ci: CI pipeline and changesets-driven release to npm

[AStaroverov]

Adds the missing GitHub flow: CI on every PR, and an automated release of webactor to npm.

CI — .github/workflows/ci.yml

Four parallel jobs on every PR and push to main:

Job	What it runs
check	pnpm lint, pnpm format:check, pnpm build (includes tsc --noEmit per package)
unit	pnpm test:unit — 130 vitest tests
e2e	pnpm test:e2e — 32 Playwright load tests
devtools	pnpm test:devtools — 32 Playwright panel/extension tests

test-results/ is uploaded as an artifact when a Playwright job fails. Setup (pnpm → Node from .node-version → pnpm install --frozen-lockfile) is factored into a composite action so it isn't repeated four times.

Release — .github/workflows/release.yml

Driven by changesets:

1. A PR lands with a .changeset/*.md describing the bump.
2. On merge to main the workflow opens/updates a chore(release): version packages PR containing the version bump and generated CHANGELOG.md.
3. Merging that PR runs pnpm release → build → changeset publish → npm + git tag + GitHub Release.

Auth is npm trusted publishing over OIDC — no NPM_TOKEN in the repo, and releases get a provenance attestation automatically.

Two defects the first CI run exposed

Both were real, both reproduce locally once you know how — neither was papered over.

Unit tests silently depended on a previous build. tests/worker/worker.mjs is a Node worker fixture, so it imports ../../dist/index.js rather than the TS source. On a machine that had never built the package the worker never came up and three tests failed on empty message arrays. rm -rf packages/webactor/dist && pnpm test:unit reproduces it. The package's test script now builds first, matching what webactor-devtools already did.

channel-storm raced with machine speed. It aborted its mid-flight opens on a setTimeout(..., 5), assuming the open request reaches the supporter within 5ms; on a 2-core runner a channel open averages 43ms, so every abort landed before the request was sent and supportsLostBeforeHandshake came back 0 instead of 100. Rather than relax the assertion, each mid-flight open now carries a unique marker and the supporter fires that abort the moment the request arrives — strictly before its own supportChannel call, on any machine.

Also in here

• MIT LICENSE plus license fields — the README claimed MIT-style usage but no license text existed, so npm would have listed the package as unlicensed.
• webactor-devtools marked private — it is never published; changesets skips it (privatePackages: { version: false }).
• Manual release / release:beta scripts removed from packages/webactor, so publishing can only happen through CI.
• prepack / postpack carry the root README and LICENSE into the package at pack time — with files: ["dist"] the tarball would otherwise ship without either.
• Repository URLs updated from AStaroverov/actorr to AStaroverov/webactor; the repo was renamed and npm's OIDC check validates the real name.
• .node-version (26.5.0), CI + npm badges, and CONTRIBUTING.md documenting the flow.

Before merging

• Configure the trusted publisher on npmjs.com — package webactor → Settings → Trusted publisher → GitHub Actions, org AStaroverov, repo webactor, workflow release.yml, environment empty. Without it the publish step fails with a 404.
• Merging publishes webactor@1.0.0 immediately. npm currently has 0.3.3 as latest, and there are no pending changesets, so the first run on main takes the publish path rather than opening a version PR.

Note for later: changeset publish shells out to pnpm publish, not npm publish. pnpm 10.x handles OIDC fine, but pnpm 11.0.8 has a regression (pnpm/pnpm#11513) — worth checking before bumping packageManager.

Verification

Every command CI runs was run locally on this branch: lint, format:check, build, unit (130 passed, verified from a cold dist/), e2e (32 passed), devtools (32 passed). changeset version was executed for real against a throwaway changeset — it produced a correct bump plus CHANGELOG.md — and reverted. pnpm pack confirmed README and LICENSE land in the tarball and the copies are cleaned up afterwards.

🤖 Generated with Claude Code