fix(growth): Evolution Engine 审计修复 — 7项全部完成

HIGH-1: 提取共享 json_utils.rs，skill_generator/workflow_composer 去重
HIGH-2: FeedbackCollector Vec→HashMap，消除 unwrap() panic 风险
HIGH-3: ProfileUpdater 改为 collect_updates() 返回字段列表，
        growth.rs 直接 async 调用 update_field()，不再用 no-op 闭包
MEDIUM-1: EvolutionMiddleware 注入后自动 drain，防止重复注入
MEDIUM-2: PatternAggregator tools 提取改为直接收集 context 值
MEDIUM-3: evolution_engine.rs 移除 4 个未使用 imports
MEDIUM-4: workflow_composer parse_response pattern 参数加下划线
MEDIUM-7: SkillCandidate 添加 version 字段（默认=1）

测试: zclaw-growth 128 tests, zclaw-runtime 86 tests, workspace 0 failures

crates/zclaw-runtime/src/middleware/evolution.rs

@@ -68,13 +68,17 @@ impl AgentMiddleware for EvolutionMiddleware {
         &self,
         ctx: &mut MiddlewareContext,
     ) -> Result<MiddlewareDecision> {
-        let pending = self.pending.read().await;
-        if pending.is_empty() {
-            return Ok(MiddlewareDecision::Continue);
-        }
+        let to_inject = {
+            let mut pending = self.pending.write().await;
+            if pending.is_empty() {
+                return Ok(MiddlewareDecision::Continue);
+            }
+            // 只取第一条（最近的）事件注入，避免信息过载
+            // drain 已注入的事件，防止重复注入
+            std::mem::take(&mut *pending)
+        };
 
-        // 只在第一条（最近的）事件上触发提示，避免信息过载
-        if let Some(evolution) = pending.first() {
+        if let Some(evolution) = to_inject.into_iter().next() {
             let injection = format!(
                 "\n\n<evolution-suggestion>\n\
                  我注意到你经常做「{pattern}」相关的事情。\n\