polish: S4 persona definitions — replace blank-slate personas

[0b00101111]

Problem

S4 personas are persona_0 through persona_99 with no identity. The LLM invents a different persona each call with no consistency. 100 calls to the same model produce votes that correlate too highly — consensus, not diversity.

Suggested Improvements

1. Define 5-10 named persona archetypes with demographics, preferences, and viewing habits
   • e.g., "college student who watches during commute, saves practical tips, shares humor"
   • e.g., "millennial parent, watches after kids sleep, engages with relatable parenting content"
2. Give each persona a content preference history ("liked videos about X, skipped Y")
3. Vary voting prompt per persona type — different evaluation criteria
4. Consider different temperature settings per persona for vote diversity
5. Once feedback data exists, calibrate persona preferences against actual performance

Context

See design/pipeline-quality-review.md

Depends on

• PR feat: dataset prep + first real pipeline run (M1-6) #49 (M1-6)
• CreatorProfile expansion (for audience context)