feat(keyboard): add English QuickType bar and system lexicon

Show verbatim/correction/completion slots, mmap a 40k-word list, and
use UITextChecker plus supplementary lexicon for conservative autocorrect.

CHANGELOG.md

@@ -8,6 +8,10 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 ## [Unreleased]
 
 ### Added
+- **English QuickType bar**: while typing a word, three equal slots show the verbatim text (quoted when unknown), the unique Space correction, and a completion. The bar stays empty before typing and between committed words. / **英文 QuickType 栏**：输入单词时，三个等宽格显示原文（生词带引号）、空格会采用的唯一纠错和补全；尚未输入及单词提交后保持空白。
+- **System English lexicon**: typing uses `UITextChecker` completions/guesses and `requestSupplementaryLexicon` contact names / text replacements. / **系统英文词库**：打字使用 `UITextChecker` 补全/猜测，以及 `requestSupplementaryLexicon` 的通讯录名与文本替换。
+- **QWERTY proximity correction**: fat-finger substitutions on neighboring keys (for example `gppd` → `good`) outrank distant edit-distance neighbors. / **邻键纠错**：相邻键的胖手指替换（如 `gppd` → `good`）优先于远键编辑距离。
+- **Larger English word list**: about 40k unigrams and truncated bigrams, derived from Peter Norvig’s public-domain n-gram counts, shipped as an mmap binary so the keyboard extension does not parse them into Swift dictionaries. / **更大英文词表**：约 4 万 unigram 与截断 bigram，来自 Peter Norvig 公有领域 n-gram 计数，以 mmap 二进制随扩展加载，避免解析进 Swift 字典。
 - **Overlapping key presses**: the typing grid tracks multiple fingers, so the next key can go down before the previous lifts. Pending letters commit in press order (not release order); Shift can be held with one finger while another types. / **叠指连打**：打字网格跟踪多指，上一键未松开也可按下下一键。未提交的字母按按下顺序出字（而非抬手顺序）；一只手指按住 Shift 时另一只可打字。
 - **Period shortcut**: in English, a second Space shortly after a Space that follows a word becomes `. ` and arms sentence Shift, matching the system "." Shortcut. / **句号快捷**：英文下，在单词后的空格上短时间内再按一次空格会变成 `. ` 并点亮句首 Shift，对齐系统「句号快捷」。
 - **Return key labels**: Go / Search / Send / Done / Next / Join and the other `UIReturnKeyType` values show their system captions on the green action key instead of collapsing to Send or a return arrow. / **回车键文案**：前往 / 搜索 / 发送 / 完成 / 下一项 / 加入等 `UIReturnKeyType` 在绿色动作键上显示系统对应文案，不再一律变成「发送」或换行箭头。
@@ -17,9 +21,15 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 - **Navigate skill**: copying text and tapping Navigate asks the LLM for one address (or origin → destination) and the host app opens driving directions — Amap if installed, then Baidu Maps, then Apple Maps. No companion Shortcut. No address stays in the current app with a keyboard tip. / **导航技能**：复制文字并点「导航」会抽取一条地址（或起点→终点），由 App 直接打开驾车导航——已装高德则用高德，否则百度，再否则 Apple 地图。不需要配套捷径。没有地址则留在当前 App，键盘上给出提示。
 - **Save to Notes skill**: the Skills tab opens a ready-made companion Shortcut named `OSGSaveToNotes` on the system Add page (one new Apple Note with an explicit title and body). After you tap Add, copying text and tapping Notes asks the LLM for a short title from the time and content; the body stays the original clipboard. / **存入备忘录技能**：技能页会打开已做好的配套捷径 `OSGSaveToNotes` 的系统添加页（新建一条带标题和正文的苹果备忘录）。点添加后，复制文字并点「备忘录」会按时间和内容生成短标题；正文保持剪贴板原文。
 - **Skills clipboard guide**: when Clipboard History is off, the Skills tab shows a card that jumps to in-app Clipboard settings and to iOS Settings for paste authorization. / **技能页剪贴板指引**：未开启剪贴板历史时，技能页展示可点击卡片，分别跳转 App 内剪贴板设置和系统设置以完成粘贴授权。
-- **Custom skills**: Skills tab `+` adds a user skill (name, about, SF Symbol, prompt, required iCloud Shortcut link with name lookup, independent Shortcut name, thinking off by default). Built-in thinking stays off and disabled. No cap on how many custom skills you can save; the keyboard still holds at most 8. / **自定义技能**：技能页右上角 `+` 可添加用户技能（名称、介绍、SF Symbol、提示词、必填 iCloud 捷径链接并自动读取名称、可与技能名分开的捷径名、思考默认关）。内置技能思考固定关闭且不可开。自定义数量不设上限，键盘仍最多启用 8 个。
+- **Custom skills**: Skills tab `+` adds a user skill (name, about, SF Symbol, prompt, optional iCloud Shortcut link with name lookup, independent Shortcut name, thinking off by default). Without a link, the model result is reviewed and inserted directly; with a link, it keeps the Shortcut export flow. Built-in thinking stays off and disabled. No cap on how many custom skills you can save; the keyboard still holds at most 8. / **自定义技能**：技能页右上角 `+` 可添加用户技能（名称、介绍、SF Symbol、提示词、可选 iCloud 捷径链接并自动读取名称、可与技能名分开的捷径名、思考默认关）。不填链接时，模型结果经确认后直接插入；填写链接时继续走捷径导出流程。内置技能思考固定关闭且不可开。自定义数量不设上限，键盘仍最多启用 8 个。
+
+### Changed
+- **Keyboard input tabs**: the four-tab capsule is centered independently of the side controls, with equal 42 pt hit widths and a translucent-black light-mode track. AI and Voice use enlarged `sparkle` and `waveform.mid` symbols, and the leading logo is 16 pt tall. / **键盘输入标签**：四标签胶囊不受两侧控件影响并在键盘上独立居中，点击宽度统一为 42 pt，浅色模式轨道使用半透明黑色；AI 与语音使用放大的 `sparkle` 和 `waveform.mid` 图标，左侧 Logo 高度为 16 pt。
+- **English lexicon mmap**: the 40k-word English table loads only in English and stays file-mapped; Chinese typing no longer pulls it in on keyboard appear. / **英文词表 mmap**：4 万词英文表仅在英文加载且走文件映射；中文打字不再在唤起键盘时一并灌入。
+- **English autocorrect conservatism**: Title Case / short / ALL CAPS tokens are not replaced, except same-length transpositions (`Teh` → `The`). Machine-applied corrections no longer boost the replacement; rejecting them learns the original. / **英文自动更正更克制**：Title Case / 短词 / 全大写默认不改，仅保留同长换位（`Teh` → `The`）。机器改写不再给新词加分；拒绝纠错会学会原文。
 
 ### Fixed
+- **Keyboard switch crash**: `requestSupplementaryLexicon` completion hops to the main actor before writing session state, so switching to OSG Keyboard no longer traps on `com.apple.TextInput.lexicon-request`. / **切换键盘崩溃**：`requestSupplementaryLexicon` 回调先回到主线程再写会话状态，切换到 OSG Keyboard 不再在 `com.apple.TextInput.lexicon-request` 上触发隔离断言。
 - **Save to Notes Shortcut**: bind the combined first-line title and clipboard body through iPhone Create Note’s real `WFCreateNoteInput` field. The previous `contents` binding was ignored and left an enter-content sheet or a title-only note. / **存入备忘录捷径**：通过 iPhone「创建备忘录」真正的 `WFCreateNoteInput` 字段绑定首行标题与剪贴板正文；旧版 `contents` 绑定会被忽略，导致弹出内容填写框或只生成标题。
 - **Extract tasks Shortcut**: receive Shortcut Input as Text, then split lines and add each title to Reminders — the previous recipe could finish successfully without creating items. / **提取待办捷径**：先把快捷指令输入收成文本，再按行写入提醒；旧配方会成功跑完但不创建条目。
 - **Skill reorder feedback**: long-press lifts a skill card and the grid slides live under the finger, matching Home Screen rearrange. / **技能拖动排序**：长按拎起技能卡片，网格随手指实时让位，接近主屏幕图标重排。

NOTICE-TYPING.md

@@ -30,11 +30,16 @@ commit, SHA-256 and output SHA-256.
 - Eight opt-in fuzzy-pinyin rule groups.
 - Offline English typing data under
   `OSGKeyboardShared/Resources/Typing/English/`:
-  - `english_lexicon.tsv` — curated word list with synthetic relative
+  - `english_lexicon.bin` — mmap-friendly binary of the top 40k alphabetic
-    frequency ranks for autocomplete / autocorrect
+    unigrams (log-scaled ranks) plus truncated bigrams; this is what the
-  - `english_bigrams.tsv` — light next-word candidates
+    keyboard extension loads
-  - Not derived from GPL/LGPL dictionaries; ranks are ordering weights,
+  - `english_lexicon.tsv` / `english_bigrams.tsv` — build inputs derived from
-    not a single third-party corpus dump
+    Peter Norvig’s public-domain `count_1w.txt` / `count_2w.txt`
+    (https://norvig.com/ngrams/; not GPL/LGPL dictionaries). Not copied into
+    the app bundle.
+  - Rebuild with `python3 Scripts/typing/build_english_lexicon.py`
+    (add `--from-tsv` to compile the binary from existing TSV without network)
+
 
 ## Reference only
 

OSGKeyboard/OSGKeyboardApp.swift

@@ -31,6 +31,7 @@ struct OSGKeyboardApp: App {
             #if DEBUG
             if ProcessInfo.processInfo.arguments.contains("--whats-new-host") {
                 // Approach A: Notes-like host only; real keyboard extension overlays it.
+                // Also used by `--keyboard-appear-stress=` (pass both flags).
                 Self.makeWhatsNewHostView()
             } else if ProcessInfo.processInfo.arguments.contains("--edit-demo") {
                 EditDemoView()
@@ -79,8 +80,13 @@ struct OSGKeyboardApp: App {
         let scenario = whatsNewScenario(from: args) ?? .edit
         let language = whatsNewLanguage(from: args)
         let seed = whatsNewSeedText(for: scenario, language: language)
+        let appearStressCount = keyboardAppearStressCount(from: args)
         WhatsNewDemoScenario.clear()
-        WhatsNewDemoScenario.arm(scenario, seedText: seed, language: language)
+        // Stress must not arm What's New playback — that drives keys on the
+        // extension while we are tearing it down.
+        if appearStressCount == 0 {
+            WhatsNewDemoScenario.arm(scenario, seedText: seed, language: language)
+        }
         if let defaults = AppGroup.defaultsIfAvailable {
             defaults.set(true, forKey: AppGroupConfiguration.Keys.hasCompletedOnboarding)
             // Force extension ExtL10n / SharedL10n into the demo language.
@@ -98,13 +104,33 @@ struct OSGKeyboardApp: App {
             }
             defaults.synchronize()
         }
+        if appearStressCount > 0, let defaults = AppGroup.defaultsIfAvailable {
+            // Hit the crash path: typing surface + English supplementary lexicon.
+            defaults.set("english", forKey: "typing.input.defaultInputMode")
+            defaults.set(true, forKey: "typing.input.rememberLastSurface")
+            defaults.set("typing", forKey: "typing.input.lastSurface")
+            defaults.set("english", forKey: "typing.input.lastTypingLanguage")
+            defaults.synchronize()
+        }
         return NotesHostDemoView(
             scenario: scenario,
             seedText: seed,
-            language: language
+            language: language,
+            appearStressCount: appearStressCount
         )
     }
 
+    private static func keyboardAppearStressCount(from args: [String]) -> Int {
+        if let paired = args.first(where: { $0.hasPrefix("--keyboard-appear-stress=") }) {
+            return Int(paired.dropFirst("--keyboard-appear-stress=".count)) ?? 0
+        }
+        if let idx = args.firstIndex(of: "--keyboard-appear-stress"),
+           args.index(after: idx) < args.endIndex {
+            return Int(args[args.index(after: idx)]) ?? 0
+        }
+        return 0
+    }
+
     private static func whatsNewScenario(from args: [String]) -> WhatsNewDemoScenario? {
         if let paired = args.first(where: { $0.hasPrefix("--whats-new-scenario=") }) {
             let raw = String(paired.dropFirst("--whats-new-scenario=".count))

OSGKeyboard/Resources/Typing/Licenses/LICENSE-NORVIG-PD.txt

@@ -0,0 +1,7 @@
+Peter Norvig’s n-gram count files (https://norvig.com/ngrams/)
+
+Norvig states: “I hereby release all these files into the public domain.”
+
+OSGKeyboard does not redistribute the raw count files. `Scripts/typing/build_english_lexicon.py`
+derives compact log-scaled unigram ranks and a truncated bigram list, then
+compiles `english_lexicon.bin` for the keyboard extension to mmap.

OSGKeyboard/Resources/Typing/Licenses/NOTICE.txt

@@ -30,6 +30,7 @@ all required copyright and permission notices must remain with distributions.
 
 English typing lexicon
 ----------------------
-english_lexicon.tsv and english_bigrams.tsv are OSG-curated word lists with
+english_lexicon.bin (from english_lexicon.tsv / english_bigrams.tsv) is an
-synthetic relative frequency ranks for offline autocomplete / autocorrect /
+OSG-curated mmap ranking table with synthetic relative frequency ranks for
-next-word ranking. They are not derived from GPL/LGPL dictionaries.
+offline autocomplete / autocorrect / next-word ranking. It is not derived
+from GPL/LGPL dictionaries.

OSGKeyboard/Services/OpenSourceLicenseCatalog.swift

@@ -105,7 +105,7 @@ enum OpenSourceLicenseCatalog {
             id: "english-typing-lexicon",
             name: "OSG English typing lexicon",
             licenseName: "Project-owned notice",
-            purpose: "Offline English autocomplete, autocorrect, and next-word ranking lists curated by OSGKeyboard (english_lexicon.tsv / english_bigrams.tsv). Not derived from GPL/LGPL dictionaries; relative ranks are ordering weights only.",
+            purpose: "Offline English autocomplete, autocorrect, and next-word ranking (english_lexicon.bin, compiled from TSV). Log-scaled ranks derived from Peter Norvig’s public-domain n-gram counts; not GPL/LGPL dictionaries.",
             url: URL(string: "https://github.com/hkgood/OSGKeyboard/blob/main/NOTICE-TYPING.md"),
             licenseText: englishLexiconNoticeText,
             platforms: [.iOS]
@@ -191,16 +191,17 @@ enum OpenSourceLicenseCatalog {
     static let englishLexiconNoticeText = """
     OSG English typing lexicon (project-owned notice)
 
-    english_lexicon.tsv and english_bigrams.tsv are curated by OSGKeyboard for
+    english_lexicon.bin (compiled from english_lexicon.tsv / english_bigrams.tsv)
-    offline English autocomplete, autocorrect, and next-word ranking inside the
+    is the mmap ranking table for offline English autocomplete, autocorrect,
-    iOS keyboard extension.
+    and next-word prediction.
 
-    These lists are not derived from GPL or LGPL dictionaries. Relative
+    Unigram ranks and truncated bigrams are derived from Peter Norvig’s
-    frequency values are synthetic ordering weights for ranking only, not
+    public-domain n-gram count files (https://norvig.com/ngrams/). OSGKeyboard
-    verbatim counts from a single third-party corpus.
+    does not ship the raw corpus. Relative frequency values are log-scaled
+    ordering weights, not verbatim Google counts.
 
     See NOTICE-TYPING.md in the OSGKeyboard repository for the full typing
-    keyboard attribution map (Chinese Rime stack vs OSG-owned English data).
+    keyboard attribution map (Chinese Rime stack vs English data).
     """
 
     static let bsd3Text = """

OSGKeyboard/Views/AIAgentSkillsView.swift

@@ -334,8 +334,17 @@ struct AIAgentSkillsView: View {
     }
 
     private func saveDraft(_ draft: SkillEditorDraft) throws {
-        guard let url = AIShortcutShareLink.parse(draft.shortcutLink) else {
+        let rawShortcutLink = draft.shortcutLink.trimmingCharacters(
-            throw AIUserSkillValidationError.invalidShortcutLink
+            in: .whitespacesAndNewlines
+        )
+        let shortcutURL: URL?
+        if rawShortcutLink.isEmpty {
+            shortcutURL = nil
+        } else {
+            guard let parsedURL = AIShortcutShareLink.parse(rawShortcutLink) else {
+                throw AIUserSkillValidationError.invalidShortcutLink
+            }
+            shortcutURL = parsedURL
         }
         let skill = AIUserSkill(
             id: draft.id,
@@ -343,7 +352,7 @@ struct AIAgentSkillsView: View {
             summary: draft.summary,
             systemImage: draft.systemImage,
             prompt: draft.prompt,
-            shortcutICloudURL: url,
+            shortcutICloudURL: shortcutURL,
             shortcutName: draft.shortcutName,
             thinkingEnabled: draft.thinkingEnabled
         )
@@ -565,7 +574,7 @@ private struct SkillEditorDraft: Identifiable, Equatable {
             summary: skill.summary,
             systemImage: skill.systemImage,
             prompt: skill.prompt,
-            shortcutLink: skill.shortcutICloudURL.absoluteString,
+            shortcutLink: skill.shortcutICloudURL?.absoluteString ?? "",
             shortcutName: skill.shortcutName,
             thinkingEnabled: skill.thinkingEnabled
         )
@@ -957,11 +966,18 @@ private struct SkillEditorSheet: View {
     }
 
     private var canSave: Bool {
-        !name.trimmingCharacters(in: .whitespacesAndNewlines).isEmpty
+        let trimmedShortcutLink = shortcutLink.trimmingCharacters(
+            in: .whitespacesAndNewlines
+        )
+        let validShortcutConfiguration = trimmedShortcutLink.isEmpty
+            || (
+                AIShortcutShareLink.parse(trimmedShortcutLink) != nil
+                    && !shortcutName.trimmingCharacters(in: .whitespacesAndNewlines).isEmpty
+            )
+        return !name.trimmingCharacters(in: .whitespacesAndNewlines).isEmpty
             && !prompt.trimmingCharacters(in: .whitespacesAndNewlines).isEmpty
             && prompt.count <= AIUserSkillLimits.maximumPromptCharacters
-            && !shortcutName.trimmingCharacters(in: .whitespacesAndNewlines).isEmpty
+            && validShortcutConfiguration
-            && AIShortcutShareLink.parse(shortcutLink) != nil
     }
 
     private var currentDraft: SkillEditorDraft {

OSGKeyboard/Views/NotesHostDemoView.swift

@@ -14,6 +14,7 @@ struct NotesHostDemoView: View {
     let scenario: WhatsNewDemoScenario
     let seedText: String
     let language: WhatsNewDemoScenario.Language
+    var appearStressCount: Int = 0
 
     private var title: String {
         switch (scenario, language) {
@@ -43,7 +44,7 @@ struct NotesHostDemoView: View {
                                 .fill(Color(uiColor: .secondarySystemGroupedBackground))
                         )
                 } else {
-                    NotesHostTextView(text: seedText)
+                    NotesHostTextView(text: seedText, appearStressCount: appearStressCount)
                         .padding(16)
                         .frame(maxWidth: .infinity, maxHeight: .infinity, alignment: .topLeading)
                         .background(
@@ -62,6 +63,7 @@ struct NotesHostDemoView: View {
             language == .en ? Locale(identifier: "en") : Locale(identifier: "zh-Hans")
         )
         .task {
+            guard appearStressCount == 0 else { return }
             // Refresh TTL while armed; stop once the extension consumes / plays.
             WhatsNewDemoScenario.arm(scenario, seedText: seedText, language: language)
             for _ in 0..<25 {
@@ -78,6 +80,7 @@ struct NotesHostDemoView: View {
 /// the real custom keyboard extension.
 private struct NotesHostTextView: UIViewRepresentable {
     let text: String
+    var appearStressCount: Int = 0
 
     func makeUIView(context: Context) -> UITextView {
         let view = UITextView()
@@ -91,8 +94,15 @@ private struct NotesHostTextView: UIViewRepresentable {
         view.textContainer.lineFragmentPadding = 0
         view.returnKeyType = .default
         view.delegate = context.coordinator
+        view.accessibilityIdentifier = "notes.host.textView"
+        context.coordinator.appearStressCount = appearStressCount
+        context.coordinator.textView = view
         DispatchQueue.main.asyncAfter(deadline: .now() + 0.35) {
-            view.becomeFirstResponder()
+            if context.coordinator.appearStressCount > 0 {
+                context.coordinator.startAppearStressIfNeeded()
+            } else {
+                view.becomeFirstResponder()
+            }
         }
         return view
     }
@@ -101,6 +111,8 @@ private struct NotesHostTextView: UIViewRepresentable {
         if uiView.text != text, !context.coordinator.userEdited {
             uiView.text = text
         }
+        // Stress owns first-responder; don't fight resignFirstResponder.
+        guard appearStressCount == 0 else { return }
         if !uiView.isFirstResponder {
             DispatchQueue.main.async {
                 _ = uiView.becomeFirstResponder()
@@ -112,10 +124,139 @@ private struct NotesHostTextView: UIViewRepresentable {
 
     final class Coordinator: NSObject, UITextViewDelegate {
         var userEdited = false
+        var appearStressCount = 0
+        weak var textView: UITextView?
+        private var started = false
+        private var waitingForShow = false
+        private var waitingForHide = false
+        private var showWaiter: CheckedContinuation<Bool, Never>?
+        private var hideWaiter: CheckedContinuation<Bool, Never>?
+        /// Invalidates leftover timeout tasks from a finished wait.
+        private var waitGeneration = 0
 
         func textViewDidChange(_ textView: UITextView) {
             userEdited = true
         }
+
+        func startAppearStressIfNeeded() {
+            guard appearStressCount > 0, !started else { return }
+            started = true
+            NotificationCenter.default.addObserver(
+                self,
+                selector: #selector(keyboardDidShow),
+                name: UIResponder.keyboardDidShowNotification,
+                object: nil
+            )
+            NotificationCenter.default.addObserver(
+                self,
+                selector: #selector(keyboardDidHide),
+                name: UIResponder.keyboardDidHideNotification,
+                object: nil
+            )
+            Task { @MainActor [weak self] in
+                await self?.runAppearStress()
+            }
+        }
+
+        deinit {
+            NotificationCenter.default.removeObserver(self)
+        }
+
+        @objc private func keyboardDidShow(_ notification: Notification) {
+            finishWait(show: true, success: true)
+        }
+
+        @objc private func keyboardDidHide(_ notification: Notification) {
+            finishWait(show: false, success: true)
+        }
+
+        private func finishWait(show: Bool, success: Bool) {
+            if show {
+                guard waitingForShow, let pending = showWaiter else { return }
+                waitingForShow = false
+                showWaiter = nil
+                waitGeneration += 1
+                pending.resume(returning: success)
+            } else {
+                guard waitingForHide, let pending = hideWaiter else { return }
+                waitingForHide = false
+                hideWaiter = nil
+                waitGeneration += 1
+                pending.resume(returning: success)
+            }
+        }
+
+        @MainActor
+        private func runAppearStress() async {
+            let total = appearStressCount
+            OSGDiag.log("keyboard.stress begin count=\(total)", category: "boot")
+            guard let textView else {
+                OSGDiag.log("keyboard.stress FAIL textView gone", category: "boot")
+                return
+            }
+            guard await becomeAndWaitForShow(textView, timeoutNanoseconds: 8_000_000_000) else {
+                OSGDiag.log("keyboard.stress FAIL first-show timeout", category: "boot")
+                return
+            }
+            OSGDiag.log("keyboard.stress first-show ok", category: "boot")
+            var passed = 0
+            for cycle in 1...total {
+                guard await resignAndWaitForHide(textView, timeoutNanoseconds: 5_000_000_000) else {
+                    OSGDiag.log("keyboard.stress FAIL cycle=\(cycle) hide timeout", category: "boot")
+                    break
+                }
+                try? await Task.sleep(nanoseconds: 350_000_000)
+                guard await becomeAndWaitForShow(textView, timeoutNanoseconds: 8_000_000_000) else {
+                    OSGDiag.log("keyboard.stress FAIL cycle=\(cycle) show timeout", category: "boot")
+                    break
+                }
+                passed += 1
+                OSGDiag.log("keyboard.stress cycle=\(passed)/\(total) ok", category: "boot")
+                try? await Task.sleep(nanoseconds: 200_000_000)
+            }
+            OSGDiag.log("keyboard.stress done passed=\(passed)/\(total)", category: "boot")
+            try? await Task.sleep(nanoseconds: 250_000_000)
+            exit(passed == total ? 0 : 1)
+        }
+
+        @MainActor
+        private func becomeAndWaitForShow(_ textView: UITextView, timeoutNanoseconds: UInt64) async -> Bool {
+            await waitForKeyboard(show: true, timeoutNanoseconds: timeoutNanoseconds) {
+                textView.becomeFirstResponder()
+            }
+        }
+
+        @MainActor
+        private func resignAndWaitForHide(_ textView: UITextView, timeoutNanoseconds: UInt64) async -> Bool {
+            await waitForKeyboard(show: false, timeoutNanoseconds: timeoutNanoseconds) {
+                textView.resignFirstResponder()
+            }
+        }
+
+        @MainActor
+        private func waitForKeyboard(
+            show: Bool,
+            timeoutNanoseconds: UInt64,
+            trigger: () -> Void
+        ) async -> Bool {
+            await withCheckedContinuation { continuation in
+                waitGeneration += 1
+                let generation = waitGeneration
+                if show {
+                    waitingForShow = true
+                    showWaiter = continuation
+                } else {
+                    waitingForHide = true
+                    hideWaiter = continuation
+                }
+                trigger()
+                Task { @MainActor [weak self] in
+                    try? await Task.sleep(nanoseconds: timeoutNanoseconds)
+                    guard let self, generation == waitGeneration else { return }
+                    self.finishWait(show: show, success: false)
+                }
+            }
+        }
     }
 }
 

OSGKeyboard/en.lproj/Localizable.strings

@@ -477,10 +477,10 @@
 "skills.editor.prompt" = "Processing prompt";
 "skills.editor.thinking" = "Thinking";
 "skills.editor.thinkingHint" = "Off by default. Turn on only when you want slower, deeper reasoning for this skill.";
-"skills.editor.shortcut" = "Shortcut";
+"skills.editor.shortcut" = "Shortcut (Optional)";
 "skills.editor.linkPlaceholder" = "https://www.icloud.com/shortcuts/…";
-"skills.editor.shortcutNamePlaceholder" = "Shortcut name (can differ from the skill name)";
+"skills.editor.shortcutNamePlaceholder" = "Shortcut name (required with a link)";
-"skills.editor.shortcutHint" = "Paste an iCloud share link. The published Shortcut name is filled in automatically and you can change it. Don’t rename it in the Shortcuts app after adding.";
+"skills.editor.shortcutHint" = "Leave the link empty to process text and insert the result after review. With a link, the Shortcut name is filled in automatically. Don’t rename it after adding.";
 "skills.editor.lookingUp" = "Looking up Shortcut name…";
 "skills.editor.resolvedName" = "Will run: %@";
 "skills.editor.lookupFailed" = "Couldn’t read the Shortcut name. Check the link, or type the name yourself.";

OSGKeyboard/zh-Hans.lproj/Localizable.strings

@@ -476,10 +476,10 @@
 "skills.editor.prompt" = "文本处理提示词";
 "skills.editor.thinking" = "思考";
 "skills.editor.thinkingHint" = "默认关闭。仅在需要该技能更慢、更深的推理时开启。";
-"skills.editor.shortcut" = "捷径";
+"skills.editor.shortcut" = "捷径（可选）";
 "skills.editor.linkPlaceholder" = "https://www.icloud.com/shortcuts/…";
-"skills.editor.shortcutNamePlaceholder" = "捷径名称（可与技能名称不同）";
+"skills.editor.shortcutNamePlaceholder" = "捷径名称（填写链接时必填）";
-"skills.editor.shortcutHint" = "粘贴 iCloud 分享链接。发布名称会自动填入，也可以自行修改。添加到「快捷指令」后请勿改名。";
+"skills.editor.shortcutHint" = "不填链接时仅处理文字，结果确认后插入。填写链接时会自动读取捷径名称；添加到「快捷指令」后请勿改名。";
 "skills.editor.lookingUp" = "正在读取捷径名称…";
 "skills.editor.resolvedName" = "将运行：%@";
 "skills.editor.lookupFailed" = "无法读取捷径名称。请检查链接，或手动填写名称。";

OSGKeyboardExt/KeyboardViewController.swift

@@ -42,6 +42,7 @@ public final class KeyboardViewController: UIInputViewController {
     private var typingSession: TypingSessionController {
         if let typingSessionStorage { return typingSessionStorage }
         let created = TypingSessionController()
+        created.systemLexicon = UIKitEnglishSystemLexicon()
         typingSessionStorage = created
         return created
     }
@@ -233,6 +234,7 @@ public final class KeyboardViewController: UIInputViewController {
         if state.surface == .typing {
             OSGDiag.log("KVC.viewWillAppear enterTypingMode", category: "boot")
             typingSession.enterTypingMode()
+            refreshEnglishSupplementaryLexicon()
         }
         clipboardCapture.keyboardDidAppear()
         OSGDiag.log(
@@ -633,6 +635,7 @@ public final class KeyboardViewController: UIInputViewController {
         state.surface = surface
         if surface == .typing {
             typingSession.enterTypingMode()
+            refreshEnglishSupplementaryLexicon()
         } else {
             typingSession.leaveTypingMode()
         }
@@ -754,6 +757,33 @@ public final class KeyboardViewController: UIInputViewController {
         }
     }
 
+    /// Contacts and user text replacements, without a Contacts permission.
+    ///
+    /// `requestSupplementaryLexicon` replies on `com.apple.TextInput.lexicon-request`,
+    /// not the main actor. Touching `TypingSessionController` there traps in Swift 6
+    /// (`_dispatch_assert_queue_fail`) and the extension is killed on appear.
+    private func refreshEnglishSupplementaryLexicon() {
+        requestSupplementaryLexicon { @Sendable lexicon in
+            Task { @MainActor [weak self] in
+                self?.applySupplementaryLexicon(lexicon)
+            }
+        }
+    }
+
+    private func applySupplementaryLexicon(_ lexicon: UILexicon) {
+        typingSessionStorage?.supplementaryWords = lexicon.entries.compactMap { entry -> String? in
+            let text = entry.documentText.trimmingCharacters(in: .whitespacesAndNewlines)
+            guard !text.isEmpty else { return nil }
+            let parts = text.split { $0.isWhitespace || $0 == "," }
+            guard parts.count == 1 else { return nil }
+            let token = String(parts[0])
+            guard token.allSatisfy({ $0.isLetter || $0 == "'" || $0 == "’" || $0 == "-" }) else {
+                return nil
+            }
+            return token
+        }
+    }
+
     private static func typingAutocapitalizationMode(
         for type: UITextAutocapitalizationType
     ) -> TypingAutocapitalizationMode {

OSGKeyboardExt/Typing/TypingRootView.swift

@@ -142,29 +142,31 @@ struct TypingRootView: View {
     }
 
     private var idleTopBar: some View {
-        HStack(spacing: Spacing.xs) {
+        ZStack {
-            KeyboardBrandLogo(action: state.openSettings)
-            // Globe key now lives at the bottom-left of the keyboard (matching
-            // iOS system layout); see the typingKeySurface ForEach.
-
-            if let err = typing.lastError {
-                typingErrorLabel(err)
-            }
-
-            // iOS-style editing cluster (undo / redo / copy / cut) — iPad only,
-            // where the top bar has room to mirror the system shortcut row.
-            if state.usesIPadLayoutMetrics {
-                editingToolbar
-            }
-
-            Spacer(minLength: 0)
-
             KeyboardTopControls(
                 state: state,
                 typing: typing,
                 palette: palette,
                 onInsert: onInsert
             )
+
+            HStack(spacing: Spacing.xs) {
+                KeyboardBrandLogo(action: state.openSettings)
+                // Globe key now lives at the bottom-left of the keyboard (matching
+                // iOS system layout); see the typingKeySurface ForEach.
+
+                if let err = typing.lastError {
+                    typingErrorLabel(err)
+                }
+
+                // iOS-style editing cluster (undo / redo / copy / cut) — iPad only,
+                // where the top bar has room to mirror the system shortcut row.
+                if state.usesIPadLayoutMetrics {
+                    editingToolbar
+                }
+
+                Spacer(minLength: 0)
+            }
         }
         .padding(.horizontal, KeyboardTopBarMetrics.nestedHorizontalInset)
     }
@@ -206,47 +208,11 @@ struct TypingRootView: View {
                 editingToolbar
                     .padding(.leading, KeyboardTopBarMetrics.nestedHorizontalInset)
             }
-            ScrollView(.horizontal, showsIndicators: false) {
+            if typing.language == .english {
-                HStack(spacing: Spacing.xs) {
+                englishQuickTypeBar
-                    if typing.composition.candidates.isEmpty {
+            } else {
-                        selectedCandidateLabel(text: typing.composition.preedit)
+                chineseCandidateStrip
-                    } else if typing.isCandidatePanelExpanded {
-                        candidateChip(text: typing.composition.candidates[0].text) {
-                            apply(typing.selectCandidate(at: 0))
-                        }
-                    } else {
-                        ForEach(
-                            Array(
-                                typing.composition.candidates
-                                    .prefix(TypingLayoutMetrics.collapsedBarCandidateLimit)
-                                    .enumerated()
-                            ),
-                            id: \.element.id
-                        ) { index, candidate in
-                            if index == 0 {
-                                candidateChip(text: candidate.text) {
-                                    apply(typing.selectCandidate(at: index))
-                                }
-                            } else {
-                                Text(candidate.text)
-                                    .font(.system(size: 20, weight: .regular))
-                                    .foregroundStyle(palette.textPrimary)
-                                    .padding(.horizontal, 10)
-                                    .frame(height: 40)
-                                    .contentShape(Rectangle())
-                                    .onTapGesture {
-                                        apply(typing.selectCandidate(at: index))
-                                    }
-                                    .accessibilityAddTraits(.isButton)
-                                    .accessibilityLabel(candidate.text)
-                            }
-                        }
-                    }
-                }
-                .padding(.leading, KeyboardTopBarMetrics.nestedHorizontalInset)
-                .padding(.trailing, Spacing.xs)
             }
-            .scrollBounceBehavior(.basedOnSize, axes: .horizontal)
 
             if typing.canExpandCandidatePanel {
                 expandChevronButton
@@ -256,6 +222,90 @@ struct TypingRootView: View {
         }
     }
 
+    /// Three equal QuickType slots. Space applies only `role == .correction`.
+    private var englishQuickTypeBar: some View {
+        HStack(spacing: 0) {
+            ForEach(
+                Array(
+                    typing.composition.candidates
+                        .prefix(EnglishSuggestionEngine.slotCount)
+                        .enumerated()
+                ),
+                id: \.element.id
+            ) { index, candidate in
+                if index > 0 {
+                    Rectangle()
+                        .fill(palette.dividerStrong)
+                        .frame(width: 1, height: 18)
+                }
+                englishQuickTypeSlot(candidate, index: index)
+            }
+        }
+        .frame(maxWidth: .infinity)
+        .padding(.leading, KeyboardTopBarMetrics.nestedHorizontalInset)
+    }
+
+    private func englishQuickTypeSlot(_ candidate: TypingCandidate, index: Int) -> some View {
+        let label = candidate.isQuoted ? "\"\(candidate.text)\"" : candidate.text
+        let weight: Font.Weight = candidate.role == .correction ? .semibold : .regular
+        return Text(label)
+            .font(.system(size: 17, weight: weight))
+            .foregroundStyle(palette.textPrimary)
+            .lineLimit(1)
+            .minimumScaleFactor(0.75)
+            .frame(maxWidth: .infinity, minHeight: 40)
+            .contentShape(Rectangle())
+            .onTapGesture {
+                apply(typing.selectCandidate(at: index))
+            }
+            .accessibilityAddTraits(.isButton)
+            .accessibilityLabel(candidate.text)
+    }
+
+    private var chineseCandidateStrip: some View {
+        ScrollView(.horizontal, showsIndicators: false) {
+            HStack(spacing: Spacing.xs) {
+                if typing.composition.candidates.isEmpty {
+                    selectedCandidateLabel(text: typing.composition.preedit)
+                } else if typing.isCandidatePanelExpanded {
+                    candidateChip(text: typing.composition.candidates[0].text) {
+                        apply(typing.selectCandidate(at: 0))
+                    }
+                } else {
+                    ForEach(
+                        Array(
+                            typing.composition.candidates
+                                .prefix(TypingLayoutMetrics.collapsedBarCandidateLimit)
+                                .enumerated()
+                        ),
+                        id: \.element.id
+                    ) { index, candidate in
+                        if index == 0 {
+                            candidateChip(text: candidate.text) {
+                                apply(typing.selectCandidate(at: index))
+                            }
+                        } else {
+                            Text(candidate.text)
+                                .font(.system(size: 20, weight: .regular))
+                                .foregroundStyle(palette.textPrimary)
+                                .padding(.horizontal, 10)
+                                .frame(height: 40)
+                                .contentShape(Rectangle())
+                                .onTapGesture {
+                                    apply(typing.selectCandidate(at: index))
+                                }
+                                .accessibilityAddTraits(.isButton)
+                                .accessibilityLabel(candidate.text)
+                        }
+                    }
+                }
+            }
+            .padding(.leading, KeyboardTopBarMetrics.nestedHorizontalInset)
+            .padding(.trailing, Spacing.xs)
+        }
+        .scrollBounceBehavior(.basedOnSize, axes: .horizontal)
+    }
+
     /// Opaque chip like the translation control so ▼ never shares pixels with text.
     private var expandChevronButton: some View {
         Button {

OSGKeyboardExt/Views/AIKeyboardView.swift

@@ -132,15 +132,17 @@ struct AIKeyboardView: View {
             )
             .padding(.horizontal, KeyboardTopBarMetrics.nestedHorizontalInset)
         } else {
-            HStack(spacing: Spacing.xs) {
+            ZStack {
-                KeyboardBrandLogo(action: state.openSettings)
-                Spacer(minLength: 0)
                 KeyboardTopControls(
                     state: state,
                     typing: typing,
                     palette: palette,
                     onInsert: onInsert
                 )
+                HStack {
+                    KeyboardBrandLogo(action: state.openSettings)
+                    Spacer(minLength: 0)
+                }
             }
             .padding(.horizontal, KeyboardTopBarMetrics.nestedHorizontalInset)
         }

OSGKeyboardExt/Views/KeyboardRootView.swift

@@ -238,17 +238,19 @@ public struct KeyboardRootView: View {
                     onDismiss: state.dismissClipboardSuggestion
                 )
             } else {
-                HStack(spacing: Spacing.xs) {
+                ZStack {
-                    KeyboardBrandLogo(action: state.openSettings)
-                    // Globe key now lives at the bottom-left of the keyboard (matching
-                    // iOS system layout); see micActionRow's bottom HStack.
-                    Spacer(minLength: 0)
                     KeyboardTopControls(
                         state: state,
                         typing: typing,
                         palette: palette,
                         onInsert: onInsert
                     )
+                    HStack {
+                        KeyboardBrandLogo(action: state.openSettings)
+                        // Globe key now lives at the bottom-left of the keyboard (matching
+                        // iOS system layout); see micActionRow's bottom HStack.
+                        Spacer(minLength: 0)
+                    }
                 }
             }
         }

OSGKeyboardExt/Views/KeyboardTopControls.swift

@@ -23,8 +23,10 @@ enum KeyboardTopBarMetrics {
     static let horizontalInset: CGFloat = 12
     /// TypingRootView already contributes 8 pt around the entire key surface.
     static let nestedHorizontalInset: CGFloat = horizontalInset - KeyboardChromeLayout.horizontalInset
-    static let logoHeight: CGFloat = 22
+    static let logoHeight: CGFloat = 16
     static let logoWidth: CGFloat = logoHeight * 952 / 291
+    /// Equal hit width for AI / Voice / Chinese / English input tabs.
+    static let inputTabWidth: CGFloat = 42
     /// Shared footprint for top-trailing chips (clipboard, cancel/X, translation).
     static let trailingChipSize: CGFloat = 34
     static let trailingChipIconSize: CGFloat = 15
@@ -113,39 +115,45 @@ struct KeyboardTopControls: View {
     let onInsert: (String) -> Void
 
     var body: some View {
-        HStack(spacing: 6) {
+        ZStack {
-            // 分段轨道：不透明灰底；选中项用白/升高键面滑动，避免半透明发淡。
+            inputTabSwitcher
-            HStack(spacing: 2) {
+            if state.canShowClipboardEntry {
-                ForEach(KeyboardInputTab.allCases, id: \.self) { tab in
+                HStack {
-                    tabButton(tab)
+                    Spacer(minLength: 0)
+                    KeyboardClipboardMenuButton(
+                        palette: palette,
+                        action: state.openClipboardPanel
+                    )
+                    .equatable()
                 }
             }
-            .padding(2)
+        }
-            .background(tabTrackFill, in: Capsule())
+        .frame(maxWidth: .infinity)
-            .overlay(
+    }
-                Capsule().stroke(palette.divider, lineWidth: 0.5)
-            )
 
-            if state.canShowClipboardEntry {
+    private var inputTabSwitcher: some View {
-                KeyboardClipboardMenuButton(
+        // 分段轨道固定在键盘水平中心，不受两侧 Logo / 剪贴板入口影响。
-                    palette: palette,
+        HStack(spacing: 2) {
-                    action: state.openClipboardPanel
+            ForEach(KeyboardInputTab.allCases, id: \.self) { tab in
-                )
+                tabButton(tab)
-                .equatable()
             }
         }
+        .padding(2)
+        .background(tabTrackFill, in: Capsule())
+        .overlay(
+            Capsule().stroke(palette.divider, lineWidth: 0.5)
+        )
     }
 
     private func tabButton(_ tab: KeyboardInputTab) -> some View {
         let selected = isSelected(tab)
-        let width: CGFloat = tab == .english || tab == .ai ? 34 : 42
 
         return Button {
             withAnimation(Motion.soft) {
                 select(tab)
             }
         } label: {
-            tabLabel(tab, selected: selected, width: width)
+            tabLabel(tab, selected: selected)
         }
         .buttonStyle(TopControlPressStyle(pressedFill: pressedFill))
         .disabled(tab != .voice && !state.canEnterTypingSurface)
@@ -157,13 +165,11 @@ struct KeyboardTopControls: View {
     @ViewBuilder
     private func tabLabel(
         _ tab: KeyboardInputTab,
-        selected: Bool,
+        selected: Bool
-        width: CGFloat
     ) -> some View {
-        let label = Text(tab.title)
+        let label = tabContent(tab, selected: selected)
-            .font(.system(size: 12, weight: selected ? .semibold : .medium))
             .foregroundStyle(selected ? palette.textPrimary : palette.textSecondary)
-            .frame(width: width, height: 30)
+            .frame(width: KeyboardTopBarMetrics.inputTabWidth, height: 30)
 
         if selected {
             let namespace = sharedSelectionNamespace ?? fallbackSelectionNamespace
@@ -179,6 +185,20 @@ struct KeyboardTopControls: View {
         }
     }
 
+    @ViewBuilder
+    private func tabContent(_ tab: KeyboardInputTab, selected: Bool) -> some View {
+        if tab == .ai {
+            Image(systemName: "sparkle")
+                .font(.system(size: 15, weight: selected ? .semibold : .medium))
+        } else if tab == .voice {
+            Image(systemName: "waveform.mid")
+                .font(.system(size: 15, weight: selected ? .semibold : .medium))
+        } else {
+            Text(tab.title)
+                .font(.system(size: 12, weight: selected ? .semibold : .medium))
+        }
+    }
+
     private func tabOpacity(_ tab: KeyboardInputTab) -> Double {
         guard tab != .voice, !state.canEnterTypingSurface else { return 1 }
         if case .recording = state.phase {
@@ -191,10 +211,10 @@ struct KeyboardTopControls: View {
         colorScheme == .dark ? Color(white: 0.22) : Color(white: 0.84)
     }
 
-    /// 分段轨道底色：不透明，且与选中键面（NativeKeyboardKeyColors.fill）拉开明度，
+    /// 分段轨道底色与选中键面（NativeKeyboardKeyColors.fill）拉开明度；
-    /// 深色下压暗、浅色下提亮，让滑动的选中项始终清晰可辨。
+    /// 浅色模式叠加半透明黑色，在不同宿主键盘底色上维持可见对比。
     private var tabTrackFill: Color {
-        colorScheme == .dark ? Color(white: 0.12) : Color(white: 0.87)
+        colorScheme == .dark ? Color(white: 0.12) : Color.black.opacity(0.12)
     }
 
     private func isSelected(_ tab: KeyboardInputTab) -> Bool {

OSGKeyboardExtTests/EnglishTypingTests.swift

@@ -10,18 +10,42 @@ final class EnglishTypingTests: XCTestCase {
     func testLexiconLoadsAndCompletesPrefix() {
         let lexicon = EnglishLexicon()
         lexicon.prepare()
-        XCTAssertGreaterThan(lexicon.wordCount, 1_000)
+        XCTAssertTrue(lexicon.isLoaded)
+        XCTAssertGreaterThan(lexicon.wordCount, 10_000)
         let hits = lexicon.completions(prefix: "hel", limit: 5)
         XCTAssertTrue(hits.contains("hello") || hits.contains("help") || hits.contains("held"))
+        lexicon.unload()
+        XCTAssertFalse(lexicon.isLoaded)
+        XCTAssertEqual(lexicon.wordCount, 0)
+    }
+
+    @MainActor
+    func testChineseTypingDoesNotLoadEnglishLexicon() {
+        EnglishLexicon.shared.unload()
+        let typing = TypingSessionController()
+        _ = typing.setLanguage(.chinese)
+        typing.enterTypingMode()
+        XCTAssertFalse(EnglishLexicon.shared.isLoaded)
+        _ = typing.setLanguage(.english)
+        XCTAssertTrue(EnglishLexicon.shared.isLoaded)
+        _ = typing.setLanguage(.chinese)
+        XCTAssertFalse(EnglishLexicon.shared.isLoaded)
+        typing.leaveTypingMode()
     }
 
     func testCorrectionFindsNearbyWord() {
-        let lexicon = EnglishLexicon()
+        let engine = EnglishSuggestionEngine()
-        lexicon.prepare()
+        engine.prepare()
-        // "teh" is a classic typo for "the".
+        // "teh" leaks into web unigrams; the engine must still treat it as a typo.
-        let correction = lexicon.bestCorrection(for: "teh")
+        let decision = engine.correctionDecision(
-        XCTAssertEqual(correction, "the")
+            for: "teh",
-        XCTAssertNil(lexicon.bestCorrection(for: "the"))
+            personalTerms: [],
+            learnedBoosts: [:]
+        )
+        XCTAssertEqual(decision?.replacement, "the")
+        XCTAssertNil(
+            engine.correctionDecision(for: "the", personalTerms: [], learnedBoosts: [:])
+        )
     }
 
     func testSuggestionEngineSkipsPersonalDictionaryTypos() {
@@ -45,7 +69,22 @@ final class EnglishTypingTests: XCTestCase {
                 learnedBoosts: [:]
             )
         )
-        XCTAssertEqual(composition.candidates.first?.text, "OSGKeyboard")
+        XCTAssertEqual(composition.candidates.first?.role, .verbatim)
+        XCTAssertEqual(composition.candidates.first?.text, "osg")
+        XCTAssertTrue(composition.candidates.contains { $0.text == "OSGKeyboard" })
+    }
+
+    func testSuggestionEngineReturnsNoCandidatesWithoutCurrentWord() {
+        let engine = EnglishSuggestionEngine()
+        engine.prepare()
+        let composition = engine.compositionWhileTyping(
+            EnglishSuggestionContext(
+                previousWord: "hello",
+                personalTerms: ["OSGKeyboard"]
+            )
+        )
+
+        XCTAssertEqual(composition, .empty)
     }
 
     func testAutocapitalizationAtFieldStartAndAfterSentence() {
@@ -208,7 +247,7 @@ final class EnglishTypingTests: XCTestCase {
         XCTAssertEqual(typing.composition.preedit, "boa")
         XCTAssertTrue(
             typing.composition.candidates.contains {
-                $0.text.compare("boat", options: .caseInsensitive) == .orderedSame
+                $0.text.compare("board", options: .caseInsensitive) == .orderedSame
             }
         )
 
@@ -232,14 +271,14 @@ final class EnglishTypingTests: XCTestCase {
             preceding += output.text
             typing.syncAutocapitalization(accountingForInsert: output.text)
         }
-        guard let boatIndex = typing.composition.candidates.firstIndex(where: {
+        guard let boardIndex = typing.composition.candidates.firstIndex(where: {
-            $0.text.compare("boat", options: .caseInsensitive) == .orderedSame
+            $0.text.compare("board", options: .caseInsensitive) == .orderedSame
         }) else {
-            return XCTFail("expected boat completion")
+            return XCTFail("expected board completion")
         }
 
         preceding = "board"
-        let output = typing.selectCandidate(at: boatIndex)
+        let output = typing.selectCandidate(at: boardIndex)
 
         XCTAssertEqual(output, .none)
         XCTAssertEqual(typing.composition.preedit.lowercased(), "board")
@@ -283,22 +322,17 @@ final class EnglishTypingTests: XCTestCase {
 
     @MainActor
     func testAutocorrectUndoRestoresOriginal() {
-        let typing = TypingSessionController()
+        let typing = makeIsolatedEnglishSession(suite: "english.undo.test")
-        typing.suggestionsEnabled = true
-        _ = typing.setLanguage(.english)
-        typing.enterTypingMode()
 
         for ch in ["t", "e", "h"] {
             _ = typing.handleKey(ch)
         }
         let spaced = typing.handleSpace()
-        // Either corrected to "the " or left as-is if lexicon missing in test bundle.
+        XCTAssertEqual(spaced.deleteCount, 3)
-        if spaced.deleteCount > 0 {
+        XCTAssertTrue(spaced.text.hasPrefix("the"))
-            XCTAssertTrue(spaced.text.hasPrefix("the"))
+        let undone = typing.handleKey("⌫")
-            let undone = typing.handleKey("⌫")
+        XCTAssertEqual(undone.text, "teh")
-            XCTAssertEqual(undone.text, "teh")
+        XCTAssertEqual(undone.deleteCount, spaced.text.count)
-            XCTAssertEqual(undone.deleteCount, spaced.text.count)
-        }
     }
 
     @MainActor
@@ -370,6 +404,177 @@ final class EnglishTypingTests: XCTestCase {
         XCTAssertFalse(PeriodShortcut.shouldArm(afterSpaceFollowing: "hello "))
     }
 
+    @MainActor
+    func testEnglishQuickTypePutsVerbatimFirstAndMarksCorrection() {
+        let typing = makeIsolatedEnglishSession(suite: "english.quicktype.bar.test")
+
+        for character in ["t", "e", "h"] {
+            _ = typing.handleKey(character)
+        }
+        XCTAssertEqual(typing.composition.candidates.first?.role, .verbatim)
+        XCTAssertEqual(
+            typing.composition.candidates.first?.text.lowercased(),
+            "teh"
+        )
+        XCTAssertTrue(
+            typing.composition.candidates.contains {
+                $0.role == .correction && $0.text.lowercased() == "the"
+            }
+        )
+        XCTAssertLessThanOrEqual(typing.composition.candidates.count, EnglishSuggestionEngine.slotCount)
+    }
+
+    @MainActor
+    func testEnglishSpaceAppliesCorrectionSlotOnly() {
+        let typing = makeIsolatedEnglishSession(suite: "english.quicktype.space.test")
+
+        for character in ["t", "e", "h"] {
+            _ = typing.handleKey(character)
+        }
+        let spaced = typing.handleSpace()
+        XCTAssertEqual(spaced.deleteCount, 3)
+        XCTAssertTrue(spaced.text.lowercased().hasPrefix("the"))
+        XCTAssertTrue(typing.composition.candidates.isEmpty)
+    }
+
+    @MainActor
+    func testEnglishSpaceKeepsVerbatimWhenNoCorrection() {
+        let typing = makeIsolatedEnglishSession(suite: "english.quicktype.verbatim.test")
+
+        for character in ["h", "e", "l"] {
+            _ = typing.handleKey(character)
+        }
+        let spaced = typing.handleSpace()
+        XCTAssertEqual(spaced, .insert(" "))
+        XCTAssertTrue(typing.composition.candidates.isEmpty)
+    }
+
+    func testTitleCaseNamesAreNotAutocorrected() {
+        let engine = EnglishSuggestionEngine()
+        engine.prepare()
+        XCTAssertNil(engine.correctionDecision(for: "Rocky", personalTerms: [], learnedBoosts: [:]))
+        XCTAssertNil(engine.correctionDecision(for: "Wang", personalTerms: [], learnedBoosts: [:]))
+        XCTAssertNil(engine.correctionDecision(for: "Chen", personalTerms: [], learnedBoosts: [:]))
+        XCTAssertNil(engine.correctionDecision(for: "Li", personalTerms: [], learnedBoosts: [:]))
+    }
+
+    func testTitleCaseTranspositionStillCorrects() {
+        let engine = EnglishSuggestionEngine()
+        engine.prepare()
+        XCTAssertEqual(
+            engine.correctionDecision(for: "Teh", personalTerms: [], learnedBoosts: [:])?.replacement,
+            "The"
+        )
+    }
+
+    func testProximityCorrectsAdjacentKeyTypos() {
+        let engine = EnglishSuggestionEngine()
+        engine.prepare()
+        XCTAssertEqual(
+            engine.correctionDecision(for: "gppd", personalTerms: [], learnedBoosts: [:])?.replacement,
+            "good"
+        )
+    }
+
+    func testRealWordFormIsNotCorrectedToFrom() {
+        let engine = EnglishSuggestionEngine()
+        engine.prepare()
+        XCTAssertNil(engine.correctionDecision(for: "form", personalTerms: [], learnedBoosts: [:]))
+    }
+
+    func testSupplementaryLexiconBlocksAutocorrect() {
+        let engine = EnglishSuggestionEngine()
+        engine.prepare()
+        XCTAssertNil(
+            engine.correctionDecision(
+                for: "teh",
+                personalTerms: [],
+                learnedBoosts: [:],
+                systemWords: ["teh"]
+            )
+        )
+    }
+
+    func testLearnedDefenseBlocksAutocorrect() {
+        let engine = EnglishSuggestionEngine()
+        engine.prepare()
+        XCTAssertNil(
+            engine.correctionDecision(
+                for: "teh",
+                personalTerms: [],
+                learnedBoosts: ["teh": 5]
+            )
+        )
+    }
+
+    @MainActor
+    func testAutocorrectDoesNotBoostReplacement() {
+        let suite = "english.learning.polarity.test"
+        let defaults = UserDefaults(suiteName: suite)!
+        defaults.removePersistentDomain(forName: suite)
+        let store = EnglishLearningStore(defaults: defaults)
+        let typing = TypingSessionController(learningStore: store)
+        typing.suggestionsEnabled = true
+        _ = typing.setLanguage(.english)
+        typing.enterTypingMode()
+
+        for character in ["t", "e", "h"] {
+            _ = typing.handleKey(character)
+        }
+        _ = typing.handleSpace()
+        XCTAssertEqual(store.boost(for: "the"), 0)
+        XCTAssertEqual(store.boost(for: "teh"), 0)
+    }
+
+    @MainActor
+    func testRejectingAutocorrectLearnsOriginal() {
+        let suite = "english.learning.defense.test"
+        let defaults = UserDefaults(suiteName: suite)!
+        defaults.removePersistentDomain(forName: suite)
+        let store = EnglishLearningStore(defaults: defaults)
+        let typing = TypingSessionController(learningStore: store)
+        typing.suggestionsEnabled = true
+        _ = typing.setLanguage(.english)
+        typing.enterTypingMode()
+
+        for character in ["t", "e", "h"] {
+            _ = typing.handleKey(character)
+        }
+        let spaced = typing.handleSpace()
+        if spaced.deleteCount > 0 {
+            _ = typing.handleKey("⌫")
+            XCTAssertGreaterThanOrEqual(store.boost(for: "teh"), 5)
+        }
+    }
+
+    func testQWERTYNeighborsIncludeDiagonals() {
+        let aroundG = EnglishQWERTYProximity.neighbors(of: "g", includingSelf: true)
+        XCTAssertTrue(aroundG.contains("t"))
+        XCTAssertTrue(aroundG.contains("f"))
+        XCTAssertTrue(aroundG.contains("h"))
+        XCTAssertTrue(aroundG.contains("b"))
+        XCTAssertFalse(aroundG.contains("q"))
+    }
+
+    func testBigramsPredictNextWords() {
+        let lexicon = EnglishLexicon()
+        lexicon.prepare()
+        let next = lexicon.nextWords(after: "thank", limit: 4)
+        XCTAssertTrue(next.contains("you"))
+    }
+
+    @MainActor
+    private func makeIsolatedEnglishSession(suite: String) -> TypingSessionController {
+        let defaults = UserDefaults(suiteName: suite)!
+        defaults.removePersistentDomain(forName: suite)
+        let store = EnglishLearningStore(defaults: defaults)
+        let typing = TypingSessionController(learningStore: store)
+        typing.suggestionsEnabled = true
+        _ = typing.setLanguage(.english)
+        typing.enterTypingMode()
+        return typing
+    }
+
     @MainActor
     private func apply(_ typing: TypingSessionController, _ output: TypingOutput) {
         typing.syncAutocapitalization(

OSGKeyboardShared/Models/AIUserSkill.swift

@@ -14,7 +14,8 @@ public struct AIUserSkill: Codable, Equatable, Identifiable, Sendable {
     public var summary: String
     public var systemImage: String
     public var prompt: String
-    public var shortcutICloudURL: URL
+    /// Optional iCloud share URL. Nil means the skill only transforms text.
+    public var shortcutICloudURL: URL?
     /// Name used by `shortcuts://run-shortcut?name=`. Independent of `name`.
     public var shortcutName: String
     /// Per-skill reasoning. Built-in skills are always off; custom defaults off.
@@ -28,8 +29,8 @@ public struct AIUserSkill: Codable, Equatable, Identifiable, Sendable {
         summary: String = "",
         systemImage: String = AIUserSkillLimits.defaultSystemImage,
         prompt: String,
-        shortcutICloudURL: URL,
+        shortcutICloudURL: URL? = nil,
-        shortcutName: String,
+        shortcutName: String = "",
         thinkingEnabled: Bool = false,
         createdAt: Date = Date(),
         updatedAt: Date? = nil
@@ -49,15 +50,16 @@ public struct AIUserSkill: Codable, Equatable, Identifiable, Sendable {
     public var isUserCreated: Bool { id.hasPrefix("user.") }
 
     public func asClipboardSkill() -> AIClipboardSkill {
-        AIClipboardSkill(
+        let exportsToShortcut = shortcutICloudURL != nil
+        return AIClipboardSkill(
             id: id,
             systemImage: systemImage,
             titleKey: "",
             cardTitleKey: "",
             descriptionKey: "",
-            kind: .export,
+            kind: exportsToShortcut ? .export : .transform,
             isDefault: false,
-            shortcutName: shortcutName,
+            shortcutName: exportsToShortcut ? shortcutName : nil,
             shortcutICloudURL: shortcutICloudURL,
             customName: name,
             customSummary: summary,
@@ -149,9 +151,13 @@ public struct AIUserSkillCatalog: Codable, Equatable, Sendable {
                 maximum: AIUserSkillLimits.maximumPromptCharacters
             )
         }
-        guard !shortcutName.isEmpty else { throw AIUserSkillValidationError.emptyShortcutName }
+        if let shortcutURL = skill.shortcutICloudURL {
-        guard AIShortcutShareLink.isValid(skill.shortcutICloudURL) else {
+            guard !shortcutName.isEmpty else {
-            throw AIUserSkillValidationError.invalidShortcutLink
+                throw AIUserSkillValidationError.emptyShortcutName
+            }
+            guard AIShortcutShareLink.isValid(shortcutURL) else {
+                throw AIUserSkillValidationError.invalidShortcutLink
+            }
         }
         guard !icon.isEmpty else { throw AIUserSkillValidationError.emptyIcon }
 

OSGKeyboardShared/Services/AIAgentSkillLayoutStore.swift

@@ -130,13 +130,26 @@ public final class AIAgentSkillLayoutStore: ObservableObject {
     }
 
     public func saveUserSkill(_ skill: AIUserSkill) throws {
-        let previousURL = userCatalog.skill(id: skill.id)?.shortcutICloudURL
+        let previousSkill = userCatalog.skill(id: skill.id)
+        let previousURL = previousSkill?.shortcutICloudURL
+        let previousLayout = layout.sanitized(catalog: mergedCatalog)
         var catalog = userCatalog
         try catalog.upsert(skill)
         commitUserCatalog(catalog)
-        if previousURL != nil, previousURL != skill.shortcutICloudURL {
+        guard previousSkill != nil, previousURL != skill.shortcutICloudURL else {
-            dropShortcutConfirmation(for: skill.id)
+            return
         }
+        let keepsKeyboardSlot = skill.shortcutICloudURL == nil
+        commitLayout(
+            AIAgentSkillLayout(
+                enabledIDs: keepsKeyboardSlot
+                    ? previousLayout.enabledIDs
+                    : previousLayout.enabledIDs.filter { $0 != skill.id },
+                confirmedShortcutIDs: previousLayout.confirmedShortcutIDs.filter {
+                    $0 != skill.id
+                }
+            )
+        )
     }
 
     public func deleteUserSkill(id: String) {
@@ -152,16 +165,6 @@ public final class AIAgentSkillLayoutStore: ObservableObject {
         )
     }
 
-    private func dropShortcutConfirmation(for id: String) {
-        let current = layout.sanitized(catalog: mergedCatalog)
-        commitLayout(
-            AIAgentSkillLayout(
-                enabledIDs: current.enabledIDs.filter { $0 != id },
-                confirmedShortcutIDs: current.confirmedShortcutIDs.filter { $0 != id }
-            )
-        )
-    }
-
     private func commitLayout(_ layout: AIAgentSkillLayout) {
         persistLayout(layout)
         self.layout = loadLayout()

OSGKeyboardShared/Typing/EnglishLexicon.swift

@@ -2,180 +2,359 @@
 // OSGKeyboard · Shared
 //
 // Offline English word list + bigrams for the typing extension.
-// Loaded once, kept compact for the keyboard RSS budget.
+// The 40k-word table is a mmap'd binary (`english_lexicon.bin`); dirty heap
+// stays near zero until a lookup materializes a handful of result strings.
+// TSV files in the repo are the build input, not the runtime format.
 
 import Foundation
 
+public struct EnglishScoredCorrection: Equatable, Sendable {
+    public var word: String
+    public var spatialCost: Int
+    public var frequency: Int
+    public var isTransposition: Bool
+    public var isShortening: Bool
+
+    public init(
+        word: String,
+        spatialCost: Int,
+        frequency: Int,
+        isTransposition: Bool,
+        isShortening: Bool
+    ) {
+        self.word = word
+        self.spatialCost = spatialCost
+        self.frequency = frequency
+        self.isTransposition = isTransposition
+        self.isShortening = isShortening
+    }
+}
+
 /// Ranked English lexicon used by autocomplete / autocorrect / next-word.
 public final class EnglishLexicon: @unchecked Sendable {
     public static let shared = EnglishLexicon()
 
-    /// Lowercased word → relative frequency (higher is more common).
+    private var mapped: Data?
-    private var frequencies: [String: Int] = [:]
+    private var header: FileHeader?
-    /// Sorted lowercased words for prefix binary search.
-    private var sortedWords: [String] = []
-    /// previous(lower) → next-word candidates (lower).
-    private var bigrams: [String: [String]] = [:]
     private var loaded = false
     private let lock = NSLock()
 
     public init() {}
 
+    /// True after a successful mmap. Tests use this to prove Chinese typing
+    /// does not pull the English table into the extension.
+    public var isLoaded: Bool {
+        lock.lock()
+        defer { lock.unlock() }
+        return loaded
+    }
+
     public func prepare() {
         lock.lock()
         defer { lock.unlock() }
         guard !loaded else { return }
-        loadLexicon()
+        loadMappedLexicon()
-        loadBigrams()
-        loaded = true
     }
 
-    /// Release in-memory tables when leaving the typing surface (jetsam recovery).
+    /// Release the mapped file when leaving English / the typing surface.
     public func unload() {
         lock.lock()
         defer { lock.unlock() }
-        frequencies.removeAll(keepingCapacity: false)
+        mapped = nil
-        sortedWords.removeAll(keepingCapacity: false)
+        header = nil
-        bigrams.removeAll(keepingCapacity: false)
         loaded = false
     }
 
     public var wordCount: Int {
-        prepareIfNeeded()
+        lock.lock()
-        return sortedWords.count
+        defer { lock.unlock() }
+        return header?.unigramCount ?? 0
     }
 
     public func frequency(of word: String) -> Int {
-        prepareIfNeeded()
+        withMap { buf, header in
-        return frequencies[word.lowercased()] ?? 0
+            guard let index = lookupIndex(asciiLowered(word), header: header, buf: buf) else {
+                return 0
+            }
+            return frequency(at: index, header: header, buf: buf)
+        } ?? 0
     }
 
     public func contains(_ word: String) -> Bool {
-        prepareIfNeeded()
+        withMap { buf, header in
-        return frequencies[word.lowercased()] != nil
+            lookupIndex(asciiLowered(word), header: header, buf: buf) != nil
+        } ?? false
+    }
+
+    /// Highest-frequency unigrams, for next-word fallback when no bigram hits.
+    public func topWords(limit: Int = 6) -> [String] {
+        guard limit > 0 else { return [] }
+        return withMap { buf, header in
+            let count = min(limit, header.unigramCount)
+            var words: [String] = []
+            words.reserveCapacity(count)
+            for rank in 0..<count {
+                let index = Int(
+                    readU16(buf, header.freqRankOffset + rank * 2)
+                )
+                guard index < header.unigramCount else { continue }
+                if let word = string(at: index, header: header, buf: buf) {
+                    words.append(word)
+                }
+            }
+            return words
+        } ?? []
     }
 
     /// Prefix completions, highest frequency first.
     public func completions(prefix: String, limit: Int = 8) -> [String] {
-        prepareIfNeeded()
+        let needle = asciiLowered(prefix)
-        let needle = prefix.lowercased()
         guard !needle.isEmpty, limit > 0 else { return [] }
-
+        return withMap { buf, header in
-        var results: [(String, Int)] = []
+            var scored: [(Int, Int)] = []
-        var index = lowerBound(needle)
+            var index = lowerBound(needle, header: header, buf: buf)
-        while index < sortedWords.count {
+            while index < header.unigramCount {
-            let word = sortedWords[index]
+                guard let bytes = wordBytes(at: index, header: header, buf: buf) else { break }
-            guard word.hasPrefix(needle) else { break }
+                guard hasPrefix(bytes, needle) else { break }
-            if word != needle {
+                if !bytesEqual(bytes, needle) {
-                results.append((word, frequencies[word] ?? 0))
+                    scored.append((index, frequency(at: index, header: header, buf: buf)))
+                }
+                index += 1
+                // Soft cap scan to keep keystroke path cheap.
+                if scored.count >= limit * 8 { break }
             }
-            index += 1
+            scored.sort { lhs, rhs in
-            // Soft cap scan to keep keystroke path cheap.
+                if lhs.1 != rhs.1 { return lhs.1 > rhs.1 }
-            if results.count >= limit * 8 { break }
+                return lhs.0 < rhs.0
-        }
+            }
-        results.sort { lhs, rhs in
+            return scored.prefix(limit).compactMap { pair in
-            if lhs.1 != rhs.1 { return lhs.1 > rhs.1 }
+                string(at: pair.0, header: header, buf: buf)
-            return lhs.0 < rhs.0
+            }
-        }
+        } ?? []
-        return Array(results.prefix(limit).map(\.0))
     }
 
-    /// Best edit-distance ≤ 2 correction, or nil when the typed word is fine.
+    /// Nearby words scored by QWERTY proximity + frequency. Does not decide
-    /// Uses Damerau–Levenshtein so adjacent swaps (teh → the) count as 1.
+    /// whether autocorrect should fire — the suggestion engine does.
-    /// Scans only same-initial-letter candidates (not the full frequency table).
+    public func scoredCorrections(for typed: String, limit: Int = 6) -> [EnglishScoredCorrection] {
-    public func bestCorrection(for typed: String) -> String? {
+        let needle = asciiLowered(typed)
-        prepareIfNeeded()
+        guard needle.count >= 3, let firstByte = needle.first, limit > 0 else { return [] }
-        let needle = typed.lowercased()
+        let first = Character(UnicodeScalar(firstByte))
-        guard needle.count >= 2, let first = needle.first else { return nil }
+        var initials = Set(EnglishQWERTYProximity.neighbors(of: first, includingSelf: true))
-        if frequencies[needle] != nil { return nil }
+        initials.insert(first)
 
-        var best: (word: String, distance: Int, freq: Int)?
+        return withMap { buf, header in
-        var index = lowerBound(String(first))
+            var best: [ScoredIndex] = []
-        while index < sortedWords.count {
+            best.reserveCapacity(limit)
-            let word = sortedWords[index]
+            for initial in initials {
-            guard word.first == first else { break }
+                guard let letter = initial.asciiLetterIndex else { continue }
-            defer { index += 1 }
+                let rangeOffset = header.initialOffset + letter * 4
-            guard abs(word.count - needle.count) <= 2 else { continue }
+                let start = Int(readU16(buf, rangeOffset))
-            let freq = frequencies[word] ?? 0
+                let count = Int(readU16(buf, rangeOffset + 2))
-            let distance = damerauLevenshtein(needle, word, max: 2)
+                guard start >= 0, count >= 0, start + count <= header.unigramCount else { continue }
-            guard distance > 0, distance <= 2 else { continue }
+                for index in start..<(start + count) {
-            if let current = best {
+                    guard let bytes = wordBytes(at: index, header: header, buf: buf) else { continue }
-                if distance < current.distance
+                    let delta = abs(bytes.count - needle.count)
-                    || (distance == current.distance && freq > current.freq) {
+                    guard delta <= 2, !bytesEqual(bytes, needle) else { continue }
-                    best = (word, distance, freq)
+                    guard let alignment = EnglishQWERTYProximity.align(
+                        typedASCII: needle,
+                        candidateASCII: bytes
+                    ) else { continue }
+                    guard alignment.cost > 0 else { continue }
+                    insertBest(
+                        ScoredIndex(
+                            index: index,
+                            spatialCost: alignment.cost,
+                            frequency: frequency(at: index, header: header, buf: buf),
+                            isTransposition: alignment.isTransposition,
+                            isShortening: alignment.isShortening
+                        ),
+                        into: &best,
+                        limit: limit
+                    )
                 }
-            } else {
-                best = (word, distance, freq)
             }
-        }
+            return best.compactMap { scored in
-        guard let best else { return nil }
+                guard let word = string(at: scored.index, header: header, buf: buf) else {
-        // Distance-2 corrections need a common word so rare near-misses don't win.
+                    return nil
-        if best.distance == 2, best.freq < 200 { return nil }
+                }
-        return best.word
+                return EnglishScoredCorrection(
+                    word: word,
+                    spatialCost: scored.spatialCost,
+                    frequency: scored.frequency,
+                    isTransposition: scored.isTransposition,
+                    isShortening: scored.isShortening
+                )
+            }
+        } ?? []
+    }
+
+    /// Best proximity correction, or nil when the typed word is already known.
+    public func bestCorrection(for typed: String) -> String? {
+        if contains(typed) { return nil }
+        return scoredCorrections(for: typed, limit: 1).first?.word
     }
 
     public func nextWords(after previous: String, limit: Int = 6) -> [String] {
-        prepareIfNeeded()
+        guard limit > 0 else { return [] }
-        let key = previous.lowercased()
+        let needle = asciiLowered(previous)
-        guard let list = bigrams[key] else { return [] }
+        return withMap { buf, header in
-        return Array(list.prefix(limit))
+            guard let prevIndex = lookupIndex(needle, header: header, buf: buf) else {
+                return []
+            }
+            guard let group = lookupBigramGroup(prevIndex: prevIndex, header: header, buf: buf) else {
+                return []
+            }
+            let count = min(limit, group.nextCount)
+            var words: [String] = []
+            words.reserveCapacity(count)
+            for offset in 0..<count {
+                let index = Int(readU16(buf, header.bigramNextOffset + (group.firstNext + offset) * 2))
+                if let word = string(at: index, header: header, buf: buf) {
+                    words.append(word)
+                }
+            }
+            return words
+        } ?? []
     }
 
-    // MARK: - Private
+    // MARK: - Mapped file
 
-    private func prepareIfNeeded() {
+    private struct FileHeader {
-        if !loaded { prepare() }
+        var unigramCount: Int
+        var bigramGroupCount: Int
+        var stringPoolOffset: Int
+        var stringPoolSize: Int
+        var unigramOffset: Int
+        var freqRankOffset: Int
+        var initialOffset: Int
+        var bigramIndexOffset: Int
+        var bigramNextOffset: Int
+        var fileSize: Int
+
+        static let magic = "OSGENG01"
+        static let version = 1
+        static let headerSize = 64
+        static let initialCount = 26
+
+        static func parse(_ data: Data) -> FileHeader? {
+            guard data.count >= headerSize else { return nil }
+            return data.withUnsafeBytes { buf -> FileHeader? in
+                let magicBytes = UnsafeRawBufferPointer(rebasing: buf[0..<8])
+                let magic = String(bytes: magicBytes, encoding: .ascii)
+                guard magic == Self.magic else { return nil }
+                guard Int(readU32(buf, 8)) == version else { return nil }
+                let unigramCount = Int(readU32(buf, 12))
+                let bigramGroupCount = Int(readU32(buf, 16))
+                let stringPoolOffset = Int(readU32(buf, 20))
+                let stringPoolSize = Int(readU32(buf, 24))
+                let unigramOffset = Int(readU32(buf, 28))
+                let freqRankOffset = Int(readU32(buf, 32))
+                let initialOffset = Int(readU32(buf, 36))
+                let bigramIndexOffset = Int(readU32(buf, 40))
+                let bigramNextOffset = Int(readU32(buf, 44))
+                let fileSize = data.count
+
+                guard unigramCount >= 0, unigramCount <= 200_000 else { return nil }
+                guard bigramGroupCount >= 0, bigramGroupCount <= 100_000 else { return nil }
+                guard region(unigramOffset, unigramCount * 8, in: fileSize),
+                      region(freqRankOffset, unigramCount * 2, in: fileSize),
+                      region(initialOffset, initialCount * 4, in: fileSize),
+                      region(bigramIndexOffset, bigramGroupCount * 8, in: fileSize),
+                      region(stringPoolOffset, stringPoolSize, in: fileSize)
+                else {
+                    return nil
+                }
+
+                return FileHeader(
+                    unigramCount: unigramCount,
+                    bigramGroupCount: bigramGroupCount,
+                    stringPoolOffset: stringPoolOffset,
+                    stringPoolSize: stringPoolSize,
+                    unigramOffset: unigramOffset,
+                    freqRankOffset: freqRankOffset,
+                    initialOffset: initialOffset,
+                    bigramIndexOffset: bigramIndexOffset,
+                    bigramNextOffset: bigramNextOffset,
+                    fileSize: fileSize
+                )
+            }
+        }
+
+        private static func region(_ offset: Int, _ size: Int, in fileSize: Int) -> Bool {
+            offset >= 0 && size >= 0 && offset <= fileSize && size <= fileSize - offset
+        }
     }
 
-    private func loadLexicon() {
+    private struct ScoredIndex {
+        var index: Int
+        var spatialCost: Int
+        var frequency: Int
+        var isTransposition: Bool
+        var isShortening: Bool
+    }
+
+    private struct BigramGroup {
+        var nextCount: Int
+        var firstNext: Int
+    }
+
+    private func loadMappedLexicon() {
         guard let url = Bundle(for: EnglishLexicon.self)
-            .url(forResource: "english_lexicon", withExtension: "tsv", subdirectory: nil)
+            .url(forResource: "english_lexicon", withExtension: "bin")
-            ?? Bundle(for: EnglishLexicon.self)
+            ?? Bundle.main.url(forResource: "english_lexicon", withExtension: "bin")
-            .url(forResource: "english_lexicon", withExtension: "tsv")
-            ?? Bundle.main.url(forResource: "english_lexicon", withExtension: "tsv")
         else {
             return
         }
-        guard let data = try? String(contentsOf: url, encoding: .utf8) else { return }
+        // `.mappedIfSafe` keeps the 40k table on file-backed pages. Jetsam
-        var map: [String: Int] = [:]
+        // charges dirty heap, not these clean mapped pages.
-        for line in data.split(whereSeparator: \.isNewline) {
+        guard let data = try? Data(contentsOf: url, options: [.mappedIfSafe]),
-            let parts = line.split(separator: "\t", maxSplits: 1)
+              let parsed = FileHeader.parse(data)
-            guard parts.count == 2,
-                  let freq = Int(parts[1]) else { continue }
-            let word = String(parts[0]).lowercased()
-            guard !word.isEmpty else { continue }
-            map[word] = freq
-        }
-        frequencies = map
-        sortedWords = map.keys.sorted()
-    }
-
-    private func loadBigrams() {
-        guard let url = Bundle(for: EnglishLexicon.self)
-            .url(forResource: "english_bigrams", withExtension: "tsv")
-            ?? Bundle.main.url(forResource: "english_bigrams", withExtension: "tsv")
         else {
             return
         }
-        guard let data = try? String(contentsOf: url, encoding: .utf8) else { return }
+        mapped = data
-        var map: [String: [String]] = [:]
+        header = parsed
-        for line in data.split(whereSeparator: \.isNewline) {
+        loaded = true
-            let parts = line.split(separator: "\t", maxSplits: 1)
-            guard parts.count == 2 else { continue }
-            let prev = String(parts[0]).lowercased()
-            let nexts = parts[1].split(whereSeparator: \.isWhitespace).map { String($0).lowercased() }
-            guard !prev.isEmpty, !nexts.isEmpty else { continue }
-            map[prev] = nexts
-        }
-        bigrams = map
     }
 
-    private func lowerBound(_ prefix: String) -> Int {
+    private func withMap<T>(_ body: (UnsafeRawBufferPointer, FileHeader) -> T) -> T? {
+        lock.lock()
+        defer { lock.unlock() }
+        guard loaded, let data = mapped, let header else { return nil }
+        return data.withUnsafeBytes { buf in
+            body(buf, header)
+        }
+    }
+
+    private func lookupIndex(
+        _ needle: [UInt8],
+        header: FileHeader,
+        buf: UnsafeRawBufferPointer
+    ) -> Int? {
+        let index = lowerBound(needle, header: header, buf: buf)
+        guard index < header.unigramCount,
+              let bytes = wordBytes(at: index, header: header, buf: buf),
+              bytesEqual(bytes, needle)
+        else {
+            return nil
+        }
+        return index
+    }
+
+    private func lowerBound(
+        _ needle: [UInt8],
+        header: FileHeader,
+        buf: UnsafeRawBufferPointer
+    ) -> Int {
         var low = 0
-        var high = sortedWords.count
+        var high = header.unigramCount
         while low < high {
             let mid = (low + high) / 2
-            if sortedWords[mid] < prefix {
+            guard let bytes = wordBytes(at: mid, header: header, buf: buf) else {
+                high = mid
+                continue
+            }
+            if compare(bytes, needle) < 0 {
                 low = mid + 1
             } else {
                 high = mid
@@ -184,40 +363,140 @@ public final class EnglishLexicon: @unchecked Sendable {
         return low
     }
 
-    /// Damerau–Levenshtein with early exit when distance would exceed `max`.
+    private func lookupBigramGroup(
-    private func damerauLevenshtein(_ a: String, _ b: String, max: Int) -> Int {
+        prevIndex: Int,
-        let aChars = Array(a)
+        header: FileHeader,
-        let bChars = Array(b)
+        buf: UnsafeRawBufferPointer
-        let aCount = aChars.count
+    ) -> BigramGroup? {
-        let bCount = bChars.count
+        var low = 0
-        if abs(aCount - bCount) > max { return max + 1 }
+        var high = header.bigramGroupCount
-
+        while low < high {
-        var prevPrev = [Int](repeating: 0, count: bCount + 1)
+            let mid = (low + high) / 2
-        var prev = Array(0...bCount)
+            let midPrev = Int(readU16(buf, header.bigramIndexOffset + mid * 8))
-        for i in 1...aCount {
+            if midPrev < prevIndex {
-            var current = [Int](repeating: 0, count: bCount + 1)
+                low = mid + 1
-            current[0] = i
+            } else {
-            var rowMin = current[0]
+                high = mid
-            for j in 1...bCount {
-                let cost = aChars[i - 1] == bChars[j - 1] ? 0 : 1
-                var value = min(
-                    prev[j] + 1,
-                    current[j - 1] + 1,
-                    prev[j - 1] + cost
-                )
-                // Adjacent transposition
-                if i > 1, j > 1,
-                   aChars[i - 1] == bChars[j - 2],
-                   aChars[i - 2] == bChars[j - 1] {
-                    value = min(value, prevPrev[j - 2] + 1)
-                }
-                current[j] = value
-                rowMin = min(rowMin, value)
             }
-            if rowMin > max { return max + 1 }
-            prevPrev = prev
-            prev = current
         }
-        return prev[bCount]
+        guard low < header.bigramGroupCount else { return nil }
+        let offset = header.bigramIndexOffset + low * 8
+        guard Int(readU16(buf, offset)) == prevIndex else { return nil }
+        return BigramGroup(
+            nextCount: Int(readU16(buf, offset + 2)),
+            firstNext: Int(readU32(buf, offset + 4))
+        )
+    }
+
+    private func frequency(at index: Int, header: FileHeader, buf: UnsafeRawBufferPointer) -> Int {
+        Int(readU16(buf, header.unigramOffset + index * 8 + 6))
+    }
+
+    private func wordBytes(
+        at index: Int,
+        header: FileHeader,
+        buf: UnsafeRawBufferPointer
+    ) -> UnsafeBufferPointer<UInt8>? {
+        guard index >= 0, index < header.unigramCount else { return nil }
+        let record = header.unigramOffset + index * 8
+        let poolOff = Int(readU32(buf, record))
+        let length = Int(buf[record + 4])
+        let start = header.stringPoolOffset + poolOff
+        guard length >= 0,
+              start >= header.stringPoolOffset,
+              start + length <= header.stringPoolOffset + header.stringPoolSize,
+              start + length <= header.fileSize,
+              let base = buf.baseAddress
+        else {
+            return nil
+        }
+        return UnsafeBufferPointer(
+            start: base.advanced(by: start).assumingMemoryBound(to: UInt8.self),
+            count: length
+        )
+    }
+
+    private func string(
+        at index: Int,
+        header: FileHeader,
+        buf: UnsafeRawBufferPointer
+    ) -> String? {
+        guard let bytes = wordBytes(at: index, header: header, buf: buf) else { return nil }
+        return String(bytes: bytes, encoding: .ascii)
+    }
+
+    private func insertBest(_ scored: ScoredIndex, into best: inout [ScoredIndex], limit: Int) {
+        if let existing = best.firstIndex(where: { $0.index == scored.index }) {
+            if isOrderedBefore(scored, best[existing]) {
+                best[existing] = scored
+                best.sort(by: isOrderedBefore)
+            }
+            return
+        }
+        if best.count < limit {
+            best.append(scored)
+            best.sort(by: isOrderedBefore)
+            return
+        }
+        if let last = best.last, isOrderedBefore(scored, last) {
+            best[best.count - 1] = scored
+            best.sort(by: isOrderedBefore)
+        }
+    }
+
+    private func isOrderedBefore(_ lhs: ScoredIndex, _ rhs: ScoredIndex) -> Bool {
+        if lhs.spatialCost != rhs.spatialCost { return lhs.spatialCost < rhs.spatialCost }
+        if lhs.frequency != rhs.frequency { return lhs.frequency > rhs.frequency }
+        return lhs.index < rhs.index
+    }
+}
+
+private func readU16(_ buf: UnsafeRawBufferPointer, _ offset: Int) -> UInt16 {
+    UInt16(littleEndian: buf.loadUnaligned(fromByteOffset: offset, as: UInt16.self))
+}
+
+private func readU32(_ buf: UnsafeRawBufferPointer, _ offset: Int) -> UInt32 {
+    UInt32(littleEndian: buf.loadUnaligned(fromByteOffset: offset, as: UInt32.self))
+}
+
+private func asciiLowered(_ string: String) -> [UInt8] {
+    string.utf8.map { byte in
+        (byte >= 65 && byte <= 90) ? byte + 32 : byte
+    }
+}
+
+private func compare(_ word: UnsafeBufferPointer<UInt8>, _ needle: [UInt8]) -> Int {
+    let count = min(word.count, needle.count)
+    for index in 0..<count {
+        let left = word[index]
+        let right = needle[index]
+        if left < right { return -1 }
+        if left > right { return 1 }
+    }
+    if word.count < needle.count { return -1 }
+    if word.count > needle.count { return 1 }
+    return 0
+}
+
+private func hasPrefix(_ word: UnsafeBufferPointer<UInt8>, _ prefix: [UInt8]) -> Bool {
+    guard word.count >= prefix.count else { return false }
+    for index in prefix.indices where word[index] != prefix[index] {
+        return false
+    }
+    return true
+}
+
+private func bytesEqual(_ word: UnsafeBufferPointer<UInt8>, _ needle: [UInt8]) -> Bool {
+    guard word.count == needle.count else { return false }
+    for index in needle.indices where word[index] != needle[index] {
+        return false
+    }
+    return true
+}
+
+private extension Character {
+    var asciiLetterIndex: Int? {
+        guard let value = utf8.first, value >= 97, value <= 122 else { return nil }
+        return Int(value - 97)
     }
 }

OSGKeyboardShared/Typing/EnglishQWERTYProximity.swift

@@ -0,0 +1,210 @@
+// EnglishQWERTYProximity.swift
+// OSGKeyboard · Shared
+//
+// Spatial cost for English autocorrect. Adjacent (including diagonal) keys
+// are cheap; far substitutions are expensive. Inspired by AOSP LatinIME's
+// proximity weighting — formula only, no Android code.
+
+import Foundation
+
+public struct EnglishAlignment: Equatable, Sendable {
+    /// Weighted edit cost. `0` means identical.
+    public var cost: Int
+    public var isTransposition: Bool
+    public var isShortening: Bool
+}
+
+public enum EnglishQWERTYProximity: Sendable {
+    /// Two adjacent substitutions, or one farther miss, still eligible.
+    public static let maxAutocorrectCost = 34
+    public static let adjacentCost = 10
+    public static let nearCost = 22
+    public static let farCost = 34
+    public static let insDelCost = 18
+    public static let transpositionCost = 10
+
+    /// US QWERTY, staggered rows matching the on-screen letter grid.
+    private static let coordinates: [Character: (x: Double, y: Double)] = {
+        let rows: [[Character]] = [
+            Array("qwertyuiop"),
+            Array("asdfghjkl"),
+            Array("zxcvbnm")
+        ]
+        let offsets: [Double] = [0, 0.5, 1.5]
+        var map: [Character: (x: Double, y: Double)] = [:]
+        for (rowIndex, row) in rows.enumerated() {
+            let origin = offsets[rowIndex]
+            for (column, letter) in row.enumerated() {
+                map[letter] = (origin + Double(column), Double(rowIndex))
+            }
+        }
+        return map
+    }()
+
+    public static func neighbors(of letter: Character, includingSelf: Bool) -> [Character] {
+        let needle = Character(letter.lowercased())
+        guard let origin = coordinates[needle] else {
+            return includingSelf ? [needle] : []
+        }
+        var hits: [Character] = []
+        for (candidate, point) in coordinates {
+            let distance = chebyshev(origin, point)
+            if distance == 0 {
+                if includingSelf { hits.append(candidate) }
+            } else if distance <= 1.01 {
+                hits.append(candidate)
+            }
+        }
+        return hits
+    }
+
+    public static func keyDistance(_ a: Character, _ b: Character) -> Int {
+        let left = Character(a.lowercased())
+        let right = Character(b.lowercased())
+        if left == right { return 0 }
+        guard let origin = coordinates[left], let other = coordinates[right] else {
+            return farCost
+        }
+        let distance = chebyshev(origin, other)
+        if distance <= 1.01 { return adjacentCost }
+        if distance <= 2.01 { return nearCost }
+        return farCost
+    }
+
+    public static func align(typed: String, candidate: String) -> EnglishAlignment? {
+        let source = asciiLowered(typed)
+        let targetBytes = asciiLowered(candidate)
+        return targetBytes.withUnsafeBufferPointer { pointer in
+            align(typedASCII: source, candidateASCII: pointer)
+        }
+    }
+
+    /// Same cost model as `align(typed:candidate:)`, but the candidate stays in
+    /// a mapped file — no Swift `String` per scanned word.
+    public static func align(
+        typedASCII: [UInt8],
+        candidateASCII: UnsafeBufferPointer<UInt8>
+    ) -> EnglishAlignment? {
+        let source = typedASCII
+        let target = candidateASCII
+        let delta = abs(source.count - target.count)
+        guard delta <= 2 else { return nil }
+        if delta == 0, bytesEqual(source, target) {
+            return EnglishAlignment(cost: 0, isTransposition: false, isShortening: false)
+        }
+
+        if source.count == target.count, isAdjacentTransposition(source, target) {
+            return EnglishAlignment(
+                cost: transpositionCost,
+                isTransposition: true,
+                isShortening: false
+            )
+        }
+
+        if source.count == target.count {
+            var cost = 0
+            for index in source.indices {
+                cost += keyDistance(source[index], target[index])
+                if cost > maxAutocorrectCost { return nil }
+            }
+            return EnglishAlignment(
+                cost: cost,
+                isTransposition: false,
+                isShortening: false
+            )
+        }
+
+        let cost = bandedEditCost(source, target)
+        guard cost <= maxAutocorrectCost else { return nil }
+        return EnglishAlignment(
+            cost: cost,
+            isTransposition: false,
+            isShortening: target.count < source.count
+        )
+    }
+
+    private static func keyDistance(_ a: UInt8, _ b: UInt8) -> Int {
+        if a == b { return 0 }
+        guard a >= 97, a <= 122, b >= 97, b <= 122 else { return farCost }
+        return keyDistance(Character(UnicodeScalar(a)), Character(UnicodeScalar(b)))
+    }
+
+    private static func isAdjacentTransposition(
+        _ source: [UInt8],
+        _ target: UnsafeBufferPointer<UInt8>
+    ) -> Bool {
+        guard source.count == target.count, source.count >= 2 else { return false }
+        var mismatch = -1
+        for index in source.indices where source[index] != target[index] {
+            if mismatch == -1 {
+                mismatch = index
+            } else if index == mismatch + 1,
+                      source[mismatch] == target[index],
+                      source[index] == target[mismatch] {
+                for rest in (index + 1)..<source.count where source[rest] != target[rest] {
+                    return false
+                }
+                return true
+            } else {
+                return false
+            }
+        }
+        return false
+    }
+
+    /// Banded Levenshtein with proximity substitutions and a Damerau swap.
+    private static func bandedEditCost(
+        _ source: [UInt8],
+        _ target: UnsafeBufferPointer<UInt8>
+    ) -> Int {
+        let aCount = source.count
+        let bCount = target.count
+        var previous = Array(0...bCount).map { $0 * insDelCost }
+        var older = previous
+        for i in 1...aCount {
+            var current = [Int](repeating: 0, count: bCount + 1)
+            current[0] = i * insDelCost
+            var rowMin = current[0]
+            for j in 1...bCount {
+                let substitution = previous[j - 1] + keyDistance(source[i - 1], target[j - 1])
+                var value = min(
+                    previous[j] + insDelCost,
+                    current[j - 1] + insDelCost,
+                    substitution
+                )
+                if i > 1, j > 1,
+                   source[i - 1] == target[j - 2],
+                   source[i - 2] == target[j - 1] {
+                    value = min(value, older[j - 2] + transpositionCost)
+                }
+                current[j] = value
+                rowMin = min(rowMin, value)
+            }
+            if rowMin > maxAutocorrectCost { return maxAutocorrectCost + 1 }
+            older = previous
+            previous = current
+        }
+        return previous[bCount]
+    }
+
+    private static func asciiLowered(_ string: String) -> [UInt8] {
+        string.utf8.map { byte in
+            (byte >= 65 && byte <= 90) ? byte + 32 : byte
+        }
+    }
+
+    private static func bytesEqual(_ source: [UInt8], _ target: UnsafeBufferPointer<UInt8>) -> Bool {
+        guard source.count == target.count else { return false }
+        for index in source.indices where source[index] != target[index] {
+            return false
+        }
+        return true
+    }
+
+    private static func chebyshev(
+        _ a: (x: Double, y: Double),
+        _ b: (x: Double, y: Double)
+    ) -> Double {
+        max(abs(a.x - b.x), abs(a.y - b.y))
+    }
+}

OSGKeyboardShared/Typing/EnglishSuggestionEngine.swift

@@ -1,8 +1,8 @@
 // EnglishSuggestionEngine.swift
 // OSGKeyboard · Shared
 //
-// Builds TypingComposition for English: completions while composing,
+// Builds a 3-slot English QuickType board: verbatim / correction / completion
-// high-confidence corrections on commit, next-word predictions after.
+// (or next-word after commit). Space applies only the correction slot.
 
 import Foundation
 
@@ -12,19 +12,29 @@ public struct EnglishSuggestionContext: Sendable {
     public var personalTerms: [String]
     public var learnedBoosts: [String: Int]
     public var includeOriginalAfterCorrection: String?
+    /// Contacts / text replacements from `UILexicon`.
+    public var systemWords: [String]
+    public var systemCompletions: [String]
+    public var systemGuesses: [String]
 
     public init(
         currentWord: String = "",
         previousWord: String = "",
         personalTerms: [String] = [],
         learnedBoosts: [String: Int] = [:],
-        includeOriginalAfterCorrection: String? = nil
+        includeOriginalAfterCorrection: String? = nil,
+        systemWords: [String] = [],
+        systemCompletions: [String] = [],
+        systemGuesses: [String] = []
     ) {
         self.currentWord = currentWord
         self.previousWord = previousWord
         self.personalTerms = personalTerms
         self.learnedBoosts = learnedBoosts
         self.includeOriginalAfterCorrection = includeOriginalAfterCorrection
+        self.systemWords = systemWords
+        self.systemCompletions = systemCompletions
+        self.systemGuesses = systemGuesses
     }
 }
 
@@ -47,6 +57,10 @@ public struct EnglishCorrectionDecision: Equatable, Sendable {
 
 /// Pure ranking / candidate builder — no UITextDocumentProxy access.
 public struct EnglishSuggestionEngine: Sendable {
+    public static let slotCount = 3
+    /// In-vocabulary words only yield to a much more common transposition / neighbor.
+    public static let inVocabularyFrequencyGap = 250
+
     private let lexicon: EnglishLexicon
 
     public init(lexicon: EnglishLexicon = .shared) {
@@ -57,112 +71,238 @@ public struct EnglishSuggestionEngine: Sendable {
         lexicon.prepare()
     }
 
-    /// Suggestions while the user is mid-word.
+    /// Suggestions only while the user is actively typing an English word.
     public func compositionWhileTyping(_ context: EnglishSuggestionContext) -> TypingComposition {
         let prefix = context.currentWord
-        guard !prefix.isEmpty else {
+        guard !prefix.isEmpty else { return .empty }
-            return nextWordComposition(context)
+        return makeBoard(context).composition
-        }
-
-        var ranked: [(text: String, score: Int, id: String)] = []
-        var seen = Set<String>()
-
-        func append(_ raw: String, baseScore: Int, tag: String, preserveCase: Bool = false) {
-            let display = preserveCase ? raw : matchCase(of: prefix, to: raw)
-            let key = display.lowercased()
-            guard seen.insert(key).inserted else { return }
-            let boost = context.learnedBoosts[key] ?? 0
-            let personalBoost = context.personalTerms.contains { $0.lowercased() == key } ? 5_000 : 0
-            ranked.append((display, baseScore + boost + personalBoost, "\(tag)|\(key)"))
-        }
-
-        for term in context.personalTerms where term.lowercased().hasPrefix(prefix.lowercased())
-            && term.lowercased() != prefix.lowercased() {
-            append(term, baseScore: 8_000 + term.count, tag: "personal", preserveCase: true)
-        }
-
-        for word in lexicon.completions(prefix: prefix, limit: 12) {
-            append(word, baseScore: lexicon.frequency(of: word), tag: "complete")
-        }
-
-        ranked.sort { lhs, rhs in
-            if lhs.score != rhs.score { return lhs.score > rhs.score }
-            return lhs.text.count < rhs.text.count
-        }
-
-        let candidates = ranked.prefix(8).map {
-            TypingCandidate(id: $0.id, text: $0.text, engineIndex: 0)
-        }
-        return TypingComposition(preedit: prefix, candidates: Array(candidates))
     }
 
     /// Decide whether to autocorrect on space / punctuation.
     public func correctionDecision(
         for typed: String,
         personalTerms: [String],
-        learnedBoosts: [String: Int]
+        learnedBoosts: [String: Int],
+        previousWord: String = "",
+        systemWords: [String] = [],
+        systemGuesses: [String] = []
     ) -> EnglishCorrectionDecision? {
-        let trimmed = typed
+        let context = EnglishSuggestionContext(
-        guard trimmed.count >= 2 else { return nil }
+            currentWord: typed,
-        let lower = trimmed.lowercased()
+            previousWord: previousWord,
-
+            personalTerms: personalTerms,
-        if personalTerms.contains(where: { $0.lowercased() == lower }) { return nil }
+            learnedBoosts: learnedBoosts,
-        if (learnedBoosts[lower] ?? 0) >= 5 { return nil }
+            systemWords: systemWords,
-        if shouldSkipAutocorrect(trimmed) { return nil }
+            systemGuesses: systemGuesses
-        if lexicon.contains(lower) { return nil }
-
-        guard let correction = lexicon.bestCorrection(for: lower) else { return nil }
-        // Personal dictionary wins over lexicon corrections.
-        if personalTerms.contains(where: { $0.lowercased() == correction }) {
-            return EnglishCorrectionDecision(original: trimmed, replacement: matchCase(of: trimmed, to: correction))
-        }
-        let typedBoost = learnedBoosts[lower] ?? 0
-        let correctionFreq = lexicon.frequency(of: correction) + (learnedBoosts[correction] ?? 0)
-        // High-confidence gate: correction must clearly beat defending the typo.
-        guard correctionFreq >= 80, correctionFreq > typedBoost + 40 else { return nil }
-        return EnglishCorrectionDecision(
-            original: trimmed,
-            replacement: matchCase(of: trimmed, to: correction)
         )
+        return makeBoard(context).decision
     }
 
     public func nextWordComposition(_ context: EnglishSuggestionContext) -> TypingComposition {
-        var ranked: [(text: String, score: Int, id: String)] = []
+        var ranked: [(text: String, score: Int, role: TypingCandidateRole, quoted: Bool)] = []
         var seen = Set<String>()
 
-        func append(_ raw: String, baseScore: Int, tag: String) {
+        func append(_ raw: String, baseScore: Int, role: TypingCandidateRole, quoted: Bool = false) {
             let key = raw.lowercased()
             guard seen.insert(key).inserted else { return }
             let boost = context.learnedBoosts[key] ?? 0
-            let personalBoost = context.personalTerms.contains { $0.lowercased() == key } ? 2_000 : 0
+            let personalBoost = isPersonal(key, in: context) ? 2_000 : 0
-            ranked.append((raw, baseScore + boost + personalBoost, "\(tag)|\(key)"))
+            ranked.append((raw, baseScore + boost + personalBoost, role, quoted))
         }
 
         if let original = context.includeOriginalAfterCorrection {
-            append(original, baseScore: 20_000, tag: "original")
+            append(original, baseScore: 20_000, role: .verbatim, quoted: true)
         }
 
         if !context.previousWord.isEmpty {
             for (index, word) in lexicon.nextWords(after: context.previousWord, limit: 8).enumerated() {
-                append(word, baseScore: 1_000 - index * 10, tag: "next")
+                append(word, baseScore: 1_200 - index * 10, role: .nextWord)
             }
         }
 
         for term in context.personalTerms.prefix(4) {
-            append(term, baseScore: 500, tag: "personal")
+            append(term, baseScore: 500, role: .nextWord)
+        }
+
+        if ranked.filter({ $0.role == .nextWord }).isEmpty {
+            for (index, word) in lexicon.topWords(limit: 6).enumerated() {
+                append(word, baseScore: 200 - index, role: .nextWord)
+            }
         }
 
         ranked.sort { $0.score > $1.score }
-        let candidates = ranked.prefix(8).map {
+        let candidates = ranked.prefix(Self.slotCount).map {
-            TypingCandidate(id: $0.id, text: $0.text, engineIndex: 0)
+            TypingCandidate(
+                id: "\($0.role.rawValue)|\($0.text.lowercased())",
+                text: $0.text,
+                role: $0.role,
+                isQuoted: $0.quoted
+            )
         }
         return TypingComposition(preedit: "", candidates: Array(candidates))
     }
 
-    // MARK: - Helpers
+    public func isKnownWord(_ word: String, personalTerms: [String], systemWords: [String]) -> Bool {
+        let lower = word.lowercased()
+        if lexicon.contains(lower) { return true }
+        if personalTerms.contains(where: { $0.lowercased() == lower }) { return true }
+        if systemWords.contains(where: { $0.lowercased() == lower }) { return true }
+        return false
+    }
 
-    private func shouldSkipAutocorrect(_ typed: String) -> Bool {
+    // MARK: - Board
-        if typed.count <= 1 { return true }
+
+    private struct Board {
+        var composition: TypingComposition
+        var decision: EnglishCorrectionDecision?
+    }
+
+    private func makeBoard(_ context: EnglishSuggestionContext) -> Board {
+        let typed = context.currentWord
+        let decision = makeCorrectionDecision(context)
+        var slots: [TypingCandidate] = []
+        var seen = Set<String>()
+
+        func add(_ text: String, role: TypingCandidateRole, quoted: Bool = false) {
+            let key = text.lowercased()
+            guard seen.insert(key).inserted else { return }
+            slots.append(
+                TypingCandidate(
+                    id: "\(role.rawValue)|\(key)",
+                    text: text,
+                    role: role,
+                    isQuoted: quoted
+                )
+            )
+        }
+
+        let known = isKnownWord(
+            typed,
+            personalTerms: context.personalTerms,
+            systemWords: context.systemWords
+        )
+        add(typed, role: .verbatim, quoted: !known)
+
+        if let decision {
+            add(decision.replacement, role: .correction)
+        }
+
+        for term in context.personalTerms where term.lowercased().hasPrefix(typed.lowercased())
+            && term.lowercased() != typed.lowercased() {
+            add(term, role: .completion)
+            if slots.count >= Self.slotCount { break }
+        }
+
+        for word in context.systemCompletions {
+            let display = matchCase(of: typed, to: word)
+            add(display, role: .completion)
+            if slots.count >= Self.slotCount { break }
+        }
+
+        for word in lexicon.completions(prefix: typed, limit: 8) {
+            add(matchCase(of: typed, to: word), role: .completion)
+            if slots.count >= Self.slotCount { break }
+        }
+
+        let composition = TypingComposition(
+            preedit: typed,
+            candidates: Array(slots.prefix(Self.slotCount))
+        )
+        return Board(composition: composition, decision: decision)
+    }
+
+    private func makeCorrectionDecision(_ context: EnglishSuggestionContext) -> EnglishCorrectionDecision? {
+        let typed = context.currentWord
+        guard typed.count >= 3 else { return nil }
+        let lower = typed.lowercased()
+
+        if isProtectedToken(typed) { return nil }
+        if isPersonal(lower, in: context) { return nil }
+        if context.systemWords.contains(where: { $0.lowercased() == lower }) { return nil }
+        if (context.learnedBoosts[lower] ?? 0) >= 5 { return nil }
+
+        let inLexicon = lexicon.contains(lower)
+        let typedFreq = lexicon.frequency(of: lower) + (context.learnedBoosts[lower] ?? 0)
+
+        var pool = lexicon.scoredCorrections(for: lower, limit: 8)
+        for guess in context.systemGuesses {
+            let word = guess.lowercased()
+            guard word != lower else { continue }
+            if pool.contains(where: { $0.word == word }) { continue }
+            guard let alignment = EnglishQWERTYProximity.align(typed: lower, candidate: word) else { continue }
+            pool.append(
+                EnglishScoredCorrection(
+                    word: word,
+                    spatialCost: alignment.cost,
+                    frequency: max(lexicon.frequency(of: word), 1),
+                    isTransposition: alignment.isTransposition,
+                    isShortening: alignment.isShortening
+                )
+            )
+        }
+
+        var best: (EnglishScoredCorrection, Int)?
+        for candidate in pool {
+            guard allowsAutocorrect(
+                typed: typed,
+                replacement: candidate.word,
+                inLexicon: inLexicon,
+                typedFreq: typedFreq,
+                candidate: candidate
+            ) else { continue }
+            var score = candidate.frequency * 2 - candidate.spatialCost
+            if isPersonal(candidate.word, in: context) { score += 5_000 }
+            score += context.learnedBoosts[candidate.word] ?? 0
+            if lexicon.nextWords(after: context.previousWord).contains(candidate.word) {
+                score += 80
+            }
+            if let current = best {
+                if score > current.1 { best = (candidate, score) }
+            } else {
+                best = (candidate, score)
+            }
+        }
+
+        guard let best else { return nil }
+        let keepScore = inLexicon ? typedFreq * 2 : 0
+        guard best.1 > keepScore + 40 else { return nil }
+        return EnglishCorrectionDecision(
+            original: typed,
+            replacement: matchCase(of: typed, to: best.0.word)
+        )
+    }
+
+    private func allowsAutocorrect(
+        typed: String,
+        replacement: String,
+        inLexicon: Bool,
+        typedFreq: Int,
+        candidate: EnglishScoredCorrection
+    ) -> Bool {
+        if isTitleCase(typed) {
+            // Teh → The is a same-length transposition. Rocky → Rock is not.
+            guard candidate.isTransposition, !candidate.isShortening else { return false }
+        }
+        if inLexicon {
+            let gap = candidate.frequency - typedFreq
+            // Web-corpus dumps leak typos (`teh`, `adn`) at the floor of the
+            // list. Real words like `form` sit much higher and must not yield
+            // to `from`.
+            let looksLikeLeakedTypo = typedFreq <= 680
+            if candidate.isTransposition {
+                return looksLikeLeakedTypo && gap >= 40
+            }
+            if typed.count == replacement.count,
+               candidate.spatialCost <= EnglishQWERTYProximity.adjacentCost {
+                return looksLikeLeakedTypo && gap >= Self.inVocabularyFrequencyGap
+            }
+            return false
+        }
+        return candidate.frequency > 0
+    }
+
+    private func isProtectedToken(_ typed: String) -> Bool {
+        if typed.count <= 2 { return true }
         if typed.allSatisfy(\.isUppercase) { return true }
         if typed.contains(where: \.isNumber) { return true }
         if typed.contains("@") || typed.contains(".") || typed.contains("/") { return true }
@@ -170,6 +310,16 @@ public struct EnglishSuggestionEngine: Sendable {
         return false
     }
 
+    private func isTitleCase(_ typed: String) -> Bool {
+        guard let first = typed.first, first.isUppercase else { return false }
+        let rest = typed.dropFirst()
+        return !rest.isEmpty && rest.allSatisfy(\.isLowercase)
+    }
+
+    private func isPersonal(_ key: String, in context: EnglishSuggestionContext) -> Bool {
+        context.personalTerms.contains { $0.lowercased() == key }
+    }
+
     private func matchCase(of sample: String, to word: String) -> String {
         if sample.allSatisfy(\.isUppercase) {
             return word.uppercased()

OSGKeyboardShared/Typing/EnglishSystemLexicon.swift

@@ -0,0 +1,67 @@
+// EnglishSystemLexicon.swift
+// OSGKeyboard · Shared
+//
+// Apple's sanctioned English sources for a custom keyboard: UITextChecker
+// completions / guesses, plus UILexicon names from
+// `requestSupplementaryLexicon`. The engine stays pure; the keyboard
+// extension fills these fields on each refresh.
+
+import Foundation
+#if canImport(UIKit)
+import UIKit
+#endif
+
+@MainActor
+public protocol EnglishSystemLexiconProviding: AnyObject {
+    func completions(prefix: String, limit: Int) -> [String]
+    func guesses(for word: String, limit: Int) -> [String]
+}
+
+@MainActor
+public final class EmptyEnglishSystemLexicon: EnglishSystemLexiconProviding {
+    public init() {}
+
+    public func completions(prefix: String, limit: Int) -> [String] {
+        []
+    }
+
+    public func guesses(for word: String, limit: Int) -> [String] {
+        []
+    }
+}
+
+#if canImport(UIKit)
+/// System spellchecker. Always called from `TypingSessionController` (@MainActor).
+@MainActor
+public final class UIKitEnglishSystemLexicon: EnglishSystemLexiconProviding {
+    public var language: String
+
+    public init(language: String = "en_US") {
+        self.language = language
+    }
+
+    public func completions(prefix: String, limit: Int) -> [String] {
+        guard !prefix.isEmpty, limit > 0 else { return [] }
+        let checker = UITextChecker()
+        let range = NSRange(location: 0, length: (prefix as NSString).length)
+        let hits = checker.completions(forPartialWordRange: range, in: prefix, language: language) ?? []
+        return Array(hits.prefix(limit))
+    }
+
+    public func guesses(for word: String, limit: Int) -> [String] {
+        guard word.count >= 3, limit > 0 else { return [] }
+        let checker = UITextChecker()
+        let range = NSRange(location: 0, length: (word as NSString).length)
+        let hits = checker.guesses(forWordRange: range, in: word, language: language) ?? []
+        return Array(hits.prefix(limit))
+    }
+
+    public static func learnWord(_ word: String) {
+        let trimmed = word.trimmingCharacters(in: .whitespacesAndNewlines)
+        guard !trimmed.isEmpty else { return }
+        if !UITextChecker.hasLearnedWord(trimmed) {
+            UITextChecker.learnWord(trimmed)
+        }
+    }
+}
+#endif

OSGKeyboardShared/Typing/RimeEngineBridging.swift

@@ -21,6 +21,18 @@ public enum TypingInputLanguage: String, CaseIterable, Identifiable, Sendable {
     }
 }
 
+/// Role of an English QuickType slot. Chinese candidates stay `.completion`.
+public enum TypingCandidateRole: String, Equatable, Sendable {
+    /// The word currently being typed. Space does not replace it.
+    case verbatim
+    /// The unique slot Space will apply when autocorrect is armed.
+    case correction
+    /// Prefix completion; tap to accept, Space ignores it.
+    case completion
+    /// Next-word prediction after a committed word; tap to insert.
+    case nextWord
+}
+
 /// One candidate row item after composing.
 public struct TypingCandidate: Identifiable, Equatable, Sendable {
     public let id: String
@@ -28,17 +40,24 @@ public struct TypingCandidate: Identifiable, Equatable, Sendable {
     public let annotation: String?
     /// Absolute engine index for Chinese selection (may differ from display order).
     public let engineIndex: Int
+    public let role: TypingCandidateRole
+    /// Unknown verbatim shown in quotes, matching the system / KeyboardKit contract.
+    public let isQuoted: Bool
 
     public init(
         id: String = UUID().uuidString,
         text: String,
         annotation: String? = nil,
-        engineIndex: Int = 0
+        engineIndex: Int = 0,
+        role: TypingCandidateRole = .completion,
+        isQuoted: Bool = false
     ) {
         self.id = id
         self.text = text
         self.annotation = annotation
         self.engineIndex = engineIndex
+        self.role = role
+        self.isQuoted = isQuoted
     }
 }
 

OSGKeyboardShared/Typing/TypingSessionController.swift

@@ -27,6 +27,10 @@ public final class TypingSessionController: ObservableObject {
 
     /// When true, English suggestions / autocorrect stay off (secure fields).
     @Published public var suggestionsEnabled: Bool = true
+    /// `UITextChecker` completions / guesses. Empty in unit tests.
+    public var systemLexicon: EnglishSystemLexiconProviding = EmptyEnglishSystemLexicon()
+    /// Names and text replacements from `requestSupplementaryLexicon`.
+    public var supplementaryWords: [String] = []
 
     /// Chevron appears only for Chinese composition with at least two candidates.
     public var canExpandCandidatePanel: Bool {
@@ -128,9 +132,15 @@ public final class TypingSessionController: ObservableObject {
         )
         TypingInputConfiguration.shared.reload()
         refreshPersonalTerms()
-        // English lexicon is small; load when entering typing (not at KVC init).
+        // mmap the English table only while English is active. Chinese typing
-        englishEngine.prepare()
+        // already has Rime; loading both on appear is what jetsams the extension.
-        OSGDiag.log("typing.enter after englishPrepare \(OSGDiag.memoryTag())", category: "boot")
+        if language == .english {
+            englishEngine.prepare()
+            OSGDiag.log("typing.enter after englishPrepare \(OSGDiag.memoryTag())", category: "boot")
+        } else {
+            EnglishLexicon.shared.unload()
+            OSGDiag.log("typing.enter skip englishPrepare lang=\(language.rawValue) \(OSGDiag.memoryTag())", category: "boot")
+        }
         syncAutocapitalization()
         if FlowSessionBridge.isHostHeavy() {
             OSGDiag.log("typing.enter defer rime hostHeavy=1 — retry scheduled", category: "boot")
@@ -216,6 +226,7 @@ public final class TypingSessionController: ObservableObject {
             synchronizeEnglishDocumentContext(caretMoved: true)
         } else {
             clearEnglishWordState(keepPrevious: false)
+            EnglishLexicon.shared.unload()
             composition = engine.composition
         }
         return output
@@ -427,6 +438,9 @@ public final class TypingSessionController: ObservableObject {
                 pendingAutocorrection = nil
                 englishCurrentWord = pending.original
                 learningStore.recordDefense(of: pending.original)
+                #if canImport(UIKit)
+                UIKitEnglishSystemLexicon.learnWord(pending.original)
+                #endif
                 refreshEnglishSuggestions()
                 return .replace(deleteCount: deleteCount, with: pending.original)
             }
@@ -461,15 +475,18 @@ public final class TypingSessionController: ObservableObject {
            var decision = englishEngine.correctionDecision(
             for: word,
             personalTerms: personalTermsCache,
-            learnedBoosts: learningStore.snapshot()
+            learnedBoosts: learningStore.snapshot(),
+            previousWord: englishPreviousWord,
+            systemWords: supplementaryWords,
+            systemGuesses: systemLexicon.guesses(for: word, limit: 6)
         ) {
             decision.appliedSuffix = suffix
             pendingAutocorrection = decision
             englishPreviousWord = decision.replacement
             englishCurrentWord = ""
-            learningStore.recordAcceptance(of: decision.replacement)
+            // Machine-applied correction does not count as the user accepting
-            // Suggestions stay hidden until the user starts the next word.
+            // the replacement — otherwise names train the wrong word.
-            composition = .empty
+            refreshEnglishSuggestions(afterCommittedWord: decision.replacement)
             return .replace(
                 deleteCount: word.count,
                 with: decision.replacement + suffix
@@ -479,7 +496,17 @@ public final class TypingSessionController: ObservableObject {
         englishPreviousWord = word
         englishCurrentWord = ""
         pendingAutocorrection = nil
-        learningStore.recordAcceptance(of: word, amount: 1)
+        // Learn OOV / names the user actually committed; skip common words.
+        if !englishEngine.isKnownWord(
+            word,
+            personalTerms: personalTermsCache,
+            systemWords: supplementaryWords
+        ) {
+            learningStore.recordDefense(of: word, amount: 2)
+            #if canImport(UIKit)
+            UIKitEnglishSystemLexicon.learnWord(word)
+            #endif
+        }
         refreshEnglishSuggestions(afterCommittedWord: word)
         return suffix.isEmpty ? .none : .insert(suffix)
     }
@@ -487,7 +514,8 @@ public final class TypingSessionController: ObservableObject {
     private func selectEnglishCandidate(at index: Int) -> TypingOutput {
         guard composition.candidates.indices.contains(index) else { return .none }
         guard englishCandidateAnchorMatchesDocument() else { return .none }
-        let chosen = composition.candidates[index].text
+        let candidate = composition.candidates[index]
+        let chosen = candidate.text
 
         // Restoring original after autocorrect (no current word).
         if englishCurrentWord.isEmpty,
@@ -498,16 +526,27 @@ public final class TypingSessionController: ObservableObject {
             englishPreviousWord = pending.original
             englishCurrentWord = ""
             learningStore.recordDefense(of: pending.original)
+            #if canImport(UIKit)
+            UIKitEnglishSystemLexicon.learnWord(pending.original)
+            #endif
             refreshEnglishSuggestions(afterCommittedWord: pending.original)
             return .replace(deleteCount: deleteCount, with: pending.original + " ")
         }
 
+        if candidate.role == .verbatim {
+            learningStore.recordDefense(of: chosen)
+            #if canImport(UIKit)
+            UIKitEnglishSystemLexicon.learnWord(chosen)
+            #endif
+        } else {
+            learningStore.recordAcceptance(of: chosen)
+        }
+
         if !englishCurrentWord.isEmpty {
             let deleteCount = englishCurrentWord.count
             englishPreviousWord = chosen
             englishCurrentWord = ""
             pendingAutocorrection = nil
-            learningStore.recordAcceptance(of: chosen)
             refreshEnglishSuggestions(afterCommittedWord: chosen)
             return .replace(deleteCount: deleteCount, with: chosen + " ")
         }
@@ -516,7 +555,6 @@ public final class TypingSessionController: ObservableObject {
         englishPreviousWord = chosen
         englishCurrentWord = ""
         pendingAutocorrection = nil
-        learningStore.recordAcceptance(of: chosen)
         refreshEnglishSuggestions(afterCommittedWord: chosen)
         return .insert(chosen + " ")
     }
@@ -537,19 +575,23 @@ public final class TypingSessionController: ObservableObject {
             composition = .empty
             return
         }
-        // Idle / between words: no candidate bar. Completions start after
+        // With no active English word, keep the candidate bar empty. This also
-        // the first letter of the current word.
+        // prevents next-word predictions from appearing between committed words.
         guard !englishCurrentWord.isEmpty else {
             composition = .empty
             return
         }
         let previous = word ?? englishPreviousWord
+        let typed = englishCurrentWord
         let context = EnglishSuggestionContext(
-            currentWord: englishCurrentWord,
+            currentWord: typed,
             previousWord: previous,
             personalTerms: personalTermsCache,
             learnedBoosts: learningStore.snapshot(),
-            includeOriginalAfterCorrection: nil
+            includeOriginalAfterCorrection: pendingAutocorrection?.original,
+            systemWords: supplementaryWords,
+            systemCompletions: typed.isEmpty ? [] : systemLexicon.completions(prefix: typed, limit: 6),
+            systemGuesses: typed.count >= 3 ? systemLexicon.guesses(for: typed, limit: 6) : []
         )
         composition = englishEngine.compositionWhileTyping(context)
     }

OSGKeyboardTests/AIUserSkillTests.swift

@@ -62,6 +62,86 @@ final class AIUserSkillTests: XCTestCase {
         XCTAssertEqual(catalog.entries.first?.shortcutName, "My Tasks")
     }
 
+    func testTextOnlySkillAllowsEmptyShortcutConfiguration() throws {
+        var catalog = AIUserSkillCatalog()
+        let skill = AIUserSkill(
+            name: "Rewrite",
+            prompt: "Rewrite the clipboard"
+        )
+
+        try catalog.upsert(skill)
+
+        let saved = try XCTUnwrap(catalog.entries.first)
+        let clipboardSkill = saved.asClipboardSkill()
+        XCTAssertNil(saved.shortcutICloudURL)
+        XCTAssertEqual(saved.shortcutName, "")
+        XCTAssertEqual(clipboardSkill.kind, .transform)
+        XCTAssertFalse(clipboardSkill.requiresShortcut)
+        XCTAssertNil(clipboardSkill.shortcutName)
+    }
+
+    func testShortcutNameWithoutLinkDoesNotExport() throws {
+        var catalog = AIUserSkillCatalog()
+        let skill = AIUserSkill(
+            name: "Rewrite",
+            prompt: "Rewrite the clipboard",
+            shortcutName: "Ignored without a link"
+        )
+
+        try catalog.upsert(skill)
+
+        let saved = try XCTUnwrap(catalog.entries.first)
+        XCTAssertEqual(saved.shortcutName, "Ignored without a link")
+        XCTAssertEqual(saved.asClipboardSkill().kind, .transform)
+        XCTAssertNil(saved.asClipboardSkill().shortcutName)
+    }
+
+    func testShortcutSkillRequiresNameAndValidShareLink() {
+        var catalog = AIUserSkillCatalog()
+        XCTAssertThrowsError(
+            try catalog.upsert(
+                AIUserSkill(
+                    name: "Export",
+                    prompt: "Export it",
+                    shortcutICloudURL: sampleURL
+                )
+            )
+        ) { error in
+            XCTAssertEqual(error as? AIUserSkillValidationError, .emptyShortcutName)
+        }
+
+        XCTAssertThrowsError(
+            try catalog.upsert(
+                AIUserSkill(
+                    name: "Export",
+                    prompt: "Export it",
+                    shortcutICloudURL: URL(string: "https://example.com/not-a-shortcut"),
+                    shortcutName: "Run Me"
+                )
+            )
+        ) { error in
+            XCTAssertEqual(error as? AIUserSkillValidationError, .invalidShortcutLink)
+        }
+    }
+
+    func testExistingShortcutSkillEncodingDecodesWithOptionalURL() throws {
+        let original = AIUserSkill(
+            name: "Export",
+            prompt: "Export it",
+            shortcutICloudURL: sampleURL,
+            shortcutName: "Run Me"
+        )
+
+        let decoded = try JSONDecoder().decode(
+            AIUserSkill.self,
+            from: JSONEncoder().encode(original)
+        )
+
+        XCTAssertEqual(decoded.shortcutICloudURL, sampleURL)
+        XCTAssertEqual(decoded.asClipboardSkill().kind, .export)
+        XCTAssertTrue(decoded.asClipboardSkill().requiresShortcut)
+    }
+
     func testThinkingDefaultsOffAndBuiltinCannotEnable() {
         let user = AIUserSkill(
             name: "Custom",
@@ -186,4 +266,45 @@ final class AIUserSkillStoreTests: XCTestCase {
         XCTAssertFalse(store.layout.hasConfirmedShortcut(skill.id))
         XCTAssertFalse(store.layout.isEnabled(skill.id))
     }
+
+    func testRemovingShortcutLinkKeepsEnabledTextSkill() throws {
+        let store = AIAgentSkillLayoutStore(defaults: makeDefaults())
+        var skill = AIUserSkill(
+            name: "Custom",
+            prompt: "Do it",
+            shortcutICloudURL: URL(
+                string: "https://www.icloud.com/shortcuts/65bf33ba4206484ba78d582eaf1e9c44"
+            ),
+            shortcutName: "Run Me"
+        )
+        try store.saveUserSkill(skill)
+        XCTAssertEqual(store.confirmShortcutAndEnable(skill.id), .enabled)
+
+        skill.shortcutICloudURL = nil
+        try store.saveUserSkill(skill)
+
+        XCTAssertTrue(store.layout.isEnabled(skill.id))
+        XCTAssertFalse(store.layout.hasConfirmedShortcut(skill.id))
+        XCTAssertEqual(store.userSkill(id: skill.id)?.asClipboardSkill().kind, .transform)
+    }
+
+    func testAddingShortcutLinkDisablesTextSkillUntilConfirmed() throws {
+        let store = AIAgentSkillLayoutStore(defaults: makeDefaults())
+        var skill = AIUserSkill(
+            name: "Custom",
+            prompt: "Do it"
+        )
+        try store.saveUserSkill(skill)
+        XCTAssertEqual(store.enable(skill.id), .enabled)
+
+        skill.shortcutICloudURL = URL(
+            string: "https://www.icloud.com/shortcuts/65bf33ba4206484ba78d582eaf1e9c44"
+        )
+        skill.shortcutName = "Run Me"
+        try store.saveUserSkill(skill)
+
+        XCTAssertFalse(store.layout.isEnabled(skill.id))
+        XCTAssertFalse(store.layout.hasConfirmedShortcut(skill.id))
+        XCTAssertEqual(store.enable(skill.id), .needsShortcut)
+    }
 }

OSGKeyboardTests/EnglishTypingOnDeviceTests.swift

@@ -0,0 +1,77 @@
+// EnglishTypingOnDeviceTests.swift
+// OSGKeyboardTests
+//
+// Hosted in the main app so these can run on a physical iPhone.
+// ExtTests stay tool-hosted (simulator-only).
+
+import XCTest
+@testable import OSGKeyboardShared
+
+final class EnglishTypingOnDeviceTests: XCTestCase {
+    func testLexiconLoadsFortyThousandWords() {
+        let lexicon = EnglishLexicon()
+        lexicon.prepare()
+        XCTAssertTrue(lexicon.isLoaded)
+        XCTAssertGreaterThanOrEqual(lexicon.wordCount, 30_000)
+        XCTAssertTrue(lexicon.contains("the"))
+        XCTAssertTrue(lexicon.contains("hello"))
+        XCTAssertTrue(lexicon.contains("definitely"))
+        lexicon.unload()
+        XCTAssertFalse(lexicon.isLoaded)
+    }
+
+    func testTehAutocorrectsToThe() {
+        let engine = EnglishSuggestionEngine()
+        engine.prepare()
+        XCTAssertEqual(
+            engine.correctionDecision(for: "teh", personalTerms: [], learnedBoosts: [:])?.replacement,
+            "the"
+        )
+    }
+
+    func testTitleCaseNamesStay() {
+        let engine = EnglishSuggestionEngine()
+        engine.prepare()
+        XCTAssertNil(engine.correctionDecision(for: "Rocky", personalTerms: [], learnedBoosts: [:]))
+        XCTAssertNil(engine.correctionDecision(for: "Wang", personalTerms: [], learnedBoosts: [:]))
+    }
+
+    func testProximityGppdBecomesGood() {
+        let engine = EnglishSuggestionEngine()
+        engine.prepare()
+        XCTAssertEqual(
+            engine.correctionDecision(for: "gppd", personalTerms: [], learnedBoosts: [:])?.replacement,
+            "good"
+        )
+    }
+
+    func testFormIsNotCorrectedToFrom() {
+        let engine = EnglishSuggestionEngine()
+        engine.prepare()
+        XCTAssertNil(engine.correctionDecision(for: "form", personalTerms: [], learnedBoosts: [:]))
+    }
+
+    func testThankYouBigram() {
+        let lexicon = EnglishLexicon()
+        lexicon.prepare()
+        XCTAssertTrue(lexicon.nextWords(after: "thank").contains("you"))
+    }
+
+    @MainActor
+    func testUITextCheckerCompletionsAvailable() {
+        let system = UIKitEnglishSystemLexicon()
+        let hits = system.completions(prefix: "hel", limit: 6)
+        XCTAssertFalse(hits.isEmpty, "device UITextChecker should complete hel")
+    }
+
+    @MainActor
+    func testQuickTypeBoardMarksCorrection() {
+        let engine = EnglishSuggestionEngine()
+        engine.prepare()
+        let composition = engine.compositionWhileTyping(
+            EnglishSuggestionContext(currentWord: "teh")
+        )
+        XCTAssertEqual(composition.candidates.first?.role, .verbatim)
+        XCTAssertTrue(composition.candidates.contains { $0.role == .correction && $0.text == "the" })
+    }
+}

OSGKeyboardUITests/EnglishKeyboardDeviceUITests.swift

@@ -0,0 +1,43 @@
+// EnglishKeyboardDeviceUITests.swift
+// OSGKeyboardUITests
+//
+// Physical-device pass: Notes-like host + the real keyboard extension.
+// Skips if OSGKeyboard is not enabled as the current keyboard.
+
+import XCTest
+
+final class EnglishKeyboardDeviceUITests: XCTestCase {
+    func testOSGKeyboardAppearsOnNotesHost() throws {
+        let app = XCUIApplication()
+        app.launchArguments = [
+            "--whats-new-host",
+            "--whats-new-lang=en",
+            "--whats-new-scenario=edit"
+        ]
+        app.launch()
+
+        let textView = app.textViews["notes.host.textView"]
+        XCTAssertTrue(
+            textView.waitForExistence(timeout: 12),
+            "Notes host text view should appear"
+        )
+        if !textView.exists {
+            return
+        }
+        textView.tap()
+
+        let keyboard = XCUIApplication(bundleIdentifier: "com.osgkeyboard.ios.keyboard")
+        let appeared = keyboard.wait(for: .runningForeground, timeout: 8)
+            || keyboard.windows.firstMatch.waitForExistence(timeout: 8)
+        if !appeared {
+            throw XCTSkip(
+                "OSGKeyboard extension is not the active keyboard on this device. Enable it in Settings ▸ Keyboard, then re-run."
+            )
+        }
+
+        XCTAssertTrue(
+            keyboard.windows.firstMatch.exists,
+            "OSGKeyboard extension window should be on screen"
+        )
+    }
+}

Scripts/keyboard-appear-stress.sh

@@ -0,0 +1,131 @@
+#!/usr/bin/env bash
+# Physical-device keyboard appear/hide stress.
+# Shows and dismisses the real OSGKeyboard extension N times, then checks
+# crash reports.
+#
+# Usage:
+#   ./Scripts/keyboard-appear-stress.sh [COUNT=50] [UDID]
+set -euo pipefail
+
+ROOT="$(cd "$(dirname "$0")/.." && pwd)"
+cd "$ROOT"
+
+COUNT="${1:-50}"
+UDID="${2:-00008130-001C249C0E52001C}"
+BUNDLE="com.osgkeyboard.ios"
+OUT_DIR="${ROOT}/.tmp/keyboard-appear-stress-$(date +%Y%m%d-%H%M%S)"
+mkdir -p "$OUT_DIR"
+CONSOLE="$OUT_DIR/console.log"
+SUMMARY="$OUT_DIR/summary.txt"
+DERIVED="${ROOT}/.derivedData-device-stress"
+APP="$DERIVED/Build/Products/Debug-iphoneos/OSGKeyboard.app"
+
+echo "==> Out: $OUT_DIR"
+echo "==> Count: $COUNT  Device: $UDID"
+
+echo "==> Building Debug-iphoneos"
+xcodebuild build \
+  -project "$ROOT/OSGKeyboard.xcodeproj" \
+  -scheme OSGKeyboard \
+  -destination "platform=iOS,id=$UDID" \
+  -configuration Debug \
+  -derivedDataPath "$DERIVED" \
+  -allowProvisioningUpdates \
+  -onlyUsePackageVersionsFromResolvedFile \
+  CODE_SIGNING_ALLOWED=YES \
+  >/dev/null
+
+echo "==> Installing"
+xcrun devicectl device install app --device "$UDID" "$APP" --timeout 180 >/dev/null
+
+crash_list() {
+  xcrun devicectl device info files \
+    --device "$UDID" \
+    --domain-type systemCrashLogs \
+    --timeout 30 2>/dev/null \
+    | awk '/OSGKeyboardExt-/{print $1}'
+}
+
+BEFORE="$OUT_DIR/crashes-before.txt"
+AFTER="$OUT_DIR/crashes-after.txt"
+crash_list | sort >"$BEFORE"
+
+echo "==> Launching appear-stress count=$COUNT"
+# 50 hide/show cycles plus first show; ~0.5s each + timeouts.
+TIMEOUT_SECS=$((COUNT * 3 + 40))
+set +e
+python3 - "$TIMEOUT_SECS" "$UDID" "$BUNDLE" "$COUNT" "$CONSOLE" <<'PY'
+import subprocess, sys, time, os, signal
+timeout, udid, bundle, count, console = sys.argv[1:6]
+cmd = [
+    "xcrun", "devicectl", "device", "process", "launch",
+    "--device", udid,
+    "--console",
+    "--terminate-existing",
+    bundle,
+    "--whats-new-host",
+    "--whats-new-lang=en",
+    "--whats-new-scenario=edit",
+    f"--keyboard-appear-stress={count}",
+]
+with open(console, "w") as out:
+    proc = subprocess.Popen(cmd, stdout=out, stderr=subprocess.STDOUT)
+    try:
+        proc.wait(timeout=int(timeout))
+    except subprocess.TimeoutExpired:
+        proc.send_signal(signal.SIGTERM)
+        try:
+            proc.wait(timeout=8)
+        except subprocess.TimeoutExpired:
+            proc.kill()
+        sys.exit(124)
+    sys.exit(proc.returncode or 0)
+PY
+LAUNCH_STATUS=$?
+set -e
+
+crash_list | sort >"$AFTER"
+NEW_CRASHES="$OUT_DIR/crashes-new.txt"
+comm -13 "$BEFORE" "$AFTER" >"$NEW_CRASHES"
+
+PASSED="$(python3 - "$CONSOLE" <<'PY'
+import re, sys
+text = open(sys.argv[1], errors="replace").read()
+hits = re.findall(r"keyboard\.stress done passed=(\d+)/(\d+)", text)
+print(hits[-1][0] if hits else "")
+PY
+)"
+TOTAL="$(python3 - "$CONSOLE" <<'PY'
+import re, sys
+text = open(sys.argv[1], errors="replace").read()
+hits = re.findall(r"keyboard\.stress done passed=(\d+)/(\d+)", text)
+print(hits[-1][1] if hits else "")
+PY
+)"
+FAIL_LINE="$(grep 'keyboard.stress FAIL' "$CONSOLE" | tail -1 || true)"
+
+{
+  echo "Keyboard appear/hide stress"
+  echo "device=$UDID count=$COUNT"
+  echo "launch_exit=$LAUNCH_STATUS"
+  echo "passed=${PASSED:-0}/${TOTAL:-$COUNT}"
+  echo "fail_line=${FAIL_LINE:-none}"
+  echo "new_OSGKeyboardExt_crashes:"
+  if [[ -s "$NEW_CRASHES" ]]; then
+    cat "$NEW_CRASHES"
+  else
+    echo "  (none)"
+  fi
+  echo "console=$CONSOLE"
+} | tee "$SUMMARY"
+
+if [[ -s "$NEW_CRASHES" ]]; then
+  echo "FAIL: new OSGKeyboardExt crash reports" >&2
+  exit 1
+fi
+if [[ "${PASSED:-0}" != "$COUNT" ]]; then
+  echo "FAIL: expected $COUNT cycles, got ${PASSED:-0}" >&2
+  tail -40 "$CONSOLE" >&2
+  exit 1
+fi
+echo "PASS: $COUNT/$COUNT appear-hide cycles, no new extension crashes"

Scripts/typing/build_english_lexicon.py

@@ -0,0 +1,427 @@
+#!/usr/bin/env python3
+"""Build compact English unigram + bigram TSVs from Peter Norvig's public-domain
+n-gram counts (https://norvig.com/ngrams/).
+
+Norvig: “I hereby release all these files into the public domain.”
+We store log-scaled ranks (not raw counts) so the keyboard extension stays small
+and we are not redistributing the full Google Web Trillion Word Corpus dump.
+"""
+
+from __future__ import annotations
+
+import argparse
+import math
+import re
+import struct
+import sys
+import urllib.request
+from pathlib import Path
+
+ROOT = Path(__file__).resolve().parents[2]
+OUT_DIR = ROOT / "OSGKeyboardShared" / "Resources" / "Typing" / "English"
+UNIGRAM_URL = "https://norvig.com/ngrams/count_1w.txt"
+BIGRAM_URL = "https://norvig.com/ngrams/count_2w.txt"
+
+WORD_RE = re.compile(r"^[a-z]+(?:'[a-z]+)?$")
+MAX_UNIGRAMS = 40_000
+MAX_BIGRAMS = 8_000
+MAX_WORD_LEN = 20
+# Stop reading the 2-gram file once we have enough accepted rows; the file is
+# already sorted by descending count.
+BIGRAM_SCAN_LIMIT = 80_000
+
+# Seed collocations so next-word still works if the 2-gram download fails.
+FALLBACK_BIGRAMS: list[tuple[str, list[str]]] = [
+    ("the", "of and to in is for that with on a".split()),
+    ("of", "the a this my our their these those course them".split()),
+    ("to", "the be a do make see get go have my".split()),
+    ("and", "the a then also other more so we you".split()),
+    ("a", "lot few new good great little bit way time person".split()),
+    ("in", "the a my this order fact front time case addition".split()),
+    ("is", "a the not to that this it one more also".split()),
+    ("for", "the a example me you us this that now sure".split()),
+    ("that", "the is was I you it we they this are".split()),
+    ("i", "am have will would can do think know want was".split()),
+    ("it", "is was would will can be has had to not".split()),
+    ("on", "the a my this time top of it you that".split()),
+    ("you", "are can will would have do know want to should".split()),
+    ("with", "the a my you it this that him her them".split()),
+    ("as", "a the well much soon far long if of to".split()),
+    ("this", "is was the a time one way thing point case".split()),
+    ("we", "are have will can would do need want should were".split()),
+    ("have", "a the been to been a been the time been".split()),
+    ("be", "a the able to in on there here with as".split()),
+    ("are", "a the not you we they going to in on".split()),
+    ("not", "a the be to sure only yet even really the".split()),
+    ("but", "the I a it is also then we you not".split()),
+    ("from", "the a my this that it you now here there".split()),
+    ("at", "the a my this time least home work school night".split()),
+    ("by", "the a this that now then far me you email".split()),
+    ("or", "the a not so to it you we they this".split()),
+    ("an", "hour example email idea issue update account apple app".split()),
+    ("if", "you the I we it that this not so a".split()),
+    ("will", "be you I we the not have to a get".split()),
+    ("can", "be you I we not the a help see get".split()),
+    ("would", "be you I we like have not the a to".split()),
+    ("do", "you not the I we it that this a".split()),
+    ("there", "is are was were a the no not been have".split()),
+    ("their", "own new first last time way work house car".split()),
+    ("what", "is the a you I we do time about if".split()),
+    ("when", "the I you we it is a this that not".split()),
+    ("which", "is the a you we they of in to that".split()),
+    ("who", "is are was were the a you I we".split()),
+    ("how", "to much many long about is the a you".split()),
+    ("about", "the a this that it you to time me".split()),
+    ("into", "the a this that my it you a new".split()),
+    ("just", "a the like to be now want wanted got".split()),
+    ("like", "a the to this that it you I we".split()),
+    ("so", "I the a you we that this much many".split()),
+    ("than", "the a I you we this that it to".split()),
+    ("then", "the I you we a it to is was".split()),
+    ("them", "to a the in on with for and I".split()),
+    ("these", "are is the a days things people ones two".split()),
+    ("those", "are is the a who were days people ones".split()),
+    ("my", "own new first last time way email phone name".split()),
+    ("your", "own new email phone name time way account".split()),
+    ("our", "own new first last time team way house".split()),
+    ("going", "to be the a in on for with".split()),
+    ("want", "to a the you I we it".split()),
+    ("need", "to a the you I we it".split()),
+    ("let", "me you us the a".split()),
+    ("please", "let me you the a".split()),
+    ("thank", "you so much".split()),
+    ("thanks", "for so much".split()),
+    ("looking", "forward to for at".split()),
+    ("let", "me you us know".split()),
+]
+
+
+def fetch_lines(url: str, max_lines: int | None = None) -> list[str]:
+    req = urllib.request.Request(
+        url,
+        headers={"User-Agent": "OSGKeyboard-lexicon-builder/1.0"},
+    )
+    with urllib.request.urlopen(req, timeout=120) as response:
+        lines: list[str] = []
+        for raw in response:
+            line = raw.decode("utf-8", errors="ignore").strip()
+            if not line:
+                continue
+            lines.append(line)
+            if max_lines is not None and len(lines) >= max_lines:
+                break
+        return lines
+
+
+def parse_count_line(line: str) -> tuple[str, int] | None:
+    parts = line.split()
+    if len(parts) < 2:
+        return None
+    token = parts[0].lower()
+    try:
+        count = int(parts[-1])
+    except ValueError:
+        return None
+    return token, count
+
+
+def log_rank(count: int) -> int:
+    return max(1, int(round(math.log10(count) * 100)))
+
+
+def build_unigrams(lines: list[str]) -> dict[str, int]:
+    ranked: list[tuple[str, int]] = []
+    seen: set[str] = set()
+    for line in lines:
+        parsed = parse_count_line(line)
+        if parsed is None:
+            continue
+        word, count = parsed
+        if word in seen:
+            continue
+        if not WORD_RE.match(word) or len(word) > MAX_WORD_LEN:
+            continue
+        seen.add(word)
+        ranked.append((word, count))
+        if len(ranked) >= MAX_UNIGRAMS:
+            break
+    return {word: log_rank(count) for word, count in ranked}
+
+
+def build_bigrams(
+    lines: list[str],
+    unigrams: dict[str, int],
+) -> dict[str, list[str]]:
+    grouped: dict[str, list[tuple[str, int]]] = {}
+    accepted = 0
+    for line in lines:
+        parsed = parse_count_line(line)
+        if parsed is None:
+            continue
+        token, count = parsed
+        parts = token.split("_")
+        if len(parts) != 2:
+            # Norvig 2-grams are "word1 word2 count"
+            bits = line.lower().split()
+            if len(bits) < 3:
+                continue
+            left, right, count_s = bits[0], bits[1], bits[-1]
+            try:
+                count = int(count_s)
+            except ValueError:
+                continue
+        else:
+            left, right = parts
+        if left not in unigrams or right not in unigrams:
+            continue
+        if left == right:
+            continue
+        bucket = grouped.setdefault(left, [])
+        if any(word == right for word, _ in bucket):
+            continue
+        bucket.append((right, count))
+        accepted += 1
+        if accepted >= MAX_BIGRAMS * 3:
+            break
+
+    result: dict[str, list[str]] = {}
+    used = 0
+    for left, pairs in grouped.items():
+        pairs.sort(key=lambda item: item[1], reverse=True)
+        nxt = [word for word, _ in pairs[:8]]
+        if not nxt:
+            continue
+        result[left] = nxt
+        used += len(nxt)
+        if used >= MAX_BIGRAMS:
+            break
+    return result
+
+
+def merge_fallback(bigrams: dict[str, list[str]]) -> dict[str, list[str]]:
+    merged = dict(bigrams)
+    for left, rights in FALLBACK_BIGRAMS:
+        existing = merged.get(left, [])
+        seen = set(existing)
+        for word in rights:
+            if word not in seen:
+                existing.append(word)
+                seen.add(word)
+        merged[left] = existing[:10]
+    return merged
+
+
+def write_unigrams(path: Path, unigrams: dict[str, int]) -> None:
+    rows = sorted(unigrams.items(), key=lambda item: (-item[1], item[0]))
+    path.write_text("".join(f"{word}\t{freq}\n" for word, freq in rows), encoding="utf-8")
+
+
+def write_bigrams(path: Path, bigrams: dict[str, list[str]]) -> None:
+    rows = sorted(bigrams.items(), key=lambda item: item[0])
+    path.write_text(
+        "".join(f"{left}\t{' '.join(rights)}\n" for left, rights in rows),
+        encoding="utf-8",
+    )
+
+
+# mmap binary (`english_lexicon.bin`), little-endian. Layout:
+#   64-byte header, then unigram records, freq-rank indices, 26 initial
+#   ranges, bigram groups, packed next-word indices, ASCII string pool.
+# The keyboard maps this file; it must not parse TSV into Swift dictionaries.
+BIN_MAGIC = b"OSGENG01"
+BIN_VERSION = 1
+BIN_HEADER_SIZE = 64
+BIN_INITIAL_COUNT = 26
+
+
+def _align4(offset: int) -> int:
+    return (offset + 3) & ~3
+
+
+def write_binary(
+    path: Path,
+    unigrams: dict[str, int],
+    bigrams: dict[str, list[str]],
+) -> None:
+    words = sorted(unigrams.keys())
+    index_of = {word: index for index, word in enumerate(words)}
+
+    pool = bytearray()
+    records: list[tuple[int, int, int]] = []
+    for word in words:
+        encoded = word.encode("ascii")
+        if len(encoded) > 255:
+            continue
+        freq = min(int(unigrams[word]), 65_535)
+        records.append((len(pool), len(encoded), freq))
+        pool.extend(encoded)
+
+    initials = [(0, 0)] * BIN_INITIAL_COUNT
+    cursor = 0
+    while cursor < len(words):
+        first = words[cursor][0]
+        if "a" <= first <= "z":
+            start = cursor
+            while cursor < len(words) and words[cursor][0] == first:
+                cursor += 1
+            initials[ord(first) - ord("a")] = (start, cursor - start)
+        else:
+            cursor += 1
+
+    freq_order = sorted(
+        range(len(words)),
+        key=lambda index: (-unigrams[words[index]], words[index]),
+    )
+
+    groups: list[tuple[int, int, int]] = []
+    next_indices: list[int] = []
+    for left in sorted(bigrams.keys()):
+        prev_index = index_of.get(left)
+        if prev_index is None:
+            continue
+        rights = [index_of[word] for word in bigrams[left] if word in index_of]
+        if not rights:
+            continue
+        groups.append((prev_index, len(rights), len(next_indices)))
+        next_indices.extend(rights)
+
+    unigram_offset = _align4(BIN_HEADER_SIZE)
+    freq_offset = _align4(unigram_offset + len(records) * 8)
+    initial_offset = _align4(freq_offset + len(freq_order) * 2)
+    bigram_index_offset = _align4(initial_offset + BIN_INITIAL_COUNT * 4)
+    bigram_next_offset = _align4(bigram_index_offset + len(groups) * 8)
+    pool_offset = _align4(bigram_next_offset + len(next_indices) * 2)
+    total = pool_offset + len(pool)
+
+    blob = bytearray(total)
+    struct.pack_into(
+        "<8s14I",
+        blob,
+        0,
+        BIN_MAGIC,
+        BIN_VERSION,
+        len(records),
+        len(groups),
+        pool_offset,
+        len(pool),
+        unigram_offset,
+        freq_offset,
+        initial_offset,
+        bigram_index_offset,
+        bigram_next_offset,
+        0,
+        0,
+        0,
+        0,
+    )
+    for index, (pool_off, length, freq) in enumerate(records):
+        struct.pack_into(
+            "<IBBH",
+            blob,
+            unigram_offset + index * 8,
+            pool_off,
+            length,
+            0,
+            freq,
+        )
+    for index, word_index in enumerate(freq_order):
+        struct.pack_into("<H", blob, freq_offset + index * 2, word_index)
+    for letter, (start, count) in enumerate(initials):
+        struct.pack_into("<HH", blob, initial_offset + letter * 4, start, count)
+    for index, (prev_index, count, first_next) in enumerate(groups):
+        struct.pack_into(
+            "<HHI",
+            blob,
+            bigram_index_offset + index * 8,
+            prev_index,
+            count,
+            first_next,
+        )
+    for index, word_index in enumerate(next_indices):
+        struct.pack_into("<H", blob, bigram_next_offset + index * 2, word_index)
+    blob[pool_offset : pool_offset + len(pool)] = pool
+    path.write_bytes(blob)
+
+
+def read_unigrams_tsv(path: Path) -> dict[str, int]:
+    result: dict[str, int] = {}
+    for line in path.read_text(encoding="utf-8").splitlines():
+        if not line.strip():
+            continue
+        word, freq_s = line.split("\t", 1)
+        result[word.lower()] = int(freq_s)
+    return result
+
+
+def read_bigrams_tsv(path: Path) -> dict[str, list[str]]:
+    result: dict[str, list[str]] = {}
+    for line in path.read_text(encoding="utf-8").splitlines():
+        if not line.strip():
+            continue
+        left, rights = line.split("\t", 1)
+        result[left.lower()] = [word.lower() for word in rights.split() if word]
+    return result
+
+
+def emit_outputs(unigrams: dict[str, int], bigrams: dict[str, list[str]]) -> None:
+    OUT_DIR.mkdir(parents=True, exist_ok=True)
+    write_unigrams(OUT_DIR / "english_lexicon.tsv", unigrams)
+    write_bigrams(OUT_DIR / "english_bigrams.tsv", bigrams)
+    write_binary(OUT_DIR / "english_lexicon.bin", unigrams, bigrams)
+    print(f"Wrote {OUT_DIR / 'english_lexicon.tsv'}", file=sys.stderr)
+    print(f"Wrote {OUT_DIR / 'english_bigrams.tsv'}", file=sys.stderr)
+    print(
+        f"Wrote {OUT_DIR / 'english_lexicon.bin'} "
+        f"({(OUT_DIR / 'english_lexicon.bin').stat().st_size} bytes)",
+        file=sys.stderr,
+    )
+
+
+def main() -> int:
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument(
+        "--from-tsv",
+        action="store_true",
+        help="Compile english_lexicon.bin from existing TSV files (no network).",
+    )
+    args = parser.parse_args()
+    OUT_DIR.mkdir(parents=True, exist_ok=True)
+
+    if args.from_tsv:
+        unigram_path = OUT_DIR / "english_lexicon.tsv"
+        bigram_path = OUT_DIR / "english_bigrams.tsv"
+        if not unigram_path.is_file() or not bigram_path.is_file():
+            print("Missing english_lexicon.tsv / english_bigrams.tsv", file=sys.stderr)
+            return 1
+        unigrams = read_unigrams_tsv(unigram_path)
+        bigrams = read_bigrams_tsv(bigram_path)
+        write_binary(OUT_DIR / "english_lexicon.bin", unigrams, bigrams)
+        print(
+            f"Wrote {OUT_DIR / 'english_lexicon.bin'} "
+            f"({(OUT_DIR / 'english_lexicon.bin').stat().st_size} bytes) "
+            f"from {len(unigrams)} unigrams / {len(bigrams)} bigram keys",
+            file=sys.stderr,
+        )
+        return 0
+
+    print(f"Fetching unigrams from {UNIGRAM_URL}", file=sys.stderr)
+    unigram_lines = fetch_lines(UNIGRAM_URL, max_lines=200_000)
+    unigrams = build_unigrams(unigram_lines)
+    print(f"Kept {len(unigrams)} unigrams", file=sys.stderr)
+
+    bigrams: dict[str, list[str]] = {}
+    try:
+        print(f"Fetching bigrams from {BIGRAM_URL}", file=sys.stderr)
+        bigram_lines = fetch_lines(BIGRAM_URL, max_lines=BIGRAM_SCAN_LIMIT)
+        bigrams = build_bigrams(bigram_lines, unigrams)
+        print(f"Kept {sum(len(v) for v in bigrams.values())} bigram edges", file=sys.stderr)
+    except Exception as exc:  # noqa: BLE001 — fallback is intentional
+        print(f"Bigram download failed ({exc}); using fallback collocations", file=sys.stderr)
+
+    bigrams = merge_fallback(bigrams)
+    emit_outputs(unigrams, bigrams)
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

Tests/suite-manifest.json

@@ -120,6 +120,7 @@
       "platform": "ios",
       "tests": [
         "OSGKeyboardExtTests/EnglishTypingTests",
+        "OSGKeyboardTests/EnglishTypingOnDeviceTests",
         "OSGKeyboardExtTests/CandidatePanelExpandTests",
         "OSGKeyboardExtTests/ClipboardSuggestionLifecycleTests",
         "OSGKeyboardExtTests/KeyboardStateTests",

project.yml

@@ -139,6 +139,7 @@ targets:
       - path: OSGKeyboard/Resources/Typing/Licenses/LICENSE-JIEBA-MIT.txt
       - path: OSGKeyboard/Resources/Typing/Licenses/LICENSE-PHRASE-PINYIN-DATA-MIT.txt
       - path: OSGKeyboard/Resources/Typing/Licenses/LICENSE-PINYIN-DATA-MIT.txt
+      - path: OSGKeyboard/Resources/Typing/Licenses/LICENSE-NORVIG-PD.txt
       - path: OSGKeyboard/Resources/Typing/Licenses/THIRD_PARTY_NOTICES.md
       - path: OSGKeyboard/Resources/Typing/Licenses/third-party-notices.zip
       - path: OSGKeyboard/Resources/Typing/Licenses/LIBRIME-COMBINED-NOTICES.txt
@@ -348,9 +349,7 @@ targets:
         buildPhase: resources
       # Prefer sources+buildPhase over target-level `resources:` — the latter
       # was not emitting Copy Bundle Resources entries for this framework.
-      - path: OSGKeyboardShared/Resources/Typing/English/english_lexicon.tsv
+      - path: OSGKeyboardShared/Resources/Typing/English/english_lexicon.bin
-        buildPhase: resources
-      - path: OSGKeyboardShared/Resources/Typing/English/english_bigrams.tsv
         buildPhase: resources
       - path: OSGKeyboardShared/Resources/PolishStyles
         buildPhase: resources
@@ -582,6 +581,8 @@ targets:
         buildPhase: resources
       - path: OSGKeyboard/Resources/Typing/Licenses/LICENSE-PINYIN-DATA-MIT.txt
         buildPhase: resources
+      - path: OSGKeyboard/Resources/Typing/Licenses/LICENSE-NORVIG-PD.txt
+        buildPhase: resources
       - path: OSGKeyboard/Resources/Typing/Licenses/LICENSE.txt
         buildPhase: resources
       - path: OSGKeyboard/Resources/Typing/Licenses/NOTICE.txt