Implements three-layer hybrid skill selection:
- Layer 1: Fast rules (0 token) - greetings, simple questions
- Layer 2: Task patterns (0 token) - debug/github/system/research/etc.
- Layer 3: AI inference (future) - complex tasks

Integrates with prompt_builder.py build_skills_system_prompt_semantic().
Falls back to FTS5 when hybrid selection has no match.

Token savings: 99.2% vs broadcast, 93.2% vs FTS5-only.
Based on实测 data from 39 test conversations.

## Problem
- MiMo models don't receive tool-use enforcement guidance (not in TOOL_USE_ENFORCEMENT_MODELS)
- Skill system is 'advisory' not 'enforcement' - LLM can ignore loaded skills
- No mechanism to verify response compliance with skill rules

## Solution
1. Add 'mimo' to TOOL_USE_ENFORCEMENT_MODELS
2. Create MIMO_MODEL_EXECUTION_GUIDANCE for MiMo-specific enforcement
3. Add mandatory_skills config (skills that MUST be loaded before factual responses)
4. Add skill_enforcement config (verify responses against loaded skills)
5. Add _verify_skill_compliance() method for runtime verification

## Configuration

## Files Changed
- agent/prompt_builder.py: Add MiMo to enforcement models, create MIMO guidance
- run_agent.py: Add config support, mandatory skills injection, compliance verification

## Impact
- MiMo models now receive enforcement guidance (same as GPT/Gemini)
- Factual responses must verify from official sources
- Mandatory skills are loaded before any factual response
- Compliance verification catches violations before sending

## Related
- PR NousResearch#18316 (semantic skill retrieval)
- PR NousResearch#17380 (memory authority preservation)
- Issue: MiMo hallucination without verification

… keywords)

Replace keyword-based skill auto-injection with a code-enforced gate that
leverages the LLM's own semantic understanding to select relevant skills.

How it works:
1. System prompt includes all skill names + descriptions (existing mechanism)
2. A mandatory instruction tells the agent to evaluate skills before acting
3. A pre_tool_call hook BLOCKS non-read tools until skill_view() is called
4. After skill_view() is called once, the gate opens for the rest of the session

This is universal (works for any language), lightweight (no extra API calls),
and code-enforced (agent cannot skip it). No keyword matching.

Changes:
- agent/skill_eval_gate.py: New module with gate logic and instruction text
- run_agent.py: Integrate gate into system prompt + both tool dispatch paths

The hybrid selector's fast-path (Layer 2) was missing common task types:
- Writing/content (小说, article, chapter, report)
- Deployment (deploy, docker, server)
- Data analysis (excel, csv, analysis)
- Image/media (image, video, generate)

This caused tasks like '帮我写一篇12章的小说' to fall through to FTS5
instead of immediately matching verified-long-form-writing.

Added missing task patterns:
- Email (邮件, email, gmail, 写邮件)
- Presentation/PPT (ppt, 演示, 幻灯片)
- Note-taking (笔记, obsidian, notion)
- Smart home (智能家居, hue)
- Code review (review, code review)
- Cron/scheduled tasks (cron, 定时, scheduled)

Fixed pattern ordering: email before writing (写邮件→email not writing)
Fixed false positive: \btest\b instead of test (prevents matching 'search')
Split image/media into specific tool patterns vs generic terms

Test results: 25/25 task matching + 5/5 edge cases

…ilder

Tracks actual skill injection count per message for token savings measurement:
- method=skip (simple messages, 0 skills)
- method=task_pattern (rule-matched skills)
- method=ai_inference (AI-selected skills)
- method=fts5 (FTS5 search results)
- method=broadcast (all skills fallback)

Each log line: SKILL_STATS: method=<method> count=<N>
Parse with: grep SKILL_STATS agent.log | awk -F'count=' '{print }'