feat(fetchers): YouTubeFetcher — video metadata and transcript extraction

[chaliy]

What

Add a YouTubeFetcher that matches youtube.com/watch?v={id} and youtu.be/{id} URLs, returning video metadata and transcript text.

Why

Agents can't watch video, but frequently encounter YouTube links in research, documentation, and discussions. Extracting the transcript turns video content into LLM-consumable text. This is a unique value-add — most tools simply return the noisy YouTube HTML page.

Requirements

• Match: https://youtube.com/watch?v={id}, https://www.youtube.com/watch?v={id}, https://youtu.be/{id}
• Return: title, channel name, description, duration, view count, publish date, tags
• Extract transcript/captions via timedtext API (when available)
• Indicate transcript language and whether auto-generated
• Format field: "youtube_video"
• Handle videos without transcripts gracefully (return metadata only)

Design Notes

• YouTube oEmbed API provides basic metadata: https://www.youtube.com/oembed?url={url}
• Transcript extraction is the key differentiator — investigate timedtext/innertube APIs
• Similar dual-API strategy to existing TwitterFetcher (primary + fallback)
• Transcripts can be long — consider truncation with truncated: true for very long videos
• No official public transcript API — may need to parse from page data or use undocumented endpoints

Tier

3 — Differentiated capability