fix(growth,kernel,runtime): 穷尽审计后 7 项修复 — body 持久化 + embedding 死路径 + 安全加固

CRITICAL 修复:
- body_markdown 数据丢失: SkillManifest.body 字段 + serialize_skill_md 使用 body 替代默认内容
- embedding 检索死路径: rerank_entries 使用异步 index_entry_with_embedding + score_similarity_with_embedding (70/30 混合)
- try_write 静默丢失: pending_embedding 字段 + apply_pending_embedding() 延迟应用

IMPORTANT 修复:
- auto_mode 内存泄漏: add_pending 容量限制 100 + 溢出时丢弃最旧
- name_to_slug 空 ID: uuid fallback for empty/whitespace-only names
- compaction embedding 缺失: compaction GrowthIntegration 也接收 embedding
- kernel 未初始化警告: viking_configure_embedding warn log

验证: 934+ tests PASS, 0 failures

crates/zclaw-growth/src/retriever.rs

@@ -19,6 +19,8 @@ pub struct MemoryRetriever {
     config: RetrievalConfig,
     /// Semantic scorer for similarity computation
     scorer: RwLock<SemanticScorer>,
+    /// Pending embedding client (applied on next scorer access if try_write failed)
+    pending_embedding: std::sync::Mutex<Option<Arc<dyn crate::retrieval::semantic::EmbeddingClient>>>,
     /// Query analyzer
     analyzer: QueryAnalyzer,
     /// Memory cache
@@ -32,6 +34,7 @@ impl MemoryRetriever {
             viking,
             config: RetrievalConfig::default(),
             scorer: RwLock::new(SemanticScorer::new()),
+            pending_embedding: std::sync::Mutex::new(None),
             analyzer: QueryAnalyzer::new(),
             cache: MemoryCache::default_config(),
         }
@@ -244,19 +247,40 @@ impl MemoryRetriever {
 
         let mut scorer = self.scorer.write().await;
 
+        // Apply any pending embedding client
+        self.apply_pending_embedding(&mut scorer);
+
+        // Check if embedding is available for enhanced scoring
+        let use_embedding = scorer.is_embedding_available();
+
         // Index entries for semantic search
-        for entry in &entries {
-            scorer.index_entry(entry);
+        if use_embedding {
+            for entry in &entries {
+                scorer.index_entry_with_embedding(entry).await;
+            }
+        } else {
+            for entry in &entries {
+                scorer.index_entry(entry);
+            }
         }
 
         // Score each entry
-        let mut scored: Vec<(f32, MemoryEntry)> = entries
-            .into_iter()
-            .map(|entry| {
-                let score = scorer.score_similarity(query, &entry);
-                (score, entry)
-            })
-            .collect();
+        let mut scored: Vec<(f32, MemoryEntry)> = if use_embedding {
+            let mut results = Vec::with_capacity(entries.len());
+            for entry in entries {
+                let score = scorer.score_similarity_with_embedding(query, &entry).await;
+                results.push((score, entry));
+            }
+            results
+        } else {
+            entries
+                .into_iter()
+                .map(|entry| {
+                    let score = scorer.score_similarity(query, &entry);
+                    (score, entry)
+                })
+                .collect()
+        };
 
         // Sort by score (descending), then by importance and access count
         scored.sort_by(|a, b| {
@@ -420,7 +444,8 @@ impl MemoryRetriever {
     /// Configure embedding client for semantic similarity
     ///
     /// Stores the client for lazy application on first scorer use.
-    /// Safe to call from non-async contexts.
+    /// If the scorer lock is busy, the client is stored as pending
+    /// and applied on the next successful lock acquisition.
     pub fn set_embedding_client(
         &self,
         client: Arc<dyn crate::retrieval::semantic::EmbeddingClient>,
@@ -429,7 +454,20 @@ impl MemoryRetriever {
             *scorer = SemanticScorer::with_embedding(client);
             tracing::info!("[MemoryRetriever] Embedding client configured for semantic scorer");
         } else {
-            tracing::warn!("[MemoryRetriever] Scorer lock busy, embedding will be applied on next access");
+            tracing::warn!("[MemoryRetriever] Scorer lock busy, storing embedding client as pending");
+            if let Ok(mut pending) = self.pending_embedding.lock() {
+                *pending = Some(client);
+            }
+        }
+    }
+
+    /// Apply any pending embedding client to the scorer.
+    fn apply_pending_embedding(&self, scorer: &mut SemanticScorer) {
+        if let Ok(mut pending) = self.pending_embedding.lock() {
+            if let Some(client) = pending.take() {
+                *scorer = SemanticScorer::with_embedding(client);
+                tracing::info!("[MemoryRetriever] Pending embedding client applied to scorer");
+            }
         }
     }