feat: add search_context tool for searching compressed content

[ranxianglei]

Summary

• New search_context tool lets the model search through ALL compressed block summaries before decompressing
• Fills the gap between compress and decompress: model can find relevant blocks by keyword
• Returns hit list with relevance scores, previews, and decompress hints

Changes

• lib/compress/search.ts: Added createSearchContextTool() — searches active block summaries with term-frequency relevance (topic ×0.3, summary ×0.15), case-insensitive, ranks hits, ~3000 char output cap
• lib/compress/index.ts: Export
• index.ts: Tool registration + toolsToAdd
• lib/prompts/system.ts: Tool description

Usage

search_context({ query: "decoder accuracy" })
→ 📦 [b387] ⭐⭐⭐⭐⭐ (0.95) "Decoder experiments"
   Unified decoder PPL=43, acc=34.4%...
   → decompress(387) for full content