Integrate natural narration videos into numbered classes

[folker]

What changed

• replaces the broken hard-coded espeak/SoX builder with a repository-relative macOS natural-voice pipeline
• commits all 51 reviewed 1280×720 source frames for deterministic video builds
• corrects the VS Code frame and clipped slide counters
• regenerates captions and records voice, duration, loudness targets, and SHA-256 values
• embeds videos directly in Classes 1–4 with adjacent captions, slides, and narration transcripts
• makes local --include-media builds resolve media correctly from every page
• adds regression tests for frame/narration parity and class media integration

Why

The prior builder referenced a different absolute checkout, expected missing JSON narration, and depended on unavailable espeak and SoX binaries. Existing remastering changed loudness but retained the robotic source voice. This change makes the training videos reproducible from the repository and substantially improves pacing, pronunciation, and speech consistency.

Media QA

Generated local MP4s are 1280×720 H.264 with 48 kHz AAC. All four measure −16.2 LUFS integrated; true peaks range from −1.1 to −1.4 dBTP. Durations are 8.1, 8.6, 10.6, and 9.0 minutes. MP4 binaries remain release/institutional-media assets under the repository's existing publication policy; hashes are recorded in the manifest and build report.

Validation

• visually inspected all 51 source slides at full resolution
• sampled narration from every class and representative encoded video frames
• 66 repository tests: pass
• 6 protected-app tests: pass
• publication lint: pass
• repository validation: pass
• rendered site link check with local media: pass