M10/Phase 1: Embedding-based skill matcher

[bug-ops]

Epic: #74
Effort: M (5 SP)

Deliverables

• Embed skill descriptions (frontmatter description field) at load time using LlmProvider::embed()
• SkillMatcher struct: stores skill embeddings, performs cosine similarity search
• On each user query: embed query -> rank skills by similarity -> inject top-K into system prompt
• max_active_skills config field (default: 5) to limit injected skills
• ZEPH_SKILLS_MAX_ACTIVE env var override
• Only matched skills appear in system prompt (replaces current inject-all behavior)

Acceptance Criteria

• Skills are embedded once at startup (or on reload)
• SkillMatcher::match_skills(query, limit) returns ranked skills
• System prompt includes only top-K relevant skills per query
• max_active_skills configurable in TOML and env var
• Fallback: if embedding fails, inject all skills (current behavior)
• Unit tests for matcher ranking, edge cases (no skills, single skill)