I was thinking about this a bit differently. The goal is to make feedback available to the team as close to real time as possible, right in GitHub. A command or label on the PR would trigger bench_runner on dedicated benchmarking machines and post the results back.
Where we host the detailed results feels like a separate question. I’ve also been working on a dashboard for my runners, and I think there’s more to explore there, especially since repeated runs of unmerged PRs are a bit different from nightly tracking of main.