```

feat(AI): 集成MNN推理引擎替换MLX作为主AI运行时

- 引入MNN(alibaba) + Arm SME2 + CPU作为主AI运行时，支持A19/iPhone17的
  SME2和A17的NEON加速
- 添加MLX Swift作为兜底GPU推理方案，实现双后端切换机制
- 使用单一Qwen3.5-2B多模态模型(1.2GB)，替代原有的LLM+VL分离架构
- 实现InferenceEngine.current引擎选择逻辑，真机默认MNN，模拟器回退MLX
- 更新AIAgent架构，通过MNNLLMBridge(ObjC++) → MNNBackend进行推理
- 修改队列机制防止并发推理导致OOM，使用信号量闸门控制显存占用
- 更新文档中的技术栈说明、模块边界和周次交付计划
```

康康/Features/Capture/UnifiedCaptureFlow.swift

@@ -62,7 +62,7 @@ struct UnifiedCaptureFlow: View {
         switch phase {
         case .idle:       return String(appLoc: "拍摄报告")
         case .analyzing:  return String(appLoc: "本地识别中…")
-        case .editing:    return String(appLoc: "核对识别结果")
+        case .editing:    return String(appLoc: "核对报告信息")
         }
     }
 
@@ -86,6 +86,7 @@ struct UnifiedCaptureFlow: View {
                 parsed: parsed,
                 assets: assets,
                 warning: warning,
+                metaOnly: true,   // 归档只存原图 + meta,不逐项识别(§见 CaptureService.extractReportMeta)
                 onSave: { final in saveAll(parsed: final, assets: assets) },
                 onCancel: cancelAll,
                 onReanalyze: assets.isEmpty ? nil : { reanalyze(assets: assets) }
@@ -152,9 +153,7 @@ struct UnifiedCaptureFlow: View {
         phase = .analyzing(images: images, assets: nil)
         let timeout = analyzeTimeoutSeconds
         analyzeTask = Task {
-            // Step 1: 先把图写进 Vault。
-            // 在 UI 这一层写,而不是塞进 CaptureService.analyze —— 这样取消/失败回退时,
-            // assets 已经在 phase 里,cancelAll 能清理孤儿,editingFallback 也不必再补写。
+            // Step 1: 先把图写进 Vault(归档的核心价值就是「把原图存下来」,先保证它)。
             let assets = images.compactMap { try? FileVault.shared.writeJPEG($0) }
             // 极端情况:用户在写图过程中按了「取消」,View 已 dismiss、cancelAll 看到的
             // phase 还是 .analyzing(_, nil),清不到这批刚写完的图 — 这里手动收尾。
@@ -167,7 +166,7 @@ struct UnifiedCaptureFlow: View {
                     phase = .editing(
                         parsed: .empty(),
                         assets: [],
-                        warning: String(appLoc: "图片保存失败,手动录入并保留文本")
+                        warning: String(appLoc: "图片保存失败,请重试")
                     )
                 }
                 return
@@ -179,49 +178,40 @@ struct UnifiedCaptureFlow: View {
                 }
             }
 
-            // Step 2: VL 推理(timeout 哨兵到点 cancel 父任务,VLSession 在下一个 token break)。
+            // Step 2: 轻量 meta 提取(OCR + 文本 LLM,只抽日期/机构/类型/标题)。
+            // 不再跑多模态逐项识别 —— 那在 2B 上又慢又会 OOM 卡死。watchdog 到点 cancel。
             let watchdog = Task {
                 try? await Task.sleep(for: .seconds(timeout))
                 analyzeTask?.cancel()
             }
             defer { watchdog.cancel() }
 
-            do {
-                let parsed = try await CaptureService.shared.reanalyze(assets: assets)
-                if Task.isCancelled {
-                    await editingFallback(assets: assets,
-                                          msg: String(appLoc: "识别超时(>\(timeout)s),先手动录入"))
-                    return
-                }
+            let (meta, recognized) = await CaptureService.shared.extractReportMeta(assets: assets)
+            if Task.isCancelled {
                 await MainActor.run {
-                    phase = .editing(
-                        parsed: parsed,
-                        assets: assets,
-                        warning: parsed.isEmpty ? String(appLoc: "识别没有读出指标,请手动补充") : nil
-                    )
+                    phase = .editing(parsed: .empty(), assets: assets,
+                                     warning: String(appLoc: "识别超时,已保存原图,请手动填写信息"))
                 }
-            } catch let CaptureError.parseFailed(msg) {
-                await editingFallback(assets: assets, msg: String(appLoc: "VL 输出无法解析:\(msg)"))
-            } catch let CaptureError.inferenceFailed(msg) {
-                await editingFallback(assets: assets,
-                                      msg: Task.isCancelled
-                                          ? String(appLoc: "识别超时(>\(timeout)s),先手动录入")
-                                          : String(appLoc: "推理失败:\(msg)"))
-            } catch CaptureError.modelNotReady {
-                await editingFallback(assets: assets, msg: String(appLoc: "VL 模型未就绪,先手动录入"))
-            } catch {
-                await editingFallback(assets: assets,
-                                      msg: String(appLoc: "未知错误:\(error.localizedDescription)"))
+                return
+            }
+            await MainActor.run {
+                phase = .editing(
+                    parsed: meta,
+                    assets: assets,
+                    warning: recognized ? nil
+                        : String(appLoc: "未能自动识别报告信息,已保存原图,可手动填写日期 / 机构")
+                )
             }
         }
     }
 
-    /// 「重新识别」:复用已存 assets,不再写图,只重跑 VL。
+    /// 「重新识别信息」:复用已存 assets,不再写图,只重跑一次轻量 meta 提取。
     private func reanalyze(assets: [FileVault.SavedAsset]) {
         analyzeTask?.cancel()
-        // 这里没有原始 UIImage,AnalyzingView 显示首张缩略图即可
+        // 这里没有原始 UIImage,AnalyzingView 只把首张缩略图模糊后当背景,降采样到 600px 足够,
+        // 避免为「重新识别」把整页全分辨率原图(数十 MB)载进内存。
         let thumbnails: [UIImage] = assets.compactMap {
-            try? FileVault.shared.loadImage(relativePath: $0.relativePath)
+            try? FileVault.shared.loadDownsampledImage(relativePath: $0.relativePath, maxPixelSize: 600)
         }
         phase = .analyzing(images: thumbnails, assets: assets)
         let timeout = analyzeTimeoutSeconds
@@ -232,40 +222,19 @@ struct UnifiedCaptureFlow: View {
             }
             defer { watchdog.cancel() }
 
-            do {
-                let parsed = try await CaptureService.shared.reanalyze(assets: assets)
-                if Task.isCancelled {
-                    await editingFallback(assets: assets,
-                                         msg: String(appLoc: "识别超时(>\(timeout)s),保留旧编辑"))
-                    return
-                }
+            let (meta, recognized) = await CaptureService.shared.extractReportMeta(assets: assets)
+            if Task.isCancelled {
                 await MainActor.run {
-                    phase = .editing(
-                        parsed: parsed,
-                        assets: assets,
-                        warning: parsed.isEmpty ? String(appLoc: "重新识别没有读出新指标") : nil
-                    )
+                    phase = .editing(parsed: .empty(), assets: assets,
+                                     warning: String(appLoc: "识别超时,已保留原图"))
                 }
-            } catch CaptureError.modelNotReady {
-                await editingFallback(assets: assets, msg: String(appLoc: "VL 模型未就绪"))
-            } catch let CaptureError.parseFailed(msg) {
-                await editingFallback(assets: assets, msg: String(appLoc: "VL 输出无法解析:\(msg)"))
-            } catch let CaptureError.inferenceFailed(msg) {
-                await editingFallback(assets: assets,
-                                     msg: Task.isCancelled
-                                         ? String(appLoc: "识别超时(>\(timeout)s)")
-                                         : String(appLoc: "推理失败:\(msg)"))
-            } catch {
-                await editingFallback(assets: assets,
-                                     msg: String(appLoc: "未知错误:\(error.localizedDescription)"))
+                return
+            }
+            await MainActor.run {
+                phase = .editing(parsed: meta, assets: assets,
+                                 warning: recognized ? nil
+                                    : String(appLoc: "未能自动识别报告信息,可手动填写"))
             }
-        }
-    }
-
-    /// reanalyze 失败时回到 editing,保留 assets 但清空 parsed。
-    private func editingFallback(assets: [FileVault.SavedAsset], msg: String) async {
-        await MainActor.run {
-            phase = .editing(parsed: .empty(), assets: assets, warning: msg)
         }
     }