```

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/Services/CaptureService.swift

@@ -33,7 +33,9 @@ struct ParsedReport: Sendable {
     var isEmpty: Bool { indicators.isEmpty }
 
     /// 占位空结果,失败回退时给 UI。
-    static func empty(date: Date = .now) -> ParsedReport {
+    /// nonisolated:本工程默认 MainActor 隔离,而 CaptureService(actor)里的 extractReportMeta
+    /// 需要在 actor 上下文构造空结果 —— 纯值工厂,标 nonisolated 才能跨隔离调用(Swift 6)。
+    nonisolated static func empty(date: Date = .now) -> ParsedReport {
         ParsedReport(
             title: "",
             typeRaw: ReportType.other.rawValue,
@@ -78,6 +80,40 @@ actor CaptureService {
         try await runVL(on: assets)
     }
 
+    /// 报告归档「轻量 meta 提取」:**只保存原图,不逐项识别指标**。
+    /// 逐项多模态识别在 2B 上又慢又易 OOM(jetsam 杀进程 = 用户说的「死机」),
+    /// 故归档链路改为:Vision OCR(本地,<1s/页)→ 文本 LLM 只抽 {title,type,date,institution}(~50 token)。
+    /// 全程容错:OCR 空 / 模型未就绪 / 解析失败都返回 (空 meta, recognized:false),绝不抛、绝不阻断保存原图(§3.2)。
+    /// 返回的 indicators 恒为空 —— 归档不建指标。
+    func extractReportMeta(assets: [FileVault.SavedAsset]) async -> (meta: ParsedReport, recognized: Bool) {
+        let urls = assets.map { FileVault.shared.rootURL.appendingPathComponent($0.relativePath) }
+        let ocr = await Self.ocrReference(for: urls)
+        guard !ocr.trimmingCharacters(in: .whitespacesAndNewlines).isEmpty else {
+            return (.empty(), false)
+        }
+        do {
+            try await AIRuntime.shared.prepare()   // 文本 LLM(轻);OOM 闸门已处理 VL 卸载
+        } catch {
+            return (.empty(), false)
+        }
+        var collected = ""
+        do {
+            // meta 输出极小,256 token 足够,远小于逐项识别的 2048 —— 这是不卡死的关键。
+            let stream = await AIRuntime.shared.generate(prompt: VLPrompts.reportMetaFromText(ocr),
+                                                         maxTokens: 256)
+            for try await chunk in stream { collected += chunk.text }
+        } catch {
+            return (.empty(), false)
+        }
+        let cleaned = CaptureService.stripThink(collected)
+        guard var parsed = try? CaptureService.parseReportJSON(cleaned, pageCount: assets.count) else {
+            return (.empty(), false)
+        }
+        // 归档只存 meta + 原图,丢弃模型可能附带的任何指标。
+        parsed.indicators = []
+        return (parsed, true)
+    }
+
     /// 「拍照识别」OCR 链路:把 Vision OCR 出的纯文本交给 LLM(Qwen3-1.7B)结构化抽指标。
     /// 不建 Report、不留图;失败抛 `CaptureError`,UI 回退手动录入(§3.2)。
     /// 调用方(MainActor)先做 OCR,再把文本传进来——OCR 不需进 actor,也避免 UIImage 跨 actor。
@@ -169,8 +205,17 @@ actor CaptureService {
     private static func ocrReference(for urls: [URL]) async -> String {
         var pages: [String] = []
         for (idx, url) in urls.prefix(4).enumerated() {
-            guard let src = CGImageSourceCreateWithURL(url as CFURL, nil),
-                  let cg = CGImageSourceCreateImageAtIndex(src, 0, nil) else { continue }
+            guard let src = CGImageSourceCreateWithURL(url as CFURL, nil) else { continue }
+            // OCR 不需要全分辨率:一张 4000px 体检照全量解码 ≈48MB,正赶在 VL 推理前叠加,
+            // 易触发 jetsam。降到 ≤3000px 既省内存又加速 Vision,医检报告字号此分辨率仍清晰;
+            // 且原图仍完整交给 VL 自行读取,OCR 仅当数字「抄写员」辅助,降采样不影响最终可用信息。
+            let thumbOptions: [CFString: Any] = [
+                kCGImageSourceCreateThumbnailFromImageAlways: true,
+                kCGImageSourceCreateThumbnailWithTransform: true,
+                kCGImageSourceShouldCacheImmediately: true,
+                kCGImageSourceThumbnailMaxPixelSize: 3000
+            ]
+            guard let cg = CGImageSourceCreateThumbnailAtIndex(src, 0, thumbOptions as CFDictionary) else { continue }
             guard let text = try? await OCRService.recognizeText(in: cg),
                   !text.trimmingCharacters(in: .whitespacesAndNewlines).isEmpty else { continue }
             pages.append(urls.count > 1 ? "【第 \(idx + 1) 页】\n\(text)" : text)

康康/Services/HealthExportService.swift

@@ -450,6 +450,8 @@ struct HealthExportService {
         var reports: [Report]
         var diaries: [DiaryEntry]
         var profile: UserProfile
+        /// 药品库(用户「我有哪些药」清单)→ AI 背景 current_meds。空 → 不写该字段。
+        var medications: [Medication] = []
         /// 相关指标的趋势行(确定性计算,不进 LLM)。空 → 不渲染「## 指标趋势」段。
         var trends: [ExportTrend] = []
     }
@@ -530,6 +532,9 @@ struct HealthExportService {
         // —— Profile(单例) ——
         let profile = UserProfileStore.loadOrCreate(in: ctx)
 
+        // —— 药品库(全量,作为 AI 背景 current_meds) ——
+        let medications = (try? ctx.fetch(FetchDescriptor<Medication>())) ?? []
+
         // —— 趋势(确定性,不进 LLM) ——
         // 用全量 in-window 还原完整序列;裁剪后的 indicators 决定哪些 series 相关。
         let trends = ExportTrendBuilder.build(
@@ -546,6 +551,7 @@ struct HealthExportService {
             reports: reports,
             diaries: diaries,
             profile: profile,
+            medications: medications,
             trends: trends
         )
     }
@@ -561,6 +567,7 @@ struct HealthExportService {
         let indicators = (try? ctx.fetch(indicatorDesc)) ?? []
         let diaries = (try? ctx.fetch(diaryDesc)) ?? []
         let profile = UserProfileStore.loadOrCreate(in: ctx)
+        let medications = (try? ctx.fetch(FetchDescriptor<Medication>())) ?? []
 
         let dates = indicators.map(\.capturedAt) + diaries.map(\.createdAt)
         let fromDate = dates.min() ?? Date()
@@ -581,6 +588,7 @@ struct HealthExportService {
             reports: [],
             diaries: diaries,
             profile: profile,
+            medications: medications,
             trends: trends
         )
     }
@@ -611,7 +619,11 @@ struct HealthExportService {
         if !profile.allergies.isEmpty { profDict["allergies"] = profile.allergies }
         if !profile.chronicConditions.isEmpty { profDict["chronic"] = profile.chronicConditions }
         if !profile.familyHistory.isEmpty { profDict["family_history"] = profile.familyHistory }
-        if !profile.currentMedications.isEmpty { profDict["current_meds"] = profile.currentMedications }
+        // current_meds 改读药品库(Medication);旧 profile.currentMedications 已停用。
+        let medNames = snapshot.medications.map { m in
+            m.detailLine.isEmpty ? m.name : "\(m.name) \(m.detailLine)"
+        }
+        if !medNames.isEmpty { profDict["current_meds"] = medNames }
         root["profile"] = profDict
 
         // symptoms
@@ -681,7 +693,8 @@ struct HealthExportService {
     /// 检索结果是否「实质为空」:无症状/指标/报告/日记,且 profile 也没有任何可写字段。
     /// 为真时跳过 LLM,改用确定性「无记录」摘要,避免小模型凭先验编造病例。
     static func isEffectivelyEmpty(_ s: Snapshot) -> Bool {
-        guard s.symptoms.isEmpty, s.indicators.isEmpty, s.reports.isEmpty, s.diaries.isEmpty else {
+        guard s.symptoms.isEmpty, s.indicators.isEmpty, s.reports.isEmpty,
+              s.diaries.isEmpty, s.medications.isEmpty else {
             return false
         }
         let p = s.profile
@@ -693,7 +706,6 @@ struct HealthExportService {
             && p.allergies.isEmpty
             && p.chronicConditions.isEmpty
             && p.familyHistory.isEmpty
-            && p.currentMedications.isEmpty
     }
 
     /// 无真实记录时的确定性摘要:6 段全「无记录」,主诉仅照搬本人原话,不做任何推断。

康康/Services/ReminderService.swift

@@ -80,20 +80,24 @@ enum ReminderService {
         let title = reminder.title.trimmingCharacters(in: .whitespacesAndNewlines)
         let body = reminder.note.trimmingCharacters(in: .whitespacesAndNewlines)
         let h = reminder.hour, m = reminder.minute
-        let slots: [Slot]
-        switch reminder.frequency {
-        case .daily:
-            slots = [Slot(suffix: "daily", dc: DateComponents(hour: h, minute: m))]
-        case .weekly:
-            slots = reminder.weekdays.map { wd in
-                Slot(suffix: "w\(wd)", dc: DateComponents(hour: h, minute: m, weekday: wd))
+        // 多选频率:把每个选中频率展开成槽,合并调度(suffix 各不冲突,可单独取消)。
+        var slots: [Slot] = []
+        for f in reminder.frequencies {
+            switch f {
+            case .daily:
+                slots.append(Slot(suffix: "daily", dc: DateComponents(hour: h, minute: m)))
+            case .weekly:
+                slots += reminder.weekdays.map { wd in
+                    Slot(suffix: "w\(wd)", dc: DateComponents(hour: h, minute: m, weekday: wd))
+                }
+            case .monthly:
+                slots += reminder.monthlyDays.map { d in
+                    Slot(suffix: "m\(d)", dc: DateComponents(day: d, hour: h, minute: m))
+                }
+            case .yearly:
+                slots.append(Slot(suffix: "yearly",
+                                  dc: DateComponents(month: reminder.month, day: reminder.dayOfMonth, hour: h, minute: m)))
             }
-        case .monthly:
-            slots = [Slot(suffix: "monthly",
-                          dc: DateComponents(day: reminder.dayOfMonth, hour: h, minute: m))]
-        case .yearly:
-            slots = [Slot(suffix: "yearly",
-                          dc: DateComponents(month: reminder.month, day: reminder.dayOfMonth, hour: h, minute: m))]
         }
         await schedule(
             idBase: "\(customIdPrefix)\(reminder.id.uuidString)",
@@ -146,11 +150,13 @@ enum ReminderService {
         }
     }
 
-    /// 取消某个 idBase 下所有可能后缀的 pending 通知(daily/monthly/yearly + 7 个 weekday,不漏)。
+    /// 取消某个 idBase 下所有可能后缀的 pending 通知,不漏:
+    /// daily / yearly / 旧版 monthly + 7 个 weekday(w1...w7)+ 31 个月内日(m1...m31)。
     private static func cancelBase(_ idBase: String) {
         let center = UNUserNotificationCenter.current()
         var ids = ["\(idBase).daily", "\(idBase).monthly", "\(idBase).yearly"]
         ids += (1...7).map { "\(idBase).w\($0)" }
+        ids += (1...31).map { "\(idBase).m\($0)" }
         center.removePendingNotificationRequests(withIdentifiers: ids)
     }
 }

康康/Services/SpeechDictationService.swift

@@ -25,6 +25,16 @@ final class SpeechDictationService {
         }
     }
 
+    /// 把已有文字 `prefix` 与新听写片段 `partial` 合并成一段。纯函数,方便单测、与录音生命周期解耦。
+    /// 规则:空片段保留原文;空前缀直接用片段;前缀已以空白(空格/换行)结尾则直接拼,
+    /// 否则中间补一个空格——避免「已有内容新听写」黏成一坨,也不会在换行后多塞空格。
+    static func merge(prefix: String, partial: String) -> String {
+        if partial.isEmpty { return prefix }
+        if prefix.isEmpty { return partial }
+        if prefix.last?.isWhitespace == true { return prefix + partial }
+        return prefix + " " + partial
+    }
+
     /// 优先系统语言;系统语言不支持端侧时兜底中文(demo 机即使系统是英文也能用)。
     private static func makeRecognizer() -> SFSpeechRecognizer? {
         if let r = SFSpeechRecognizer(locale: .current), r.supportsOnDeviceRecognition {