ci: indexing bench on pull requests, base vs head on the same runner

[Chofito]

Closes #8

What

• bun run bench accepts reference repository names as well as paths. A name resolves through scripts/repos.ts (from feat(eval): A/B eval runner (Claude Code with vs without Astrograph) #12) to a fresh checkout at its pinned commit, so no one has to point it at a clone by hand.
• Two options: --baseline <file> (default .bench-baseline.json, still gitignored) and --bin <bin.ts> (the astrograph entry point to measure; default this checkout).
• New bench job in CI, on pull requests only: indexes fastify, umami, trpc, koel (JS/CJS, Next.js app, TS monorepo, PHP/Laravel) with the base branch (git worktree at $GITHUB_BASE_REF), then with the PR, and fails on >20% slower or bigger. Clones are cached by the hash of eval/repos/*.json.
• AGENTS.md: the budget rule now points at the CI job.

Why not a committed baseline (as the issue suggested)

GitHub's shared runners change hardware between runs, so a committed number would drift with the machine, not the code. Measuring base and head on the same runner, minutes apart, compares code. No baseline file to keep up to date.

t3code is a reference repository for the eval but is left out here: it indexes 23k files (it vendors two upstream repos under .repos/) in ~38 s / ~1 GB, too heavy for every PR.

Verified

• bun test, bun run typecheck, bun run check green.
• Locally: bench --save --baseline … --bin <other checkout> then bench --baseline … on the four names → all ok (fastify 673 ms / 344 MB, umami 1045 ms / 416 MB, trpc 980 ms / 426 MB, koel 1303 ms / 400 MB on an M-series Mac).
• This PR's own bench job is the end-to-end test of the workflow.

🤖 Generated with Claude Code