Benchmark PRs on machines chosen by machine:<ec2-type> labels

[alexey-milovidov]

Summary

• The automatic PR benchmark always ran on the default c6a.4xlarge, with no way to configure it. Some systems cannot work there — e.g. Hyrise (Add Hyrise #883), an in-memory DBMS whose encoded hits table alone likely exceeds the machine's 32 GB.
• A maintainer can now add one or more machine:<ec2-type> labels to a PR (e.g. machine:c6a.metal) to override the default machine — one run per machine. Adding such a label relaunches the benchmark immediately (the labeled trigger is filtered so other labels don't launch anything, and a state guard keeps closed PRs inert).
• machine:all, machine:all-amd and machine:all-arm expand to the machine sets of the daily runs (mirroring benchmark-daily.yml), deduplicated against explicitly labeled machines.
• Labels for every machine type plus the three aliases already exist on the repository.
• Security/cost posture is unchanged: labels require triage access, pull_request_target still takes the workflow from the base branch, and every launch remains gated by the benchmark-approval environment.

Test plan

• YAML validated; the jq label extraction and alias expansion tested for: a single machine, all-arm, all, all-amd + all-arm, all + explicit duplicate (deduplicated), two explicit machines, and no labels (falls back to c6a.4xlarge)
• After merge: add machine:c6a.metal to Add Hyrise #883 and check that the launched run uses c6a.metal

🤖 Generated with Claude Code