feat(skills): query rewriting before skill retrieval (AutoSkill A3)

[bug-ops]

Description

Before embedding the user's query for skill matching, apply a lightweight LLM rewrite step (fast provider, e.g. qwen3:8b) that converts the raw query into a retrieval-friendly form: removes filler words, normalizes intent into a canonical capability description.

Implemented as an optional pass in SkillMatcher::match_skills controlled by a config flag skills.query_rewrite_provider. When empty/unset, falls back to current behavior (no rewrite).

Motivation

Raw user queries are often too literal or verbose for optimal cosine similarity against skill descriptions. AutoSkill demonstrates improved retrieval accuracy with a rewrite step. Since this is on the hot path (per request), it must be fast and optional.

Acceptance Criteria

• skills.query_rewrite_provider config field (optional, accepts provider name from [[llm.providers]])
• When configured, applies rewrite before embedding in SkillMatcher
• Rewrite prompt is minimal (1 sentence instruction, fast-provider optimized)
• Latency impact documented in PR description (before/after measurement)
• Feature disabled by default

References

• AutoSkill paper: https://arxiv.org/abs/2603.01145
• Research evaluation: specs/058-autoskill-query-rewriting/spec.md