Add Layer 4 caching to reduce AI load

Implemented 5-minute short-term caching for relationship inference:

**store.rs**:
- Added relationship_cache SQLite table
- save_relationship_cache(), get_cached_relationship()
- save_all_relationships_cache(), get_cached_all_relationships()
- clear_relationship_cache() - called on memory create/update/delete
- Cache duration: 5 minutes (configurable constant)

**relationship.rs**:
- Modified infer_all_relationships() to use cache
- Added get_relationship() function with caching support
- Cache hit: return immediately
- Cache miss: compute, save to cache, return

**base.rs**:
- Updated tool_get_relationship() to use cached version
- Reduced load from O(n) scan to O(1) cache lookup

**Benefits**:
- Reduces AI load when frequently querying relationships
- Automatic cache invalidation on data changes
- Scales better with growing memory count
- No user-facing changes

**Documentation**:
- Updated ARCHITECTURE.md with caching strategy details

This addresses scalability concerns for Layer 4 as memory data grows.

docs/ARCHITECTURE.md

@@ -362,10 +362,16 @@ if user.extraversion < 0.5 {
 
 ### Design Philosophy
 
-**推測のみ、保存なし**:
+**推測ベース + 短期キャッシング**:
 - 毎回Layer 1-3.5から計算
-- キャッシュなし（シンプルさ優先）
-- 後でキャッシング追加可能
+- 5分間の短期キャッシュで負荷軽減
+- メモリ更新時にキャッシュ無効化
+
+**キャッシング戦略**:
+- SQLiteテーブル（`relationship_cache`）に保存
+- 個別エンティティ: `get_relationship(entity_id)`
+- 全体リスト: `list_relationships()`
+- メモリ作成/更新/削除時に自動クリア
 
 **独立性**:
 - Layer 1-3.5に依存