[abrichr]

Closes #1883

Summary

Full-screen Recorder startup in a macOS VM guest spent ~28 s inside video_pre_callback on take_screenshot just to learn the stream size, past the writer budget. Window-scoped recording already skips that path.

video_pre_callback now sizes the stream with platform.get_screen_dimensions. On Darwin that probe reads Quartz CGDisplayModeGetPixelWidth/Height (matches a Retina grab) and does not read the framebuffer. get_display_pixel_ratio uses the same Quartz-first path. A screenshot remains only as a fallback when metrics are unavailable. get_monitor_dims stays logical for recording metadata.

Testing

• pytest packages/capture/tests/test_video_pre_callback_dims.py: 4 passed (happy path never calls take_screenshot; metrics-unavailable falls back; frame_size still wins; Darwin single-display dims match take_screenshot.size)
• pytest packages/capture/tests/test_no_fabricated_capture.py packages/capture/tests/test_headless_import.py packages/capture/tests/test_video_pre_callback_dims.py: 33 passed
• Control on origin/main: video_pre_callback else branch still uses take_screenshot().size
• Host Mac: Quartz dims (3024, 1964) match take_screenshot.size; ratio 2.0

Untested

• Full-screen Recorder startup inside oa-macvm (tart VM was stopped; not started this turn)
• Multi-monitor virtual-desktop sizing (probe is main-display physical pixels; single-display VM/host verified)