feat: iCloud settings sync and cold-start return redesign

- Add iCloud key-value settings sync (engine/language/polish/Flow prefs);
  API keys stay on-device. New "Sync settings via iCloud" toggle.
- Redesign cold-start handoff: bottom-bar left-to-right swipe guidance,
  auto-dismiss on app switch, tap-anywhere to close, retained return link.
- Harden keyboard->app handoff with host-disconnected hint.
- Include prior Unreleased ASR fixes (route-change crash, fallback warning,
  multi-utterance recognition, local ASR diagnostics).

Release 0.5.0 (build 18).

CHANGELOG.md

@@ -7,6 +7,21 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+## [0.5.0] - 2026-07-07
+
+### Added
+- **iCloud settings sync**: engine, language, polish, and Flow preferences now stay in sync across your devices via iCloud (key-value store); a new "Sync settings via iCloud" toggle in Settings controls it. API keys stay on each device and are never uploaded. / **iCloud 设置同步**：引擎、语言、润色与 Flow 偏好现可通过 iCloud（键值存储）在多设备间自动同步；设置中新增「通过 iCloud 同步设置」开关控制此功能。API 密钥仅保留在各自设备本地，绝不上传。
+
+### Changed
+- **Cold-start return guidance**: the handoff screen now points to the bottom Home indicator with a left-to-right swipe animation (instead of the misleading "swipe up"), auto-dismisses once you switch back to your previous app, and closes on a tap anywhere; a "Return to [App]" link is still offered when available. / **冷启动返回引导**：交接页改为指向底部横条并配左向右滑动动画（不再是易误解的「向上滑」），切换回上一个 App 后自动消失，点按任意处即可关闭；可用时仍保留「返回 [App]」文本按钮。
+- **Voice session handoff robustness**: hardened the keyboard→app cold-start flow, including a clearer "Voice session disconnected" hint when the host session drops. / **语音会话交接健壮性**：强化键盘→主 App 的冷启动链路，宿主会话断开时给出更清晰的「语音会话已断开」提示。
+
+### Fixed
+- **Local ASR route-change crash**: dictating with the on-device engine no longer terminates the app with `Failed to create tap due to format mismatch`. On-device `SpeechAnalyzer` warmup reconfigures the shared audio session, which triggers a route change; the tap was reinstalled with a stale hardware format (48 kHz) that no longer matched the live node (24 kHz). The tap now binds to the node's live format (`installTap(format: nil)`) and the downsampling converter rebuilds itself from the actual buffer format, so route churn is handled without crashing. / **本地识别路由切换崩溃**：使用端侧引擎听写不再以 `Failed to create tap due to format mismatch` 崩溃退出。端侧 `SpeechAnalyzer` 预热会重配共享音频会话并触发路由变化，此前重装 tap 时用了过期的硬件采样率（48 kHz），与真实节点（24 kHz）不匹配。现在 tap 绑定节点实时格式（`installTap(format: nil)`），降采样转换器按实际缓冲区格式自适应重建，路由抖动不再导致崩溃。
+- **ASR fallback warning showed raw key**: the weak-network / missing-key fallback hint displayed its localization key (e.g. `flow.warning.polishDegraded`) instead of the translated sentence. These keys live in the `Shared.strings` table but were looked up via the main-app `Localizable` table; they are now resolved through `SharedL10n`. / **ASR 兜底提示显示变量名**：弱网 / 缺少密钥的兜底提示此前显示本地化键名（如 `flow.warning.polishDegraded`）而非译文。这些键位于 `Shared.strings`，却被按主 App 的 `Localizable` 表查找；现改为经 `SharedL10n` 解析。
+- **Flow multi-utterance recognition**: the second (and later) dictation in a session no longer returns "no speech". The session-long downsampling `AVAudioConverter` was being permanently locked by an `.endOfStream` tail-flush, starving every utterance after the first (both on-device and cloud). Trailing speech is still preserved by the live drain loop. Also added recovery from `mediaServicesWereReset`. / **Flow 连续多句识别**：同一会话内第二句及之后不再提示「未识别到语音」。此前会话级降采样 `AVAudioConverter` 被尾音冲刷的 `.endOfStream` 永久锁死，导致首句之后每句都拿不到音频（端侧与云端均受影响）。尾音仍由实时排空环节保留。并新增媒体服务重置（`mediaServicesWereReset`）后的自愈重建。
+- **Local ASR diagnostics**: add chunk-level `SpeechAnalyzer` logs and a settings switch to bypass the custom language model, so local recognition failures can be isolated between Apple assets, CLM attachment, and empty analyzer results. / **本地识别诊断**：新增 `SpeechAnalyzer` 分块级日志，并在设置中加入跳过自定义语言模型的诊断开关，用于区分 Apple 端侧资源、自定义语言模型挂载、以及分析器空结果三类问题。
+
 ## [0.4.1] - 2026-07-06
 
 ### Added