polish: S1 prompts — extract mechanics, not just labels

[0b00101111]

Problem

S1 extracts pattern labels ("question + escalating") instead of mechanics — what specific words, timing, and rhetorical moves make the hook work. Downstream stages get category names but no actionable detail.

Suggested Improvements

1. Extract verbatim hook text (first 2-3 sentences) alongside the classification
2. Add content_density field (words per second) to inform S3 pacing
3. Add engagement_driver field — is this optimized for saves, shares, or comments?
4. Analyze the transition between hook → body → payoff, not just classify each
5. Include audience inference — who is this video for?
6. Change prompt from "what is the hook_type?" to "transcribe the exact opening and explain why someone would stop scrolling"

Context

See design/pipeline-quality-review.md for full analysis from the first real pipeline run.

Depends on

• PR feat: dataset prep + first real pipeline run (M1-6) #49 (M1-6 dataset + pipeline run)