根据提供的code differences信息，由于没有具体的代码变更内容，我将生成一个通用的commit message模板：

``` docs(readme): 更新文档说明 - 添加了项目使用指南 - 完善了API接口说明 - 修正了一些文字错误 ``` 注：由于未提供具体的代码差异信息，以上为示例格式。请提供具体的代码变更内容以便生成准确的commit message。
2026-06-17 08:35:59 +08:00
parent b3777d508d
commit de19d7abcd
23 changed files with 364 additions and 154 deletions
--- a/康康/AI/AIRuntime.swift
+++ b/康康/AI/AIRuntime.swift
@@ -34,7 +34,6 @@ actor AIRuntime {

    private(set) var status: Status = .notReady
    private(set) var vlStatus: Status = .notReady
-    private(set) var lastDecodeRate: Double = 0

    /// 末次文本生成的性能统计(性能自检页消费;两后端归一)。
    private(set) var lastGenerateStats: GenerateStats?
@@ -247,6 +246,8 @@ actor AIRuntime {
                }
                // 进闸门:保证本次 LLM 解码与任何 VL 解码 / 模型加载串行,绝不并发占显存。
                await self.acquireGate(priority)
+                // defer 保证正常结束 / 异常 / 取消都释放闸门;杜绝未来新增 early-return 导致全局推理死锁。
+                defer { self.releaseGate() }
                do {
                    // session.generate 跨 actor 边界,需要 await
                    let stream = await session.generate(prompt: prompt, maxTokens: maxTokens)
@@ -256,9 +257,6 @@ actor AIRuntime {
                        try Task.checkCancellation()
                        // 后台任务让位:前台请求在排队时,下一个 token 处主动退出。
                        if self.shouldPreempt(priority) { throw CancellationError() }
-                        // Task 闭包在 generate() 内启动,继承 AIRuntime 的 actor 隔离;
-                        // 调用同 actor 的 recordRate 不需要 await
-                        self.recordRate(chunk.decodeRate)
                        continuation.yield(chunk)
                    }
                    self.lastGenerateStats = await session.lastStats
@@ -269,9 +267,6 @@ actor AIRuntime {
                } catch {
                    continuation.finish(throwing: AIRuntimeError.inferenceFailed("\(error)"))
                }
-                // 正常结束 / 异常 / 取消(checkCancellation 抛出后被上面 catch 吞掉)都会走到这,
-                // 闸门一定释放,不会死锁后续推理。
-                self.releaseGate()
            }
            // 消费者取消/流终止时取消内部 Task(与 LLMSession / HealthExportService 一致)。
            continuation.onTermination = { _ in task.cancel() }
@@ -290,6 +285,7 @@ actor AIRuntime {
                    return
                }
                await self.acquireGate(priority)
+                defer { self.releaseGate() }   // 无论正常结束 / 异常 / 取消都释放闸门,防死锁
                do {
                    let stream = await self.mnn.generate(prompt: prompt, maxTokens: maxTokens)
                    for try await chunk in stream {
@@ -297,7 +293,6 @@ actor AIRuntime {
                        // 后台任务让位:前台请求在排队时,下一个 token 处主动退出
                        //(流终止触发 MNNBackend.onTermination → bridge.cancel())。
                        if self.shouldPreempt(priority) { throw CancellationError() }
-                        self.recordRate(chunk.decodeRate)
                        continuation.yield(chunk)
                    }
                    self.lastGenerateStats = await self.mnn.lastStats
@@ -307,16 +302,11 @@ actor AIRuntime {
                } catch {
                    continuation.finish(throwing: AIRuntimeError.inferenceFailed("\(error)"))
                }
-                self.releaseGate()
            }
            continuation.onTermination = { _ in task.cancel() }
        }
    }

-    private func recordRate(_ rate: Double) {
-        if rate > 0 { lastDecodeRate = rate }
-    }
-
    // MARK: - VL

    /// 加载 VL 模型。幂等,首调真正 load。
--- a/康康/AI/InferenceEngine.swift
+++ b/康康/AI/InferenceEngine.swift
@@ -35,7 +35,8 @@ nonisolated enum InferenceEngine: String, CaseIterable, Sendable {
    }

    /// 运行时探测:CPU 是否支持 SME2(A19/iPhone17+)。用于 UI 展示加速状态。
-    static var cpuSupportsSME2: Bool { MNNLLMBridge.cpuSupportsSME2() }
+    /// CPU 能力进程内不变,缓存一次,避免每次 UI 刷新都做 sysctl 系统调用。
+    static let cpuSupportsSME2: Bool = MNNLLMBridge.cpuSupportsSME2()

    // MARK: - 用户偏好(auto / mnn / mlx)

--- a/康康/DesignSystem/VaultImage.swift
+++ b/康康/DesignSystem/VaultImage.swift
@@ -31,8 +31,9 @@ struct VaultImage<Content: View, Placeholder: View>: View {
                placeholder(loading)
            }
        }
-        // id 变了(TabView 翻到新页 / 行复用换 asset)就重新加载;同一身份重渲染不会重复读盘。
-        .task(id: relativePath) {
+        // id 变了(TabView 翻到新页 / 行复用换 asset / 同一 path 改目标尺寸)就重新加载;
+        // 把 maxPixel 纳入 id(与 FileVault 缓存 key 同构),避免同 path 切尺寸时不刷新显示旧图。
+        .task(id: "\(relativePath)@\(Int(maxPixel))") {
            loading = true
            let path = relativePath
            let mp = maxPixel
--- a/康康/Features/Archive/HealthExportListView.swift
+++ b/康康/Features/Archive/HealthExportListView.swift
@@ -117,11 +117,17 @@ struct HealthExportRow: View {
        .tjCard()
    }

-    static func relativeDate(_ d: Date) -> String {
+    /// 复用单个 formatter:RelativeDateTimeFormatter 初始化较贵,列表每行每次重绘都 new 会累积开销。
+    /// 用系统 Locale.current(与原实现一致),进程内不变,可安全缓存。
+    private static let relativeFormatter: RelativeDateTimeFormatter = {
        let f = RelativeDateTimeFormatter()
        f.locale = Locale.current
        f.unitsStyle = .full
-        return f.localizedString(for: d, relativeTo: .now)
+        return f
+    }()
+
+    static func relativeDate(_ d: Date) -> String {
+        relativeFormatter.localizedString(for: d, relativeTo: .now)
    }
 }

--- a/康康/Features/Archive/HealthExportSheet.swift
+++ b/康康/Features/Archive/HealthExportSheet.swift
@@ -19,6 +19,7 @@ struct HealthExportSheet: View {
    @State private var error: Error?
    @State private var completed: Bool = false
    @State private var copiedFlash: Bool = false
+    @State private var lastScrollAt: Date = .distantPast   // 流式滚动节流时间戳
    @State private var answeringTurnID: UUID?
    @State private var retrieval: HealthExportService.RetrievalSummary?
    @State private var turnRetrievals: [UUID: HealthExportService.RetrievalSummary] = [:]
@@ -57,7 +58,7 @@ struct HealthExportSheet: View {
            header
            ScrollViewReader { proxy in
                ScrollView {
-                    VStack(alignment: .leading, spacing: 18) {
+                    LazyVStack(alignment: .leading, spacing: 18) {
                        introSection

                        ForEach(turns) { turn in
@@ -76,15 +77,15 @@ struct HealthExportSheet: View {
                    .padding(.horizontal, 20)
                    .padding(.vertical, 16)
                }
-                .onChange(of: content) { _, _ in
-                    withAnimation(.easeOut(duration: 0.12)) {
-                        proxy.scrollTo("bottom", anchor: .bottom)
-                    }
+                // 流式期间 content / turns 每 token 都变,逐次动画滚动会造成布局抖动;
+                // 节流到 ~8Hz,并在生成完成时补一次滚动确保停在底部。
+                .onChange(of: content) { _, _ in throttledScrollToBottom(proxy) }
+                .onChange(of: turns) { _, _ in throttledScrollToBottom(proxy) }
+                .onChange(of: completed) { _, done in
+                    if done { scrollToBottom(proxy) }
                }
-                .onChange(of: turns) { _, _ in
-                    withAnimation(.easeOut(duration: 0.12)) {
-                        proxy.scrollTo("bottom", anchor: .bottom)
-                    }
+                .onChange(of: answeringTurnID) { _, id in
+                    if id == nil { scrollToBottom(proxy) }
                }
            }
            if completed {
@@ -358,7 +359,7 @@ struct HealthExportSheet: View {
            HStack(spacing: 8) {
                Image(systemName: "exclamationmark.triangle.fill")
                    .foregroundStyle(Tj.Palette.brick)
-                Text(err.localizedDescription)
+                Text(friendlyMessage(for: err))
                    .font(.tjScaled( 13))
                    .foregroundStyle(Tj.Palette.text)
            }
@@ -504,8 +505,16 @@ struct HealthExportSheet: View {
                questionFocused = true
            } catch {
                answeringTurnID = nil
-                appendToTurn(id: assistantTurn.id, text: error.localizedDescription)
                questionFocused = true
+                if error is CancellationError { return }
+                #if DEBUG
+                print("[HealthExport] answer failed: \(error)")
+                #endif
+                // 已有部分回答就保留;否则给一句友好兜底,绝不把技术异常当成 AI 的回答展示。
+                if let idx = turns.firstIndex(where: { $0.id == assistantTurn.id }),
+                   turns[idx].text.trimmingCharacters(in: .whitespacesAndNewlines).isEmpty {
+                    turns[idx].text = "这次没能回答上来,请换个说法再试一次。"
+                }
            }
        }
    }
@@ -552,6 +561,10 @@ struct HealthExportSheet: View {
                    }
                }
            } catch {
+                if error is CancellationError { return }
+                #if DEBUG
+                print("[HealthExport] export failed: \(error)")
+                #endif
                self.error = error
                self.phase = nil
            }
@@ -600,6 +613,27 @@ struct HealthExportSheet: View {
        task?.cancel()
        dismiss()
    }
+
+    // MARK: - 滚动
+
+    private func scrollToBottom(_ proxy: ScrollViewProxy) {
+        lastScrollAt = Date()
+        withAnimation(.easeOut(duration: 0.12)) {
+            proxy.scrollTo("bottom", anchor: .bottom)
+        }
+    }
+
+    /// 节流滚动:流式 token 高频触发,限制到约 8Hz,避免每 token 一次动画布局。
+    private func throttledScrollToBottom(_ proxy: ScrollViewProxy) {
+        guard Date().timeIntervalSince(lastScrollAt) > 0.12 else { return }
+        scrollToBottom(proxy)
+    }
+
+    /// 把推理过程中抛出的技术异常翻译成用户能看懂的一句话(取消不算错误)。
+    private func friendlyMessage(for error: Error) -> String {
+        if error is CancellationError { return "" }
+        return "这次没能生成成功,请稍后重试。"
+    }
 }

 // MARK: - 检索结果 chips(本地 RAG 可视化)
@@ -757,15 +791,30 @@ struct MarkdownView: View {
        return nil
    }

+    /// 行内样式缓存:流式生成时整段会被反复重渲染,缓存命中可避免对同一行重复跑昂贵的
+    /// markdown 解析。键为行文本,稳定的历史行命中缓存,只有正在增长的尾行才真正解析。
+    private final class AttrBox { let value: AttributedString; init(_ v: AttributedString) { value = v } }
+    private static let inlineCache: NSCache<NSString, AttrBox> = {
+        let c = NSCache<NSString, AttrBox>()
+        c.countLimit = 256
+        return c
+    }()
+
    private func inline(_ s: String) -> AttributedString {
+        // 快路径:整段报告里绝大多数行(标题、普通条目、纯正文)都没有内联标记,
+        // 直接走纯文本,跳过 AttributedString(markdown:) —— 这是流式期间最大的一笔开销。
+        if !s.contains(where: { $0 == "*" || $0 == "_" || $0 == "[" || $0 == "`" }) {
+            return AttributedString(s)
+        }
+        let key = s as NSString
+        if let hit = Self.inlineCache.object(forKey: key) { return hit.value }
        // **bold** / *italic* / [text](url) 走 AttributedString markdown 解析
-        if let attr = try? AttributedString(
+        let attr = (try? AttributedString(
            markdown: s,
            options: .init(interpretedSyntax: .inlineOnlyPreservingWhitespace)
-        ) {
-            return attr
-        }
-        return AttributedString(s)
+        )) ?? AttributedString(s)
+        Self.inlineCache.setObject(AttrBox(attr), forKey: key)
+        return attr
    }

    // MARK: - 行级解析
--- a/康康/Features/Capture/UnifiedCaptureFlow.swift
+++ b/康康/Features/Capture/UnifiedCaptureFlow.swift
@@ -153,8 +153,21 @@ struct UnifiedCaptureFlow: View {
        phase = .analyzing(images: images, assets: nil)
        let timeout = analyzeTimeoutSeconds
        analyzeTask = Task {
-            // 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) }
+            // Step 1: 把图写进 Vault(归档核心价值:先把原图存下来),并立刻降采样出预览缩略图。
+            // 整段放后台线程,JPEG 编码逐张包 autoreleasepool 让中间 Data / 位图及时回收 ——
+            // 既不卡主线程,又能在写完后释放全分辨率原图,不让它贯穿整个识别期常驻(jetsam 防护)。
+            let inputBox = UncheckedImageBox(images: images)
+            let written: (assets: [FileVault.SavedAsset], thumbs: UncheckedImageBox) =
+                await Task.detached(priority: .userInitiated) {
+                    let assets = inputBox.images.compactMap { img in
+                        autoreleasepool { try? FileVault.shared.writeJPEG(img) }
+                    }
+                    let thumbs = assets.compactMap {
+                        try? FileVault.shared.loadDownsampledImage(relativePath: $0.relativePath, maxPixelSize: 600)
+                    }
+                    return (assets, UncheckedImageBox(images: thumbs))
+                }.value
+            let assets = written.assets
            // 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。
            if Task.isCancelled {
@@ -171,11 +184,9 @@ struct UnifiedCaptureFlow: View {
                }
                return
            }
-            // 把 assets 暴露给 phase,使工具栏「取消」也能找到孤儿清理。
+            // 原图已落盘:phase 改持 600px 缩略图(释放全分辨率原图),同时把 assets 暴露给「取消」做孤儿清理。
            await MainActor.run {
-                if case .analyzing(let imgs, _) = phase {
-                    phase = .analyzing(images: imgs, assets: assets)
-                }
+                phase = .analyzing(images: written.thumbs.images, assets: assets)
            }

            // Step 2: 轻量 meta 提取(OCR + 文本 LLM,只抽日期/机构/类型/标题)。
@@ -287,6 +298,12 @@ struct UnifiedCaptureFlow: View {
    }
 }

+/// 跨 detached 边界安全携带 UIImage 数组:图片只读、不并发改,封装免 Sendable 报错
+/// (同 MNNBackend.MNNUncheckedBox 思路)。
+private struct UncheckedImageBox: @unchecked Sendable {
+    let images: [UIImage]
+}
+
 // MARK: - 分析中视图

 private struct AnalyzingView: View {
--- a/康康/Features/Diary/DiaryQuickSheet.swift
+++ b/康康/Features/Diary/DiaryQuickSheet.swift
@@ -204,12 +204,12 @@ struct DiaryQuickSheet: View {

                            if let note = voiceNote {
                                HStack(spacing: 6) {
-                                    Image(systemName: "info.circle")
+                                    Image(systemName: "exclamationmark.circle.fill")
                                        .font(.tjScaled(11))
-                                        .foregroundStyle(Tj.Palette.text3)
+                                        .foregroundStyle(Tj.Palette.amber)
                                    Text(note)
                                        .font(.tjScaled(11))
-                                        .foregroundStyle(Tj.Palette.text3)
+                                        .foregroundStyle(Tj.Palette.text2)
                                    Spacer(minLength: 0)
                                }
                            }
@@ -575,7 +575,10 @@ struct DiaryQuickSheet: View {
                    }
                }
            } catch {
-                voiceNote = error.localizedDescription
+                #if DEBUG
+                print("[DiaryVoice] dictation start failed: \(error)")
+                #endif
+                voiceNote = String(appLoc: "无法开始录音,请检查麦克风 / 语音识别权限")
                voicePhase = .idle
            }
        }
@@ -612,7 +615,7 @@ struct DiaryQuickSheet: View {
                guard !Task.isCancelled else { return }
                appendToContent(transcript)   // 红线 #5:整理失败回退原话,不卡死
                organizedAppended = nil
-                voiceNote = String(appLoc: "AI 整理失败,已填入原话")
+                voiceNote = String(appLoc: "AI 整理没成功,已填入未整理的原文")
            }
            withAnimation(.snappy(duration: 0.2)) { voicePhase = .idle }
        }
--- a/康康/Features/Home/HomeCalendarCard.swift
+++ b/康康/Features/Home/HomeCalendarCard.swift
@@ -46,8 +46,8 @@ struct HomeCalendarCard: View {
        return (0..<7).compactMap { calendar.date(byAdding: .day, value: $0, to: monday) }
    }

-    /// 本月有记录的天数(指标/报告/日记/症状任一)。
-    private var daysWithRecordsThisMonth: Int {
+    /// 本月有记录的天数(指标/报告/日记/症状任一)。传入已构建好的 data,避免逐天重建。
+    private func daysWithRecordsThisMonth(_ data: CalendarData) -> Int {
        guard let interval = calendar.dateInterval(of: .month, for: .now) else { return 0 }
        let count = calendar.range(of: .day, in: .month, for: .now)?.count ?? 30
        var n = 0
@@ -62,9 +62,11 @@ struct HomeCalendarCard: View {
    }

    var body: some View {
-        VStack(alignment: .leading, spacing: 12) {
-            header
-            weekStrip
+        // CalendarData 一次性构建,供周条 + 本月摘要复用 —— 杜绝逐格(7)+ 逐天(~30)反复全表重建。
+        let calData = data
+        return VStack(alignment: .leading, spacing: 12) {
+            header(calData)
+            weekStrip(calData)
        }
        .padding(14)
        .frame(maxWidth: .infinity, alignment: .leading)
@@ -77,14 +79,14 @@ struct HomeCalendarCard: View {
        }
    }

-    private var header: some View {
+    private func header(_ data: CalendarData) -> some View {
        HStack(alignment: .firstTextBaseline) {
            Text("健康日历")
                .font(.tjH2())
                .foregroundStyle(Tj.Palette.text)
            Spacer()
            HStack(spacing: 3) {
-                Text(summaryLine)
+                Text(summaryLine(data))
                    .font(.tjScaled( 12))
                    .foregroundStyle(Tj.Palette.text3)
                Image(systemName: "chevron.right")
@@ -94,20 +96,20 @@ struct HomeCalendarCard: View {
        }
    }

-    private var summaryLine: String {
-        let n = daysWithRecordsThisMonth
+    private func summaryLine(_ data: CalendarData) -> String {
+        let n = daysWithRecordsThisMonth(data)
        return n > 0 ? String(appLoc: "本月 \(n) 天有记录") : String(appLoc: "本月暂无记录")
    }

-    private var weekStrip: some View {
+    private func weekStrip(_ data: CalendarData) -> some View {
        HStack(spacing: 6) {
            ForEach(weekDays, id: \.self) { day in
-                dayCell(day)
+                dayCell(day, data)
            }
        }
    }

-    private func dayCell(_ day: Date) -> some View {
+    private func dayCell(_ day: Date, _ data: CalendarData) -> some View {
        let marks = data.marks(for: day, calendar: calendar)
        let ranges = data.ranges(touching: day, calendar: calendar)
        let isToday = calendar.isDateInToday(day)
--- a/康康/Features/Indicator/CustomMetricEditor.swift
+++ b/康康/Features/Indicator/CustomMetricEditor.swift
@@ -64,6 +64,7 @@ struct CustomMetricEditor: View {
    @State private var upper: String = ""
    @State private var icon: String = "circle.fill"
    @State private var hydrated = false
+    @State private var showDeleteConfirm = false

    private var trimmedName: String { name.trimmingCharacters(in: .whitespaces) }
    private var trimmedUnit: String { unit.trimmingCharacters(in: .whitespaces) }
@@ -227,13 +228,7 @@ struct CustomMetricEditor: View {

    private var deleteButton: some View {
        Button(role: .destructive) {
-            if let m = existing {
-                ReminderService.cancel(metricId: m.seriesKey)
-                ctx.delete(m)
-                try? ctx.save()
-                onSaved(nil)
-                dismiss()
-            }
+            showDeleteConfirm = true
        } label: {
            HStack {
                Image(systemName: "trash")
@@ -250,6 +245,21 @@ struct CustomMetricEditor: View {
        }
        .buttonStyle(.plain)
        .padding(.top, 8)
+        .alert(String(appLoc: "删除这项自定义指标?"), isPresented: $showDeleteConfirm) {
+            Button(String(appLoc: "删除"), role: .destructive) { deleteMetric() }
+            Button(String(appLoc: "取消"), role: .cancel) { }
+        } message: {
+            Text("删除后不再监测该指标,已记录的历史数据仍保留。")
+        }
+    }
+
+    private func deleteMetric() {
+        guard let m = existing else { return }
+        ReminderService.cancel(metricId: m.seriesKey)
+        ctx.delete(m)
+        try? ctx.save()
+        onSaved(nil)
+        dismiss()
    }

    private var footer: some View {
--- a/康康/Features/Indicator/IndicatorQuickSheet.swift
+++ b/康康/Features/Indicator/IndicatorQuickSheet.swift
@@ -129,7 +129,29 @@ struct IndicatorQuickSheet: View {
        selectedMonitor?.displayName ?? selectedCustom?.name
    }

+    /// 数值类指标(血压 / 长期监测 / 自定义)要求填合法数字且在合理范围;
+    /// 化验 / 手动自由值允许非数字(如「阴性」「++」「未见异常」)。
+    /// 返回 nil = 通过;否则为给用户看的一句话原因。脏数据会直接画进趋势图,这里挡在入口。
+    private var numericValidationError: String? {
+        func check(_ s: String, min: Double, max: Double, field: String) -> String? {
+            let t = s.trimmingCharacters(in: .whitespaces)
+            guard !t.isEmpty else { return nil }          // 空值交给下面的非空判断兜底
+            guard let v = Double(t), v.isFinite else { return String(appLoc: "\(field)请填数字") }
+            guard v >= min, v <= max else { return String(appLoc: "\(field)数值超出合理范围") }
+            return nil
+        }
+        if isBP {
+            return check(systolic, min: 30, max: 350, field: String(appLoc: "收缩压"))
+                ?? check(diastolic, min: 20, max: 250, field: String(appLoc: "舒张压"))
+        }
+        if isLongTermMetric {   // 长期监测预设 / 自定义:必须是正数
+            return check(value, min: 0.0001, max: 1_000_000, field: String(appLoc: "数值"))
+        }
+        return nil   // 自由输入(化验 / 手动)允许非数字
+    }
+
    private var canSubmit: Bool {
+        guard numericValidationError == nil else { return false }
        if isBP {
            return !systolic.trimmingCharacters(in: .whitespaces).isEmpty &&
                   !diastolic.trimmingCharacters(in: .whitespaces).isEmpty
@@ -162,6 +184,9 @@ struct IndicatorQuickSheet: View {
                            statusSection
                        }
                    }
+                    if let validationError = numericValidationError {
+                        validationHint(validationError)
+                    }

                    timeSection
                    noteSection
@@ -1152,6 +1177,21 @@ struct IndicatorQuickSheet: View {
        return (s.label, s.color)
    }

+    // MARK: - 校验提示
+
+    private func validationHint(_ text: String) -> some View {
+        HStack(spacing: 6) {
+            Image(systemName: "exclamationmark.circle.fill")
+                .font(.tjScaled(11))
+                .foregroundStyle(Tj.Palette.brick)
+            Text(text)
+                .font(.tjScaled(12))
+                .foregroundStyle(Tj.Palette.brick)
+            Spacer(minLength: 0)
+        }
+        .transition(.opacity)
+    }
+
    // MARK: - submit

    private func submit() {
--- a/康康/Features/Me/CustomMetricsListView.swift
+++ b/康康/Features/Me/CustomMetricsListView.swift
@@ -18,11 +18,16 @@ struct CustomMetricsListView: View {
                if metrics.isEmpty {
                    emptyState
                } else {
+                    // 一次性按 seriesKey 统计使用次数(O(indicators)),行内 O(1) 查表,
+                    // 取代「每行都重扫整张指标表」的 O(metrics × indicators) N+1。
+                    let usageCounts: [String: Int] = indicators.reduce(into: [:]) { acc, ind in
+                        if let key = ind.seriesKey, !key.isEmpty { acc[key, default: 0] += 1 }
+                    }
                    ForEach(metrics) { m in
                        Button {
                            editingTarget = CustomMetricEditTarget(metric: m)
                        } label: {
-                            row(m)
+                            row(m, usage: usageCounts[m.seriesKey] ?? 0)
                        }
                        .buttonStyle(.plain)
                    }
@@ -82,9 +87,8 @@ struct CustomMetricsListView: View {
        .frame(maxWidth: .infinity)
    }

-    private func row(_ m: CustomMonitorMetric) -> some View {
-        let count = usageCount(for: m)
-        return HStack(spacing: 12) {
+    private func row(_ m: CustomMonitorMetric, usage count: Int) -> some View {
+        HStack(spacing: 12) {
            ZStack {
                Circle().fill(Tj.Palette.leafSoft)
                Image(systemName: m.icon)
@@ -137,9 +141,6 @@ struct CustomMetricsListView: View {
        )
    }

-    private func usageCount(for m: CustomMonitorMetric) -> Int {
-        indicators.filter { $0.seriesKey == m.seriesKey }.count
-    }
 }

 #Preview {
--- a/康康/Features/Me/CustomReminderEditSheet.swift
+++ b/康康/Features/Me/CustomReminderEditSheet.swift
@@ -27,6 +27,7 @@ struct CustomReminderEditSheet: View {
    @State private var month = 1
    @State private var hydrated = false
    @State private var showAuthDeniedAlert = false
+    @State private var showDeleteConfirm = false

    /// 常用时间快捷预设(时, 分):早 / 午 / 傍晚 / 睡前。
    private let timePresets: [(h: Int, m: Int)] = [(8, 0), (12, 0), (18, 0), (22, 0)]
@@ -79,7 +80,7 @@ struct CustomReminderEditSheet: View {

                if isEditing {
                    Section {
-                        Button(role: .destructive) { deleteReminder() } label: {
+                        Button(role: .destructive) { showDeleteConfirm = true } label: {
                            Label(String(appLoc: "删除提醒"), systemImage: "trash")
                        }
                    }
@@ -110,6 +111,12 @@ struct CustomReminderEditSheet: View {
            } message: {
                Text("提醒已保存,但系统通知权限未开启,到点不会弹出。请在「设置 · 通知 · 康康」中允许。")
            }
+            .alert(String(appLoc: "删除这条提醒?"), isPresented: $showDeleteConfirm) {
+                Button(String(appLoc: "删除"), role: .destructive) { deleteReminder() }
+                Button(String(appLoc: "取消"), role: .cancel) { }
+            } message: {
+                Text("删除后该提醒不再触发。")
+            }
        }
    }

--- a/康康/Features/Profile/MedicationLibraryView.swift
+++ b/康康/Features/Profile/MedicationLibraryView.swift
@@ -177,6 +177,7 @@ private struct MedicationEditSheet: View {
    @State private var viewerStart: PhotoIndex?
    /// 「记录一次服用」:嵌套拉起 MedicationLogSheet,预选当前药。
    @State private var showLog = false
+    @State private var showDeleteConfirm = false

    private var isEditing: Bool { existing != nil }
    private var canSave: Bool {
@@ -251,7 +252,7 @@ private struct MedicationEditSheet: View {

                if isEditing {
                    Section {
-                        Button(role: .destructive) { deleteMedication() } label: {
+                        Button(role: .destructive) { showDeleteConfirm = true } label: {
                            Label(String(appLoc: "从药品库删除"), systemImage: "trash")
                        }
                    }
@@ -280,6 +281,12 @@ private struct MedicationEditSheet: View {
            .sheet(isPresented: $showLog) {
                MedicationLogSheet(preselected: existing)
            }
+            .alert(String(appLoc: "从药品库删除这种药?"), isPresented: $showDeleteConfirm) {
+                Button(String(appLoc: "删除"), role: .destructive) { deleteMedication() }
+                Button(String(appLoc: "取消"), role: .cancel) { }
+            } message: {
+                Text("关联的原图会一并永久删除,无法恢复。")
+            }
        }
    }

--- a/康康/Features/Profile/MedicationScanFlow.swift
+++ b/康康/Features/Profile/MedicationScanFlow.swift
@@ -316,12 +316,13 @@ struct MedicationScanFlow: View {
            return (items, items.isEmpty ? String(appLoc: "没读出药品,可以手动填写") : nil)
        } catch CaptureError.modelNotReady {
            return ([], String(appLoc: "AI 模型未就绪,可以手动填写"))
-        } catch let CaptureError.parseFailed(msg) {
-            return ([], String(appLoc: "解析失败:\(msg)"))
-        } catch let CaptureError.inferenceFailed(msg) {
-            return ([], String(appLoc: "识别失败:\(msg)"))
+        } catch CaptureError.parseFailed {
+            // 已回退手动录入,给用户友好提示;技术细节留在 DEBUG 日志,不抛到屏幕。
+            return ([], String(appLoc: "没认出药品信息,可检查照片清晰度后重拍,或手动填写"))
+        } catch CaptureError.inferenceFailed {
+            return ([], String(appLoc: "识别没成功,可重拍或手动填写"))
        } catch {
-            return ([], String(appLoc: "未知错误:\(error.localizedDescription)"))
+            return ([], String(appLoc: "识别没成功,可重拍或手动填写"))
        }
    }

@@ -354,9 +355,13 @@ enum MedicationArchiver {
        // 原图写加密 Vault(§5/§6:落 Application Support/Vault,目录级硬件加密)。
        // 多药共享同批原图时只挂「第一条新建的药」,避免同一 JPEG 被多个 Asset 引用、
        // 删一条 cascade 误删另一条还在用的文件。
+        // 逐张包 autoreleasepool:让每张的 JPEG Data / 临时位图在下一张编码前就回收,
+        // 压住连续编码 5 张的瞬时内存峰值(这是离散保存动作,不与推理叠加,留在主线程可接受)。
        let savedAssets = images
            .prefix(MedicationScanFlow.maxImages)
-            .compactMap { try? FileVault.shared.writeJPEG($0) }
+            .compactMap { img in
+                autoreleasepool { try? FileVault.shared.writeJPEG(img) }
+            }

        let existing = (try? ctx.fetch(FetchDescriptor<Medication>())) ?? []
        var attachedImages = false
--- a/康康/Features/Quick/QuickRegionCaptureFlow.swift
+++ b/康康/Features/Quick/QuickRegionCaptureFlow.swift
@@ -117,12 +117,13 @@ struct QuickRegionCaptureFlow: View {
            return (items, items.isEmpty ? String(appLoc: "没读出指标,挪一下框再试") : nil)
        } catch CaptureError.modelNotReady {
            return ([], String(appLoc: "AI 模型未就绪,手动补充"))
-        } catch let CaptureError.parseFailed(msg) {
-            return ([], String(appLoc: "解析失败:\(msg)"))
-        } catch let CaptureError.inferenceFailed(msg) {
-            return ([], Task.isCancelled ? nil : String(appLoc: "识别失败:\(msg)"))
+        } catch CaptureError.parseFailed {
+            // 已回退手动录入,给用户友好提示;技术细节留在 CaptureService 的 DEBUG 日志。
+            return ([], String(appLoc: "没自动认出指标,挪一下框再试,或手动填写"))
+        } catch CaptureError.inferenceFailed {
+            return ([], Task.isCancelled ? nil : String(appLoc: "识别没成功,挪一下框再试,或手动填写"))
        } catch {
-            return ([], Task.isCancelled ? nil : String(appLoc: "未知错误:\(error.localizedDescription)"))
+            return ([], Task.isCancelled ? nil : String(appLoc: "识别没成功,请手动填写"))
        }
    }

--- a/康康/Features/Timeline/IndicatorSeriesDetailView.swift
+++ b/康康/Features/Timeline/IndicatorSeriesDetailView.swift
@@ -62,13 +62,15 @@ struct IndicatorSeriesDetailView: View {

    /// 历次血压对:以 bp.systolic 为锚,按 ±5s 配 bp.diastolic(同 TimelineEntry 合并规则)。
    private var bloodPressureRecords: [Record] {
+        // 收缩压、舒张压各过滤一次(O(n)),再在小得多的舒张压子集里配对,
+        // 避免「每条收缩压都重扫整张指标表」的 O(n²) 放大。
        let sysList = indicators
            .filter { $0.seriesKey == "bp.systolic" }
            .sorted { $0.capturedAt > $1.capturedAt }
+        let diaList = indicators.filter { $0.seriesKey == "bp.diastolic" }
        var usedDia = Set<PersistentIdentifier>()
        return sysList.map { sys in
-            let dia = indicators.first {
-                $0.seriesKey == "bp.diastolic" &&
+            let dia = diaList.first {
                !usedDia.contains($0.persistentModelID) &&
                abs($0.capturedAt.timeIntervalSince(sys.capturedAt)) <= 5
            }
--- a/康康/Features/Trends/SeriesChartCard.swift
+++ b/康康/Features/Trends/SeriesChartCard.swift
@@ -4,11 +4,23 @@ import Charts
 struct SeriesChartCard: View {
    let bucket: SeriesBucket

-    private var allPoints: [(line: SeriesBucket.SeriesLine, point: SeriesBucket.Point)] {
-        bucket.lines.flatMap { line in line.points.map { (line, $0) } }
+    // bucket 不可变,这些派生量在 init 里一次性算好存为 let,避免 body 内被
+    // header / chart / daysSpanLabel 多次访问时反复 flatMap / min / max 重算。
+    private let allPoints: [(line: SeriesBucket.SeriesLine, point: SeriesBucket.Point)]
+    private let dateDomain: ClosedRange<Date>?
+    private let valueDomain: ClosedRange<Double>?
+
+    init(bucket: SeriesBucket) {
+        self.bucket = bucket
+        let pts = bucket.lines.flatMap { line in line.points.map { (line, $0) } }
+        self.allPoints = pts
+        self.dateDomain = Self.makeDateDomain(pts)
+        self.valueDomain = Self.makeValueDomain(pts, lines: bucket.lines)
    }

-    private var dateDomain: ClosedRange<Date>? {
+    private static func makeDateDomain(
+        _ allPoints: [(line: SeriesBucket.SeriesLine, point: SeriesBucket.Point)]
+    ) -> ClosedRange<Date>? {
        let dates = allPoints.map(\.point.date)
        guard let lo = dates.min(), let hi = dates.max() else { return nil }
        if lo == hi {
@@ -21,14 +33,17 @@ struct SeriesChartCard: View {
        return lo...hi
    }

-    private var valueDomain: ClosedRange<Double>? {
+    private static func makeValueDomain(
+        _ allPoints: [(line: SeriesBucket.SeriesLine, point: SeriesBucket.Point)],
+        lines: [SeriesBucket.SeriesLine]
+    ) -> ClosedRange<Double>? {
        var lo = Double.greatestFiniteMagnitude
        var hi = -Double.greatestFiniteMagnitude
        for (_, p) in allPoints {
            lo = min(lo, p.value)
            hi = max(hi, p.value)
        }
-        for line in bucket.lines {
+        for line in lines {
            if let r = line.referenceRange {
                lo = min(lo, r.lowerBound)
                hi = max(hi, r.upperBound)
--- a/康康/Features/Trends/TrendDetailView.swift
+++ b/康康/Features/Trends/TrendDetailView.swift
@@ -506,7 +506,11 @@ private struct TrendInsightCard: View {
        do {
            text = try await TrendInsightService.shared.generate(for: bucket)
        } catch {
-            failedMessage = String(appLoc: "AI 解读暂不可用(模型未就绪或繁忙)")
+            // 区分「模型没下载」与「下载了但这次推理没成功」,前者给下载引导(CLAUDE.md §4)。
+            let downloaded = ModelStore.shared.isComplete(for: .mnnLLM) || ModelStore.shared.isComplete(for: .llm)
+            failedMessage = downloaded
+                ? String(appLoc: "本地推理这次没成功,点右上「解读」重试")
+                : String(appLoc: "AI 解读需先在「我的 · 模型管理」下载模型")
        }
        running = false
    }
--- a/康康/Features/Trends/TrendsView.swift
+++ b/康康/Features/Trends/TrendsView.swift
@@ -22,36 +22,32 @@ struct TrendsView: View {
                           customMetrics: customMetrics)
    }

-    private var monitorBuckets: [SeriesBucket] {
-        seriesBuckets.filter { $0.kind == .monitor }
-    }
-    private var labBuckets: [SeriesBucket] {
-        seriesBuckets.filter { $0.kind == .lab }
-    }
-
    private func filtered(_ buckets: [SeriesBucket]) -> [SeriesBucket] {
        let q = query.trimmingCharacters(in: .whitespaces)
        guard !q.isEmpty else { return buckets }
        return buckets.filter { $0.title.localizedCaseInsensitiveContains(q) }
    }
-    private var filteredMonitor: [SeriesBucket] { filtered(monitorBuckets) }
-    private var filteredLab: [SeriesBucket] { filtered(labBuckets) }

    var body: some View {
-        NavigationStack {
+        // SeriesBucket.build 一次性算,monitor / lab / 过滤全部本地派生 ——
+        // 杜绝一次渲染(及每次搜索按键)对整张指标表重复 build ~7 次。
+        let series = seriesBuckets
+        let monitor = filtered(series.filter { $0.kind == .monitor })
+        let lab = filtered(series.filter { $0.kind == .lab })
+        return NavigationStack {
            ScrollView(showsIndicators: false) {
                VStack(alignment: .leading, spacing: 18) {
                    header.padding(.top, 4)
-                    if seriesBuckets.isEmpty {
+                    if series.isEmpty {
                        emptyState
-                    } else if filteredMonitor.isEmpty && filteredLab.isEmpty {
+                    } else if monitor.isEmpty && lab.isEmpty {
                        noMatchState
                    } else {
-                        if !filteredMonitor.isEmpty {
-                            section(title: String(appLoc: "长期监测"), buckets: filteredMonitor)
+                        if !monitor.isEmpty {
+                            section(title: String(appLoc: "长期监测"), buckets: monitor)
                        }
-                        if !filteredLab.isEmpty {
-                            section(title: String(appLoc: "化验指标趋势"), buckets: filteredLab)
+                        if !lab.isEmpty {
+                            section(title: String(appLoc: "化验指标趋势"), buckets: lab)
                        }
                    }
                }
--- a/康康/Localizable.xcstrings
+++ b/康康/Localizable.xcstrings
@@ -394,6 +394,7 @@
      }
    },
    "%lld 个建议" : {
+      "extractionState" : "stale",
      "localizations" : {
        "en" : {
          "stringUnit" : {
@@ -825,9 +826,6 @@
    },
    "%lld 项异常" : {

-    },
-    "%lld." : {
-
    },
    "%lld/%lld 就绪" : {
      "localizations" : {
@@ -1242,12 +1240,6 @@
          }
        }
      }
-    },
-    "AI 生成中 · %.1f tok/s" : {
-
-    },
-    "AI 生成中 · 本地推理" : {
-
    },
    "AI 解读" : {

@@ -1300,6 +1292,7 @@
      }
    },
    "AI 辅助 · 医生角度查漏补缺" : {
+      "extractionState" : "stale",
      "localizations" : {
        "en" : {
          "stringUnit" : {
@@ -1977,6 +1970,9 @@
          }
        }
      }
+    },
+    "主要的都帮你问到啦 · 再想想?" : {
+
    },
    "主页" : {
      "localizations" : {
@@ -2969,6 +2965,9 @@
          }
        }
      }
+    },
+    "停" : {
+
    },
    "停止生成" : {

@@ -3050,6 +3049,7 @@
      }
    },
    "先写几个字,AI 来帮忙补充" : {
+      "extractionState" : "stale",
      "localizations" : {
        "en" : {
          "stringUnit" : {
@@ -3405,6 +3405,7 @@

    },
    "再问一轮 · 让 AI 从新角度追问" : {
+      "extractionState" : "stale",
      "localizations" : {
        "en" : {
          "stringUnit" : {
@@ -5214,6 +5215,7 @@

    },
    "将追加:" : {
+      "extractionState" : "stale",
      "localizations" : {
        "en" : {
          "stringUnit" : {
@@ -5630,9 +5632,6 @@
          }
        }
      }
-    },
-    "已覆盖主要问诊维度;补充原文后可再追问" : {
-
    },
    "已识别边框 · 将自动透视校正" : {
      "extractionState" : "stale",
@@ -5658,6 +5657,7 @@
      }
    },
    "已采纳" : {
+      "extractionState" : "stale",
      "localizations" : {
        "en" : {
          "stringUnit" : {
@@ -5857,6 +5857,15 @@
          }
        }
      }
+    },
+    "康康在想想 · %.1f tok/s" : {
+
+    },
+    "康康在想想…" : {
+
+    },
+    "康康帮你记" : {
+
    },
    "康康是一款以本地优先为设计原则的个人健康随记工具。" : {
      "localizations" : {
@@ -8178,6 +8187,7 @@
      }
    },
    "更新一下原文,再让 AI 继续追问" : {
+      "extractionState" : "stale",
      "localizations" : {
        "en" : {
          "stringUnit" : {
@@ -10278,6 +10288,7 @@
      }
    },
    "第 %lld 轮 · 基于你刚才更新的文本 · %lld 条" : {
+      "extractionState" : "stale",
      "localizations" : {
        "en" : {
          "stringUnit" : {
@@ -10309,6 +10320,7 @@

    },
    "第 1 轮 · %lld 条" : {
+      "extractionState" : "stale",
      "localizations" : {
        "en" : {
          "stringUnit" : {
@@ -11140,6 +11152,7 @@
      }
    },
    "让 AI 帮我想想还能记什么" : {
+      "extractionState" : "stale",
      "localizations" : {
        "en" : {
          "stringUnit" : {
@@ -11160,6 +11173,12 @@
          }
        }
      }
+    },
+    "让康康帮你把这条记得更全" : {
+
+    },
+    "记一下" : {
+
    },
    "记剂量与时间" : {

@@ -11310,6 +11329,7 @@
      }
    },
    "记录身体状态 · 可让 AI 多轮辅助查漏补缺" : {
+      "extractionState" : "stale",
      "localizations" : {
        "en" : {
          "stringUnit" : {
@@ -11330,6 +11350,9 @@
          }
        }
      }
+    },
+    "记录身体状态 · 康康在一旁帮你想还能记点啥" : {
+
    },
    "记录身体状态、用药、感受 · 可让 AI 辅助" : {
      "extractionState" : "stale",
@@ -11895,7 +11918,6 @@
      }
    },
    "跳过" : {
-      "extractionState" : "stale",
      "localizations" : {
        "en" : {
          "stringUnit" : {
@@ -12231,6 +12253,9 @@
          }
        }
      }
+    },
+    "还想到几个想问你 · 再来一轮" : {
+
    },
    "还没有任何记录\n点底部 + 号开始" : {
      "localizations" : {
@@ -12511,6 +12536,7 @@

    },
    "采纳" : {
+      "extractionState" : "stale",
      "localizations" : {
        "en" : {
          "stringUnit" : {
--- a/康康/Persistence/FileVault.swift
+++ b/康康/Persistence/FileVault.swift
@@ -40,6 +40,9 @@ nonisolated final class FileVault: @unchecked Sendable {
    private nonisolated(unsafe) let thumbnailCache: NSCache<NSString, UIImage> = {
        let cache = NSCache<NSString, UIImage>()
        cache.countLimit = 40
+        // 关键:按「解码后真实字节数」限制总量,而非只数张数。否则 40 张全屏 2000px 大图
+        // (各 ~12-16MB)就能占数百 MB 把 App 推过内存上限;混在同一缓存里的小缩略图也会被错误计权。
+        cache.totalCostLimit = 96 * 1024 * 1024   // 96MB
        return cache
    }()

@@ -119,7 +122,9 @@ nonisolated final class FileVault: @unchecked Sendable {
            throw FileVaultError.decodeFailed
        }
        let image = UIImage(cgImage: cg)
-        thumbnailCache.setObject(image, forKey: cacheKey)
+        // cost = 解码后位图字节数,让 NSCache 按真实内存占用淘汰(大图少留、小图多留)。
+        let cost = cg.bytesPerRow * cg.height
+        thumbnailCache.setObject(image, forKey: cacheKey, cost: cost)
        return image
    }

--- a/康康/Services/CaptureService.swift
+++ b/康康/Services/CaptureService.swift
@@ -147,11 +147,14 @@ actor CaptureService {
        do {
            return try CaptureService.parseIndicatorsJSON(cleaned)
        } catch let CaptureError.parseFailed(msg) {
-            // 把模型实际输出的特征带到屏幕上,便于现场定位(原始长度 / strip 后长度 / 前缀)。
-            let rawLen = collected.count
-            let cleanLen = cleaned.count
+            #if DEBUG
+            // 仅 DEBUG:把模型实际输出特征带到屏幕便于现场定位(原始 / strip 后长度 + 前缀)。
+            // Release 绝不把字节数 / JSON 前缀这类调试串抛给用户(§10 不能让用户卡在 AI 错误屏)。
            let preview = cleaned.isEmpty ? "(strip 后为空)" : String(cleaned.prefix(60))
-            throw CaptureError.parseFailed("\(msg)〔raw \(rawLen)字/clean \(cleanLen)字·前缀:\(preview)〕")
+            throw CaptureError.parseFailed("\(msg)〔raw \(collected.count)字/clean \(cleaned.count)字·前缀:\(preview)〕")
+            #else
+            throw CaptureError.parseFailed(msg)
+            #endif
        } catch {
            throw CaptureError.parseFailed("\(error)")
        }
--- a/康康/Services/HealthExportService.swift
+++ b/康康/Services/HealthExportService.swift
@@ -155,7 +155,7 @@ struct HealthExportService {
                        // 用「全文累计 + 每 chunk 重清 + diff yield」:
                        // - thinking 阶段,UI 看到的 generated 始终为空
                        // - 看到 </think> 后,真实内容流式出现
-                        var rawAccum = ""
+                        var stripper = ThinkStripper()
                        let stream = await AIRuntime.shared.generate(
                            prompt: genPrompt,
                            maxTokens: 1024
@@ -163,21 +163,15 @@ struct HealthExportService {
                        for try await chunk in stream {
                            try Task.checkCancellation()
                            if chunk.decodeRate > 0 { lastRate = chunk.decodeRate }
-                            rawAccum += chunk.text
-                            let clean = Self.stripThinkBlocks(rawAccum)
-                            if clean.count > generated.count, clean.hasPrefix(generated) {
-                                let delta = String(clean.dropFirst(generated.count))
-                                generated = clean
+                            let delta = stripper.feed(chunk.text)
+                            if !delta.isEmpty {
                                continuation.yield(.token(TokenChunk(
                                    text: delta,
                                    decodeRate: chunk.decodeRate
                                )))
-                            } else if clean != generated {
-                                // 极少:清理后比上次还短(模型补了开标签)。让 UI 不要回退,
-                                // 直接对齐 generated = clean 但不 yield(避免显示倒退)。
-                                generated = clean
                            }
                        }
+                        generated = stripper.output
                    }

                    guard !generated.trimmingCharacters(in: .whitespacesAndNewlines).isEmpty else {
@@ -250,21 +244,16 @@ struct HealthExportService {
                        dataJSON: dataJSON
                    )

-                    var displayed = ""
-                    var rawAccum = ""
+                    var stripper = ThinkStripper()
                    let stream = await AIRuntime.shared.generate(prompt: prompt, maxTokens: 480)
                    for try await chunk in stream {
                        try Task.checkCancellation()
-                        rawAccum += chunk.text
-                        let clean = Self.stripThinkBlocks(rawAccum)
-                        if clean.count > displayed.count, clean.hasPrefix(displayed) {
-                            let delta = String(clean.dropFirst(displayed.count))
-                            displayed = clean
+                        let delta = stripper.feed(chunk.text)
+                        if !delta.isEmpty {
                            continuation.yield(.token(TokenChunk(text: delta, decodeRate: chunk.decodeRate)))
-                        } else if clean != displayed {
-                            displayed = clean
                        }
                    }
+                    let displayed = stripper.output

                    guard !displayed.trimmingCharacters(in: .whitespacesAndNewlines).isEmpty else {
                        throw ServiceError.generationFailed("模型未输出任何内容")
@@ -307,23 +296,18 @@ struct HealthExportService {
                        dataJSON: dataJSON
                    )

-                    var generated = ""
-                    var rawAccum = ""
                    var lastRate: Double = 0
+                    var stripper = ThinkStripper()
                    let stream = await AIRuntime.shared.generate(prompt: genPrompt, maxTokens: 1200)
                    for try await chunk in stream {
                        try Task.checkCancellation()
                        if chunk.decodeRate > 0 { lastRate = chunk.decodeRate }
-                        rawAccum += chunk.text
-                        let clean = Self.stripThinkBlocks(rawAccum)
-                        if clean.count > generated.count, clean.hasPrefix(generated) {
-                            let delta = String(clean.dropFirst(generated.count))
-                            generated = clean
+                        let delta = stripper.feed(chunk.text)
+                        if !delta.isEmpty {
                            continuation.yield(.token(TokenChunk(text: delta, decodeRate: chunk.decodeRate)))
-                        } else if clean != generated {
-                            generated = clean
                        }
                    }
+                    var generated = stripper.output

                    guard !generated.trimmingCharacters(in: .whitespacesAndNewlines).isEmpty else {
                        throw ServiceError.generationFailed("模型未输出任何内容")
@@ -786,3 +770,38 @@ struct HealthExportService {
        return s
    }
 }
+
+/// 流式 `<think>` 去除器:逐 chunk 喂入,返回这次应增量 yield 的 delta。
+///
+/// 旧实现每个 token 都对「整段累计文本」重跑 `stripThinkBlocks` + `count`/`hasPrefix`/`dropFirst`,
+/// 全是 O(n) grapheme 操作,1024/1200 token 长报告随长度二次方增长(且都在 MainActor 上)。
+/// 这里一旦思考段闭合(出现 `</think>`)或确定不存在(首个非空字符不是 `<`,Qwen 思考必在最前),
+/// 就切到纯增量拼接,把生成主体阶段的每 token 成本降到 O(1)。最坏情况退化为旧行为,无正确性风险。
+private struct ThinkStripper {
+    private var rawAccum = ""
+    private(set) var output = ""
+    private var resolved = false
+
+    mutating func feed(_ piece: String) -> String {
+        rawAccum += piece
+        if resolved {
+            output += piece        // 快路径:思考段已解析,直接增量
+            return piece
+        }
+        let clean = HealthExportService.stripThinkBlocks(rawAccum)
+        var delta = ""
+        if clean.count > output.count, clean.hasPrefix(output) {
+            delta = String(clean.dropFirst(output.count))
+            output = clean
+        } else if clean != output {
+            output = clean         // 清理后比上次短(模型补了开标签):对齐但不回退显示
+        }
+        // 判定能否对「后续」token 切快路径(本 token 已由上面的完整清理处理)。
+        if rawAccum.contains("</think>") {
+            resolved = true                                                   // 思考段已闭合,其后纯增量
+        } else if let c = rawAccum.first(where: { !$0.isWhitespace }), c != "<" {
+            resolved = true                                                   // 开头非 '<' → 不存在 <think>
+        }
+        return delta
+    }
+}