ci: gate pull requests with the existing test scripts

[yashranaway]

Closes backlog §D1 and §D3.

Until now the only automation was the tag-triggered release workflow, so every pull request — #4 through #9 — merged with no automated checks at all. This adds .github/workflows/ci.yml.

Design

Every job invokes the same script a developer runs locally, the property release.yml already has. CI can't drift into a separate build path.

Job	Runs	Notes
static	shell syntax (sh/bash/zsh), docs/qa/evidence checksums, whitespace vs base	evidence is the only binary media in git — prove it still matches its manifest
runtime	pnpm test:runtime	jsdom context-pruning suite; fastest signal
protocol	apps/headless/test.sh in swift:6.1-bookworm	the 27-case protocol/security suite without Swift on the runner; same image Dockerfile.linux builds with
web	eslint + next build	the site can break invisibly today
linux-e2e	Tests/linux-docker.sh	full sandboxed-Chromium E2E; uploads the QA evidence bundle as an artifact
macos	build.sh + test.sh	native compile and suite
macos-e2e	Tests/macos-e2e.sh	not per-PR: nightly cron, workflow_dispatch, or the macos-e2e label, because it needs a GUI session and mutates com.headless.app user defaults

Triggers: pull requests, pushes to main, nightly cron, manual dispatch. Concurrency group cancels superseded runs.

Verified locally

• YAML parses; the shell-syntax loop passes on all 12 sh/bash scripts (zsh ones are checked on the runner, where zsh is installed)
• sha256sum -c on the committed QA evidence passes
• pnpm test:runtime passes, reproducing the documented pruning numbers (48,428 → 895 bytes)
• pnpm install --frozen-lockfile works for both workspace filters
• web lint clean, next build succeeds (6/6 static pages)

Docker isn't reachable from my sandbox, so linux-e2e rests on the precedent that release.yml already runs linux-docker.sh on ubuntu-latest — this PR's own run is the real proof.

Docs updated in the same commit

AGENTS.md describes the gate, the roadmap marks the Phase 1 CI item done, and backlog §D1/§D3 record what shipped plus what remains.

Follow-ups (not in this PR)

• Mark the jobs required in branch protection — a repo setting, not code
• Consider arm64 Linux E2E on PRs (release covers it on tags)
• Revisit making the macOS E2E a per-PR gate once its runtime is measured