Add optional Playwright Chrome correctness driver

[chenglou]

Summary

• add an optional Chrome-only Playwright correctness driver behind CHROME_AUTOMATION_DRIVER=playwright
• pin headless Chrome to the validated screen environment instead of silently trusting default headless layout
• keep benchmark runs on the existing foreground AppleScript path and fail loudly if someone tries to use the Playwright path there

Rationale

We now have a mechanically checked-in Chrome correctness baseline in corpora/chrome-step10.json, so we can evaluate a new browser driver by generating the same machine-readable sweep and diffing it directly.

That made Chrome a good candidate for trying Playwright:

• Chrome correctness currently goes through AppleScript, which causes focus blips and is less pleasant for day-to-day work
• Firefox was already on a protocol path, so it did not have the same upside
• Safari still needs to stay real Safari, so Playwright WebKit is not a drop-in replacement there

The important constraint is that this is only for correctness, not benchmarks.

During the investigation:

• plain headless Chrome diverged from the checked-in corpus status
• headed Playwright Chrome matched corpus status, but benchmark numbers came back systematically slower than the current benchmark snapshot
• pinned headless Chrome with a validated screen environment matched corpus status exactly

So this PR takes the conservative split:

• Chrome correctness: optional Playwright path
• Chrome benchmarks: unchanged AppleScript foreground path
• no silent fallback between the two

Implementation notes

• the Playwright path is only selected when CHROME_AUTOMATION_DRIVER=playwright
• it is Chrome-only and non-foreground only
• it uses headless Chrome with --screen-info={3024x1964 devicePixelRatio=2}
• it asserts the pinned environment from inside the page before trusting the run
• benchmark callers still pass foreground: true, so the Playwright path errors immediately there

Verification

• bun run check
• CHROME_AUTOMATION_DRIVER=playwright bun run accuracy-check → 7680/7680
• CHROME_AUTOMATION_DRIVER=playwright bun run corpus-sweep --all --start=300 --end=900 --step=10 --output=/tmp/pretext-playwright-step10-pr.json
• mechanical diff of /tmp/pretext-playwright-step10-pr.json against corpora/chrome-step10.json → 0 diffs
• CHROME_AUTOMATION_DRIVER=playwright bun run benchmark-check fails loudly with the expected guardrail message

Follow-up questions

• whether to add convenience scripts for the Playwright correctness path
• whether this should stay opt-in indefinitely or become the default Chrome correctness driver later