Remove headless Chrome

[iHiD]

Follow-up to #15. Both image kinds render with satori now, so nothing reaches the Chrome path.

What goes

puppeteer-core and @sparticuz/chromium, plus the code and config that only existed to drive them: generateImage, RENDERER, NAVIGATION_TIMEOUT_MS, SELECTOR_TIMEOUT_MS, and the imageSelector/waitForSelector fields.

node_modules: 138M → 47M (@sparticuz/chromium is 60M of it, puppeteer-core 9.5M). The ECR image is currently 275MB; I'd expect roughly 180–200MB after this, but that's an estimate — worth checking against ECR once it's built.

It also closes the loop on the outage

Chrome screenshotted a rendered page, so it had to fetch exercism.org from a datacentre IP, which is why it needed the NAT address allowlisted in Cloudflare. That's the coupling that left every image timing out for four days.

Payloads now come over the internal ALB. Once this merges, nothing in this lambda depends on that allowlist entry — ip.src eq 13.40.244.182 can be removed from the Cloudflare rules.

The trade-off worth being explicit about

Until now, rolling back to Chrome was unsetting one environment variable. After this it's a revert and a deploy — a couple of minutes rather than seconds.

Solutions have been on satori for about a day (~13k/day, zero errors). Profiles have been on it for minutes. If you'd rather let profiles run overnight before giving up the cheap rollback, holding this until tomorrow costs nothing.

Two things that soften it: images already in S3 keep serving regardless, so a bad render only affects URLs nobody has requested yet; and the failure mode would show up immediately in the error rate rather than silently.

Testing

45 tests pass. Both fixtures render (profile 799ms, solution 566ms). No remaining references to puppeteer, chromium, generateImage, satoriEnabled, imageSelector, waitForSelector or the public baseUrl.

Follow-ups

• RENDERER = "satori" is still set in terraform/image_generator/lambda.tf and becomes inert. Harmless, but worth removing.
• Memory is the real prize. Satori peaks at 302MB against 2000MB allocated; Chrome was the reason for the headroom. Worth ~$10/mo. Test before cutting — Lambda CPU scales with memory and satori/resvg are CPU-bound, so the duration increase can eat the saving. 1024MB is a sensible first step.