feat(core): structured intent classification via Extractor

[bug-ops]

Problem

Skill matching currently uses embedding similarity + text heuristics to classify user intent. When multiple skills have close similarity scores, the choice can be unreliable.

Solution

Use Extractor<T> with a typed IntentClassification struct to get deterministic skill selection from the LLM:

#[derive(Deserialize, JsonSchema)]
struct IntentClassification {
    skill_name: String,
    confidence: f32,
    params: HashMap<String, String>,
}

When embedding similarity is ambiguous (top-2 delta < threshold), fall back to Extractor for disambiguation. This adds an LLM call only when needed.

Acceptance criteria

• IntentClassification struct with JsonSchema derive
• Integration point in skill matcher when top candidates are close
• Unit test with MockProvider verifying disambiguation
• No regression in existing skill matching tests

Part of Epic #549