fix(browser-actions): stop padding video recorded above the viewport size

[chubes4]

Fixes #2552.

Playwright's recorder captures at CSS-viewport pixels and pads any larger recordVideo.size with gray. #2549 defaulted the size to viewport × device-scale-factor, so viewport=432x768 device-scale-factor=2.5 produced a 1080x1920 WebM with the page in the top-left 432x768 corner. The summary also claimed a resolution the content didn't have.

This takes the issue's second option:

• The default recording size is the CSS viewport. There's no device-scale multiplication, so no padding.
• An explicit video-size is still honored, capped at 8,294,400 px. The README now says the recorder can only scale down, and that a higher-resolution output should be upscaled by an encoder afterwards.
• video.width/height in the summary match the real content.

(Capturing true device-pixel frames via CDP Page.startScreencast is the other option in the issue and would be a separate, larger change.)

Tests

npm run build clean. test:browser-video-capture 4/4, including a rewritten high-DPR test that records a full-bleed blue page at 430x932 @2.5 and asserts:

• the summary reports 430x932
• the bottom-right pixel of a decoded frame is page content (blue), not padding

The pixel is read with ffmpeg (PATH, FFMPEG_PATH, or Playwright's bundled ffmpeg), and the check is skipped with a diagnostic if none is available. The cook's original version decoded the WebM inside headless Chromium from a data: URL, but loadeddata never fires there, so the test hung indefinitely.

Also passing: presentation 2/2 + 1/1, annotations 2/2, browser-preview-routing 16/16.

Process note

Drafted by a Homeboy cook (cook-detached-764926ec). Its candidate didn't compile: three call sites still passed the removed scale argument, and the new test hung. With Chris's authorization it was corrected and re-verified by Extra Chill Bot and published outside Homeboy.

AI-assisted: drafted by an OpenCode agent, corrected and finalized by Extra Chill Bot.