feat(parse): implement video key frame extraction with metadata

[mvanhorn]

Description

Implement video processing in VideoParser using opencv-python-headless. Replaces three stubs:

• _extract_metadata(): extracts duration, resolution, fps, frame count from video files via cv2.VideoCapture
• _extract_keyframes(): captures frames at configurable intervals (default 10s), returns (timestamp, jpeg_bytes) tuples
• _generate_video_description(): produces structured markdown with metadata and keyframe timeline

Also wires real metadata into parse() so ResourceNode gets actual video dimensions instead of placeholder zeros.

Related Issue

Relates to #372

Type of Change

• Bug fix (non-breaking change that fixes an issue)
• New feature (non-breaking change that adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• Documentation update
• Refactoring (no functional changes)
• Performance improvement
• Test update

Changes Made

• Add _extract_metadata() in video.py using cv2 for duration/resolution/fps
• Add _extract_keyframes() in video.py for periodic frame capture with max_frames cap (30)
• Replace _generate_video_description() stub with metadata + keyframe timeline output
• Wire _extract_metadata() into parse() to populate real metadata values
• Add [video] optional dependency group in pyproject.toml (pip install openviking[video])
• Add 7 tests in tests/parse/test_video_keyframes.py

Testing

• Tests mock cv2.VideoCapture to verify metadata extraction, keyframe timing, and graceful fallback
• ruff format and ruff check pass
• All cv2.VideoCapture instances are released in finally blocks to prevent resource leaks

Design decisions

• opencv-python-headless over moviepy/ffmpeg: lighter dependency, pure pip install, no system binary required. Headless variant avoids pulling in GUI dependencies.
• Max 30 keyframes: caps memory usage for long videos. Configurable via the method parameter.
• Metadata in parse(): the existing parse() returned zeros for duration/width/height/fps. Now returns real values when cv2 is available, zeros otherwise (backward compatible).
• No VLM calls yet: this PR adds frame extraction only. VLM scene description per keyframe can be added in a follow-up once the extraction pipeline is validated.

This is the second part of the multimodal parsing trilogy: audio ASR (#805), image OCR (#942), and video processing.

This contribution was developed with AI assistance (Claude Code).

[github-actions]

Failed to generate code suggestions for PR

[mvanhorn]

Same build CI issue as #942 - see #942 (comment). The video extras addition to pyproject.toml triggers the build matrix, which has a pre-existing No module named pip failure. lint, tests, and CLA all pass.