[Madraka]

What

render_page's Chrome start no longer fails a busy CI runner after two 20 s attempts. The render tests no longer spend part of their 50 s render budget on a cold Chrome start. A real renderer regression still fails: every render assertion is unchanged, and there is no skip.

Changes

packages/cli/src/render-chrome.ts

• chromeRenderer(executable, options) takes three new options:
  • launches (default 3): how many times to try;
  • backoffMs (default 1 s): the pause before the second launch, doubled before each later one;
  • startMs (default 20 s): how long one launch may take to print its DevTools endpoint.
• Each launch gets a new profile.
• The error names the last failure and the number of launches, for example "Chrome did not start within 20 s (3 launches)".
• A launch that hangs or exits is killed, and its profile is removed. A failed launch clears the shared Chrome only when it is still the current one, so two waiting sessions can't kill each other's relaunch.
• A Chrome that crashes after it started is dropped. The next session launches a new one.
• A new stop() ends the renderer's Chrome.
• The launch and kill code moved into module-level launchChrome and killChrome, which keeps chromeRenderer short.

packages/cli/test/render-chrome.test.ts

• The Chrome tests share one renderer. It starts in beforeAll with a 45 s start budget and a 200 s hook timeout, and afterAll stops it. The sessions use separate browser contexts (cdp.ts), so sharing a Chrome is safe; cloudhal studio does the same in production.
• A Chrome that cannot start still fails every test, with the launch error.
• Three fake-Chrome unit tests (shell scripts, Unix only):
  • a Chrome that keeps exiting gets 3 launches, each with a new profile;
  • a hung launch is killed after startMs, every launched process is gone afterwards, and the error says how long it waited;
  • when a later launch prints an endpoint, the renderer gets past starting.

Changeset: @cloudhal/cli patch.

Tests (run locally, one at a time)

• pnpm exec vitest run test/render-chrome.test.ts in packages/cli: 7/7 passed in 24 s, with real Chrome.
• The fake-Chrome tests passed three more runs in a row.
• pnpm run typecheck (cli): clean.
• Biome on the changed files: clean.
• pnpm size:check: ok.

The first local run found a tight timing in the new hung-launch test. With a 300 ms budget, the fake's shell sometimes hadn't logged its pid yet. The budget is now 1 s, about 10× a normal shell start. The kill itself was verified outside vitest, and both fake processes were gone.

Closes #207

🤖 Generated with Claude Code