重构: F-09 batch5 llm_concurrency会话级改造(决策c)

- state.rs: per_conv HashMap<conv_id,Semaphore>(每对话permits=2) + per_conv_permits热改 + acquire_per_conv(conv_id) + release_conv
- agentic.rs: loop入口acquire_global+acquire_per_conv(conv_id)整loop持有(会话级并发上限3) + 删每轮acquire/drop
- compress/title/knowledge_inject: acquire_per_conv(conv_id)+签名conv_id
- project.rs: 扫描合成key + ai_conversation_delete release_conv
global: LLM调用并发→并发会话数上限;F-260616-12 retry permit自洽
主代兜底: cargo check --workspace 0 + test 96 + grep印证

src-tauri/src/commands/project.rs

@@ -549,7 +549,8 @@ async fn extract_description_via_llm(
     };
 
     let _g = state.llm_concurrency.acquire_global().await;
-    let _c = state.llm_concurrency.acquire_per_conv().await;
+    // F-09 B 批5: 项目扫描无 conv_id(非会话内 LLM 调用),用合成 key 共享一槽(扫描低频,同类聚合限流)。
+    let _c = state.llm_concurrency.acquire_per_conv("__project_scan__").await;
     let resp = provider.complete(request).await.map_err(err_str)?;
     // 只取 description,其它字段丢弃(批量场景不需要 project_type/stack 细化)
     let desc = parse_scan_result(&resp.text)
@@ -644,7 +645,8 @@ pub async fn scan_project_with_ai(
 
     // 4. 双层限流 + complete
     let _global_permit = state.llm_concurrency.acquire_global().await;
-    let _per_conv_permit = state.llm_concurrency.acquire_per_conv().await;
+    // F-09 B 批5: 项目扫描无 conv_id,用合成 key 共享一槽(扫描低频,同类聚合限流)。
+    let _per_conv_permit = state.llm_concurrency.acquire_per_conv("__project_scan__").await;
     let llm_result = provider.complete(request).await;
 
     // 5. 解析 + 合并 stack(LLM 失败降级纯规则)