Use hybrid retrieval to enrich Skill synthesis prompts (replace cluster-blob context)

[chigichan24]

Why

• buildSynthesisPrompt in scripts/skill-synthesizer.ts currently injects cluster-representative sessions and surrounding tool sequences as one large context blob.
• That is coarse: the prompt drags in many "loosely related" turns and few precisely relevant ones. Token budget is wasted and synthesis quality suffers.
• A retriever that picks the top-k most relevant turns for the candidate Skill yields a smaller, denser, more on-topic context — and is the single biggest lever on synthesis quality.

What

• For each SkillCandidate, query the retriever (Add embedding pipeline and chunk-level vector index for RAG-based retrieval #32) using:
  • the topic label
  • top keywords
  • the heuristic SkillCandidate.skillMarkdown itself as a query
• Replace the current "all sessions in cluster" blob with the retrieved turn snippets.
• Keep the existing path reachable behind a flag (--retrieval-context opt-in initially, default-on after evaluation).
• Compose with --preview (Add a --preview mode that runs Skill synthesis without writing files #26) so users can compare context strategies side by side.

How

• Add retriever hookup in buildSynthesisPrompt. Keep prompt structure (frontmatter ask, topic, examples) but swap the examples slot for retrieved snippets.
• Measure synthesis output before/after with the Skill evaluator (Add an automatic Skill evaluation pipeline (skill-creator-style quality checks) #20) on a fixed fixture set.

Depends on

• Add embedding pipeline and chunk-level vector index for RAG-based retrieval #32 — embedding pipeline + retriever
• Add an automatic Skill evaluation pipeline (skill-creator-style quality checks) #20 — Skill evaluator (to quantify the quality delta, not strictly blocking)
• Add a --preview mode that runs Skill synthesis without writing files #26 — --preview is the natural surface to expose this for human inspection