Run benchmark workflows on pull_request instead of pull_request_target

[tautschnig]

Since 2026-07-20, the label-gated benchmark jobs (End-to-End Benchmarks / perf-benchcomp, Compiler Benchmarks / compile-timer-{short,long}) fail within seconds on every pull request from a fork, while pushes to main stay green.

Root cause: actions/checkout v6.1.0 was published on 2026-07-20 (with coordinated backports v5.1.0 and v4.4.0) and made the floating v6 tag refuse to check out fork PR code from workflows running in the trusted pull_request_target context (changelog):

##[error]Refusing to check out fork pull request code from a 'pull_request_target' workflow. [...]

Our benchmark workflows are invoked via workflow_call from extra_jobs.yml, which runs on pull_request_target (the auto-labeler needs pull-requests: write). They check out github.event.pull_request.head.sha — fork code — and then build and execute it, which is precisely the pattern the new default blocks. Pushes to main run the same workflows via the push trigger comparing trusted before/after refs, hence unaffected.

Fix: rather than opting out via allow-unsafe-pr-checkout: true, move the benchmark invocations to a new bench-pr.yml workflow triggered by plain pull_request events, so fork code builds and runs in the untrusted, fork-scoped context as GitHub recommends. Label gating (Z-EndToEndBenchCI / Z-CompilerBenchCI) is preserved, now reading the labels from the event payload instead of the labeler job's outputs. The event-name conditions inside the reusable workflows are simplified accordingly, and extra_jobs.yml keeps only the auto-labeler.

Behavioral difference: labels added by the auto-labeler do not re-trigger workflows (events created with the default GITHUB_TOKEN never do), so on freshly opened PRs the benchmarks only start with the next push, or when someone adds the label manually (human-added labels do trigger via the labeled event type). Subsequent pushes to labeled PRs — the common case — behave exactly as before. A concurrency group cancels superseded benchmark runs.

Validated with actionlint 1.7.7 (clean) on all four touched workflow files.

By submitting this pull request, I confirm that my contribution is made under the terms of the Apache 2.0 and MIT licenses.

[Copilot]

Pull request overview

This PR moves the label-gated benchmark runs off the trusted pull_request_target path (used for auto-labeling) and onto a dedicated pull_request-triggered workflow, aligning with the new actions/checkout safety behavior and avoiding running fork PR code in a trusted context.

Changes:

• Remove benchmark reusable-workflow invocations from extra_jobs.yml, leaving only the minimal auto-labeler on pull_request_target.
• Add bench-pr.yml to run the benchmark reusable workflows on pull_request events when the corresponding Z-*BenchCI label is present.
• Simplify PR-ref selection logic in the benchmark reusable workflows (but see below: the current condition breaks workflow_call use).

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 3 comments.

File	Description
.github/workflows/extra_jobs.yml	Stops invoking benchmark workflows from the trusted pull_request_target labeler workflow.
.github/workflows/bench-pr.yml	New pull_request workflow that label-gates and calls the reusable benchmark workflows with concurrency cancellation.
.github/workflows/bench-e2e.yml	Updates PR ref-selection condition for setting OLD_REF/NEW_REF (currently problematic under workflow_call).
.github/workflows/bench-compiler.yml	Same ref-selection condition update in both compiler benchmark jobs (currently problematic under workflow_call).

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[tautschnig]

We need to get the changes in #4666 merged first to unblock the CI jobs.

[feliperodri]

@tautschnig I just approved #4666 and got it into the merge queue.