GitHub's hosted runners differ from run to run (region, CPU), so a commit's
timings only compare with timings from the same runner.
- The perf job builds the commit and master before the push and loads both
alternately (`yarn stats:compare --base <commit> --iterations 6`), instead of
waiting for staging; without staging there's nothing to cancel, so every merge
gets its run.
- `perf.mjs --base` takes a commit: checked out into a temporary worktree with its
own dependencies, built and served like this checkout. `--out *.json` writes both
builds' loads; `--wait-for` is gone. One unrecorded warm-up load per build, and
the runner's CPU in the environment.
- `post.mjs` sends the commit time (the admin orders runs by it) and the base
build; perf runs need one.
Claude Opus 5.5 (xhigh)
- `Stats` workflow on master: posts the production build's file sizes (raw, gzip, Brotli; initial load vs. lazy; bytes per source package) and, once staging serves the commit, its page-load and pan/zoom measurements to the Excalidraw+ backend; OSS_STATS_TOKEN lives in the master-only `stats` environment
- `scripts/stats/`: bundle-stats Vite plugin and wrapper config (`yarn stats:bundle`), the Playwright runner (`yarn stats:perf`) and the poster
- `yarn stats:compare`: builds the checkout, serves it with `vite preview` and compares it with staging (alternating loads, Mann-Whitney U test); prints a table or writes CSV with `--out`
- `excalidraw:app-start` mark in the app's entry, ending the runs' "HTML + JS" phase
- `playwright` 1.58.2 dev dependency (installs no browsers by itself)
Claude Opus 5.5 (xhigh)