e.g. https://github.com/SocketDev/socket-patch/actions/runs/37770694851), of which
  • 12 ubuntu cells (4 Gradle lines × agent/hosted/vendor): 144 job-min/run
  • 13 ubuntu extras (JDK ceilings, configuration-cache, isolated-projects, real-central): 160 job-min/run
  • 12 windows cells: 204 job-min/run; build ubuntu+windows: the rest.
• Mean over all sampled PR runs, cancelled included: 488 job-min/run.
• Estimated 7-day average: Gradle compat = 15,400 Linux + 10,400 Windows job-min/day. Its "Run the hosted suites" step is the Remove tsc from npm prepare script #1 step in the repo by total minutes (~20,000 job-min/day).
• Linux queue wait last 24h: p50 1.2 min, p90 12.8 min, worst hour (2026-10-07T16) p50 15.7 min. This hits ci.yml's Gradle e2e legs, which were the merge_group critical path in 16 of 21 sampled successful merge_group runs.
• Failure rate on PRs: 10 failures in 431 runs (2.3%); 3 of them on feat/gradle-support.

Where the time goes

Root cause

1. Exact duplication on PRs. ci.yml's e2e job already runs, on every PR and merge_group, the ubuntu PR tier: 4 Gradle lines (6.9.4/jdk11, 7.6.6/jdk17, 8.14.3/jdk21, 9.8.0/jdk21) × {agent suites + all gradle_hosted_ (split across 3 legs since Cut merge-group CI from ~46 to ~20 min: shard Gradle e2e and test legs, skip test-release in queue, cancel orphaned runs #1133), gradle_vendor_ + gradle_multi_project} (ci.yml:1240-1255). The ubuntu cells of gradle-compatibility.yml run the same suites, filters, Gradle versions and JDKs on the same OS.
2. Grid extras on every PR. The JDK-ceiling / configuration-cache / isolated-projects (record-only) / real-Central rows are the "rows the PR tier never runs" per the workflow header, but they still trigger on every PR touching the broad paths: list (Cargo.toml, Cargo.lock, hosted/**, vex/**, commands/apply.rs, commands/scan/**, tests/common/**, …).

Proposed fix

In .github/workflows/gradle-compatibility.yml:

1. Exclude ubuntu-latest from cells on pull_request, the same way macOS is already excluded:
   exclude: - os: ${{ github.event_name == 'pull_request' && 'ubuntu-latest' || '' }} (keep the macOS exclusion as a second entry). The build ubuntu-latest job is then needed on PRs only when extras run.
2. Gate extras to schedule / workflow_dispatch, plus PRs that touch the Gradle core (crates/socket-patch-core/src/gradle/**, patch/redirect/*gradle*, vendor/jvm/**, this workflow). E.g. a small changes job with dorny/paths-filter (pinned SHA) or a git diff --name-only step, and if: github.event_name != 'pull_request' || needs.changes.outputs.gradle_core == 'true' on extras and build (ubuntu).
3. Windows cells stay on PRs: they're the only PR coverage of Gradle on Windows besides ci.yml's single e2e_vendor_jvm_build windows leg.

Expected saving

• Ubuntu cells + extras = 304 of 536 job-min per successful run (57%). At ~62 PR runs/day × 488 job-min mean × 0.57 ≈ ~17,000 Linux job-min/day (7d average; the last 24h ran ~3× the 7d rate, so the 24h saving is higher). Minus extras on Gradle-core PRs (a minority).
• macOS: 0. Windows: 0.
• Critical path: none directly; indirectly it lowers Linux runner contention (p90 queue 12.8 min) that delays ci.yml's e2e Gradle legs on PRs and merge_group.

Coverage and risk

• Ubuntu cells: the same suites/versions/JDKs still run on every PR and merge_group via ci.yml e2e, and in this workflow nightly (17 4 * * *). One small difference: the cells' probe-report artifact names (gradle-probe-ubuntu-…) won't be produced on PRs; ci.yml uploads gradle-probe-pr-ubuntu-latest-* from the same tests.
• Extras: still run nightly, on dispatch, and on PRs that change Gradle core code. A regression that only shows on a JDK ceiling and comes from non-Gradle code would surface in the nightly instead of on the PR (≤24h later).
• Not a required check; ci-ok / clippy unaffected.

Effort

S (an exclude entry + a paths gate on one job).

ROI

17,000 Linux job-min/day × 0.5 (Linux weight) × confidence 0.7 / effort 1 ≈ 6000.