refactor: share one headless-Chrome helper between check-render and build-og-card

Branch off master and PR into master, not main. This is the v1 site.

scripts/build-og-card.mjs (#130) needs headless Chrome for exactly the same reasons scripts/check-render.mjs does, and it now carries its own copy of the same four pieces of plumbing:

• findChrome() — the CHROME_PATH override and the per-platform install locations
• launchChrome() — the flag set, and scraping the DevTools URL out of stderr
• stopChrome() — waiting for the exit rather than the signal, because Windows keeps the user-data-dir locked until Chrome has gone
• connect() / withTimeout() — the minimal CDP client over one browser-level socket

The card builder's copy is the smaller of the two: it drives one page, so it drops the event waitFor and the per-target session bookkeeping check-render needs. Nothing about the Chrome handling itself differs.

This is the shape #162 was about. There, check-canonicals and check-render had a static.yml allowlist parser each, and they had already drifted in a way that mattered: the render check read only the first cp ... _site/, so a page added in a later cp deployed but was never rendered. One parser in scripts/static-allowlist.mjs fixed it. The same argument applies here before these two copies drift — a fix to Chrome discovery, or a new flag needed for a future Chrome, currently has to be made twice, and only one of the two runs in CI, so the other can rot unnoticed.

Not urgent: the copies are identical today, and build-og-card.mjs is a manual tool run rarely rather than a check. Filed so it is a decision rather than an accident.

Proposal

Extract the shared parts into scripts/chrome.mjs, alongside scripts/static-allowlist.mjs, and have both scripts import it. check-render.mjs keeps whatever is genuinely specific to rendering every page (the viewport list, the is-loading and image-ratio measurement, the page loop).

Worth deciding as part of it: whether PAGE_TIMEOUT_MS belongs to the module or stays per-caller, since the two have no reason to agree.

Acceptance criteria

• One copy of Chrome discovery, launch, shutdown and the CDP client
• npm run check-render behaves identically — same failures on a broken image, a distorted aspect ratio and a stuck is-loading
• npm run og-card still produces a byte-comparable card
• Neither script reaches into the other

Refs #130, #162

Read scripts/check-render.mjs and scripts/build-og-card.mjs, then compare their Chrome discovery, launch, shutdown, and CDP client code. Extract the shared parts into scripts/chrome.mjs while leaving rendering-specific behavior and the callers' PAGE_TIMEOUT_MS decisions in place. Run npm run check-render and npm run og-card; done means the render failures remain identical and the card remains byte-comparable.