feat: cross-project learning — global agent leaderboard and pattern library

[rlaope]

Summary

Store global agent performance in ~/.bestwork/ so new projects benefit from accumulated experience across all projects.

Requirements

• src/observe/global-leaderboard.ts — aggregate agent scores across all projects
• src/observe/pattern-library.ts — extract reusable patterns (task → agent mapping)
• ~/.bestwork/leaderboards/agents.json — global per-agent stats
• ~/.bestwork/leaderboards/patterns.jsonl — reusable task patterns + best agents
• Modify src/harness/orchestrator.ts — look up global leaderboard before static mapping
• bestwork leaderboard and bestwork patterns CLI commands

Acceptance Criteria

• Global agent rankings across all projects
• New project bootstraps with best-known agents from global data
• Pattern library: "TypeScript refactor → sr-backend 85% success"
• bestwork leaderboard shows cross-project rankings
• Tests for aggregation and pattern extraction

Complexity

4/5 — cross-project file I/O, deduplication, pattern extraction

Dependencies

• feat: agent effectiveness scoring — per-agent success rate and speed tracking #39 (agent effectiveness scoring) — REQUIRED for global scoring
• feat: team auto-optimization — learned agent allocation from historical data #41 (team auto-optimization) — beneficial for using learned patterns