chore(runner): the handsontable test discipline, ported and enforced (DEV-2203)

[danielzytohoc]

Stacked on #184 (its ci.yml is the DAG this extends).

The handsontable monorepo's test-writing skillset, ported into this repo and machine-enforced. Adapted, not copied: the runner's stack is Playwright + node --test — no Jest, no Jasmine, no frozen-suite migration story — so the rules are re-expressed in this repo's own conventions and named from this repo's own history.

What was ported, from where

Here	Ported from (monorepo)
runner/docs/TESTING.md	.claude/skills/test-writing-discipline/SKILL.md (the discipline core, verbatim in spirit) + handsontable/.ai/TESTING.md (the which-test-where decision rules)
.claude/skills/runner-test-discipline/	test-writing-discipline — intent-first, red-means-code-wrong-by-default, banned fake-green moves, no hollow assertions, never mock the unit under test, verify with a real run
.claude/skills/runner-playwright-e2e/	handsontable-playwright-e2e — rewired to this repo's conventions: e2e/helpers.ts instead of page objects, data-* test contracts, the env-gate taxonomy, CodeMirror/Sandpack gotchas, container-pool hygiene
runner/scripts/check-test-presence.mjs + the presence CI job	.github/scripts/test-presence-gate.mjs + .ai/LOCAL-ENFORCEMENT.md (the presence gate and its escape semantics)

TESTING.md also codifies what already lived scattered in AGENTS.md and the specs — the env gates (E2E_LIVE / E2E_BASE_URL / E2E_BROKER_TOKEN / E2E_AI / E2E_STARTER_MATRIX, the latter three landing with #186/#187/#189), the house idioms (getComputedStyle not screenshots, prove-by-network-count, toPass around debounces, never read long files via .cm-content, trackSessions cleanup), and the anti-patterns with names: the hollow isMcpCreated test that asserted a copy of itself (#201), the nine days of orphaned E2E_LIVE specs no workflow ran, the failure artifact that had been empty since the workflow was added.

The presence gate

A PR that changes runner/{apps,packages,workers}/**/*.{ts,tsx} (excluding *.d.ts, *.generated.ts, public/) must also change a runner/e2e/*.spec.ts or runner/pipeline/*.test.mjs — any status: added, edited, or deleted.

Escape hatch: a trailer on any commit in the PR range, non-empty reason required. Refactor-only: <reason> — pure refactor, no behavior change. Test-plan: <reason> — the test lands in a named follow-up, or an existing spec already proves it. The gate trusts the trailer; the reviewer verifies it.

The job is PR-only (if: github.event_name == 'pull_request'): a master push arrives via workflow_call after its PR already passed the gate, and has no base to diff against anyway. No base ref resolves → clean skip, never a false block.

Housekeeping

.gitignore's blanket .claude/ becomes .claude/* + !.claude/skills/, so the skills are committed while local agent state stays ignored.

Verified

• Six synthetic gate cases in throwaway commits: src-only fails (exit 1, ::error); src+test, src-only+Refactor-only:, src-only+Test-plan:, docs-only, and *.generated.ts-only all pass.
• ci.yml parses (YAML).
• pnpm test green — 409/409, nothing product-side changed.

A parallel audit is reviewing all existing tests against this ruleset; findings land as their own PRs.

---

Note

Low Risk
Documentation, CI policy, and a git-diff gate only—no runner app, worker, or test runtime behavior changes.

Overview
Ports the handsontable monorepo’s test-writing discipline into the runner and machine-enforces that runner source changes ship matching test changes.

runner/docs/TESTING.md becomes the canonical rules: intent-first discipline, which-test-where table, E2E env-gate taxonomy, house assertion idioms, and named anti-patterns from this repo’s history. AGENTS.md now points there instead of duplicating the full ruleset.

Two committed Claude skills (runner-test-discipline, runner-playwright-e2e) summarize discipline and Playwright mechanics; .gitignore is adjusted so .claude/skills/ can be committed while other .claude state stays ignored.

runner/scripts/check-test-presence.mjs diffs the PR against its base: changes under runner/{apps,packages,workers}/** (with exclusions) require a change to e2e/*.spec.ts or pipeline/*.test.mjs, or a Refactor-only: / Test-plan: commit trailer with a non-empty reason. A new presence job in ci.yml runs this on pull requests only (parallel to unit/build/e2e).

^{Reviewed by Cursor Bugbot for commit 8f2696c. Bugbot is set up for automated code reviews on this repo. Configure here.}