Epic 2: Integrate structured output into core workflows (M28)

[bug-ops]

Context

M27 (#474) added chat_typed<T>() to LlmProvider and Extractor<T> utility, but they are not yet used by any core workflow. This epic integrates structured output where it provides measurable benefit.

Analysis

Not candidates (plain text output, no JSON needed):

• Self-learning (learning.rs) — LLM returns Markdown skill body, not structured data
• Summarization (semantic.rs) — LLM returns free-form text summary
• Legacy fenced blocks (MCP/scrape executors) — parse JSON from text response, but chat_typed operates at request level, not text extraction level

Candidates for integration:

1. Skill matching / intent classification — when the agent decides which skill to invoke, a structured response with { skill_name, confidence, reasoning } would be more reliable than regex/text heuristics
2. Orchestrator model selection — router currently selects models; structured output ensures deterministic { model, reason } responses
3. Structured summarization — instead of free-form summaries, extract { key_facts, entities, sentiment } for richer semantic memory
4. Self-learning evaluation — before improving a skill, evaluate with structured { should_improve, issues, severity } instead of heuristics

Issues

• feat(core): structured intent classification via Extractor #550 feat(core): structured intent classification via Extractor size/M
• feat(llm): structured model selection in router/orchestrator #551 feat(llm): structured model selection in router/orchestrator size/S
• feat(memory): structured summarization with typed key facts #552 feat(memory): structured summarization with typed key facts size/M
• feat(skills): structured self-learning evaluation before improvement #553 feat(skills): structured self-learning evaluation before improvement size/S

Priority

Intent classification (#550) > model selection (#551) > summarization (#552) > self-learning eval (#553)

Verification

Each issue must demonstrate that structured output produces more reliable results than the current approach, with unit tests using MockProvider.