polish: S2 aggregation — preserve verbatim examples and co-occurrence

[0b00101111]

Problem

S2 flattens specific examples into category labels during aggregation. "question + escalating" with 2 examples loses the nuance of each example. No cross-pattern analysis of which combinations produce highest engagement.

Suggested Improvements

1. Preserve 2-3 verbatim hook examples per pattern type (not just labels)
2. Add pattern co-occurrence matrix — which hook types pair with which pacing styles?
3. Weight patterns by source engagement quality if available
4. Cluster by content niche if S1 extracts audience/topic information

Priority

Lower priority — depends on S1 extracting richer data first.

Context

See design/pipeline-quality-review.md

Depends on

• S1 prompt improvements