fix: use frontmatter description for skill vectorization instead of overview

[ZaynJarvis]

Context

PR #228 fixed two issues with skill search ranking:

1. Retriever visited set bug — correctly fixed, most relevant results were being dropped
2. Skill embedding text — changed from frontmatter description to LLM-generated overview

This PR reverts change (2) while keeping (1). Skills should embed using the frontmatter description (abstract), not the overview.

Why revert to abstract/frontmatter

Skills are not resources. The two have fundamentally different retrieval patterns:

• Resources are document collections where users search for content within files. Using overview/content for embedding makes sense — users query with natural language about what is inside the documents.

• Skills are tools selected by matching a short description. In system prompts, agents see only name + description to decide which skill to activate. The embedding should match this same selection surface — the frontmatter description — so vector search aligns with how skills are actually discovered and used.

Using overview introduces unnecessary indirection:

• The overview is an LLM-generated summary that may emphasize different aspects than what users/agents query for
• The frontmatter description is human-authored and intentionally crafted for skill discovery (includes trigger keywords, use cases)
• Embedding the description directly means search results reflect the same text the skill author optimized for matching

Recall testing confirms abstract works well:

Query	#1 Result	Correct?	Score
"adding memory"	adding-memory	✓	0.472
"search context"	searching-context	✓	0.551
"RAG semantic search"	openviking	✓	0.646
"remember this"	adding-memory	✓	0.317
"add file to knowledge base"	adding-resource	✓	0.448

5/5 semantic queries rank correctly with abstract-only embedding. The retriever fix from #228 was the real improvement — it ensures all skills appear in results regardless of embedding text quality.

Changes

• skill_processor.py: Revert vectorization text from overview back to context.abstract (frontmatter description)