Add oxfmt formatter with auto-format PR workflow

[darylcecile]

Now that the project is open source, contributor PRs have been silently reformatting hand-written code (see #2, where the bulk of the diff was collapsed object literals being re-expanded to multi-line and reordered imports). This change locks in the existing style with a fast formatter and makes contributors' lives easier by running it for them in CI.

Approach

• .oxfmtrc.json — oxfmt@0.52 config tuned to the existing codebase: 2-space indent, double quotes, semicolons, trailing commas, printWidth: 100. Two non-default options matter:
  • objectWrap: "collapse" — defeats Prettier-style magic-trailing-comma expansion, which was the dominant noise in feat(vm): add maxSteps execution budget #2.
  • sortImports: true — keeps import statement order canonical.
• package.json — adds bun run format and bun run format:check; adds oxfmt as a devDependency.
• AGENTS.md — documents the new commands and adds a "don't hand-reformat" rule for agents.
• .github/workflows/format-pr.yml — runs bun run format on every PR and pushes the result back as a github-actions[bot] commit, so contributors never need to think about it.

How the rules were validated

I diffed the base and head of #2, ran our formatter on both sides, then compared. Drop in noise per file:

File	raw diff lines	remaining after format
src/boundary.ts	334	2
test/interpreter-evaluator.test.ts	507	4
src/vm.ts	715	208 (real PR content)
test/vm.test.ts	358	80 (real PR content)

The remaining diffs are genuine semantic changes, not formatting.

Things to know before merging

• One-time normalization needed. The codebase was never formatter-managed. bun run format:check on main will fail today (28 files need touching). Recommend running bun run format on main as a single dedicated commit before this CI starts gating PRs, so future PRs only show real changes.
• Fork PRs are not auto-formatted. The workflow is gated on head.repo == github.repository. Fork PRs only get a read-only GITHUB_TOKEN, so the push step would fail with 403. The alternatives (pull_request_target with PR-head checkout, or a PAT) are known pwn-request vectors and I did not want to opt into them silently. Fork contributors run bun run format locally. Happy to layer a format:check job on top so they get a failing status — say the word.
• No infinite loop risk. Pushes by GITHUB_TOKEN do not trigger workflow runs (docs).
• One small style drift oxfmt can't enforce today: sorting named specifiers within an import statement (e.g. { VMError, VMErrorCode, createCapability } vs { createCapability, VMError, VMErrorCode }). There's no oxfmt option for it; would need an ESLint rule layered on top.