PsychQuant · kiki830621 · Jul 10, 2026 · Jul 10, 2026
diff --git a/CHANGELOG.md b/CHANGELOG.md
@@ -1,7 +1,59 @@
 # Changelog
 
+All notable changes to bestASR are documented here. The format follows
+[Keep a Changelog](https://keepachangelog.com/en/1.1.0/); versions follow SemVer.
+
 ## [Unreleased]
 
+## [0.12.0] - 2026-07-10
+
+### Added
+
+- **MCP async job mode (#86)**: `transcribe` accepts an opt-in `async` flag and
+  returns a `job_id` immediately; new read-only tools `transcribe_status` and
+  `transcribe_result` (bounded long-poll, 25 s cap) poll it. Jobs live in a
+  bounded in-memory registry (TTL eviction + a global sweep on every start) and
+  share the same single-flight serialization as synchronous transcribes.
+- **macOS GUI dual-track bundle (#87)**: new SwiftUI `bestasr-gui` app (drag &
+  drop / file picker, language/effort/format pickers persisted across launches,
+  honest stage+elapsed progress, transcript preview + reveal-in-Finder) and
+  `scripts/release-app.sh`, which assembles, signs, notarizes, and **staples**
+  a `bestASR.app` carrying the GUI, `bestasr-mcp`, and the CLI as
+  `bestasr-cli` (default APFS is case-insensitive — a `bestasr` entry would
+  overwrite the `bestASR` GUI executable). First offline-Gatekeeper-verifiable
+  bestASR artifact, published as `bestASR-0.12.0.zip` on the v0.12.0 release.
+- **LibriSpeech English benchmark corpora (#88)**: test-clean + dev-clean,
+  8 corpora / 48 utterances, source tarballs and converted artifacts digest-
+  pinned end to end; `references/asr-benchmark-landscape.md` records the cited
+  dataset/license/methodology survey behind the pick.
+
+### Fixed
+
+- **External-adapter watchdog hang (#91)**: a spurious `Process.isRunning`
+  false right after launch could skip the timeout watchdog entirely, leaving an
+  unbounded `waitUntilExit()` (a 1-hour CI hang). The loop is now gated on a
+  `terminationHandler`-driven exit latch installed before `run()`, so it can
+  exit only via real process exit or the SIGTERM→SIGKILL deadline branch.
+- **bash 3.2 empty-array crash in release/install scripts**: expanding an empty
+  `"${BUILD_ENV[@]}"` under `set -u` aborts on stock macOS bash — which killed
+  the build exactly on the recommended Xcode-toolchain path. Guarded in
+  `release-app.sh`, `release-mcp.sh`, and `install.sh`; the app version parse
+  is now scoped to the `BestASRVersion` enum and semver-asserted.
+
+## [0.11.0] - 2026-07-08
+
+### Added
+
+- **MCP server surface (#80, #84)**: `bestasr-mcp` speaks MCP over stdio
+  (official swift-sdk), linking BestASRKit directly so engine pipeline caches
+  persist across tool calls; v1 tools: transcribe / recommend / list_backends /
+  list_models / corpus_add. Tool errors are loud and typed; transcribes are
+  single-flight serialized.
+- **Plugin bundles the MCP server (#85)**: the Claude Code plugin auto-downloads
+  a Developer ID-signed, notarized `bestasr-mcp` from GitHub Releases
+  (che-mcps wrapper pattern); `scripts/release-mcp.sh` builds, signs,
+  smoke-tests under hardened runtime, notarizes, and publishes it.
+
 ### Added
 - mlx-audio catalog measured (#65): seven families live-probed, revision-pinned, and benchmarked — canary 1b / granite-speech 2b / voxtral-realtime 4b hit en WER 3.8% (front-tier), vibevoice-asr 9b reaches zh CER 17.7%; nemotron-asr and moonshine verified; qwen2-audio measured (chat-style output inflates WER honestly). qwen3-asr and mega-asr fail in the mlx_audio loader ("All arrays must have the same shape"); distil-whisper lacks its processor config; mms / voxtral mini-3b / qwen3-forcedaligner have no mlx conversion. mlx candidates are now addressed family/size end-to-end (bare-size collision trapped the benchmark report).
 
@@ -17,9 +69,6 @@
 ### Changed
 - **Breaking (output format)**: diarized speaker prefixes are now human-readable — SRT/VTT cues read `Speaker 1: text` (was `[SPEAKER_1] text`) and txt lines use the same `Speaker N: ` form; enrolled names render as `Name: `. JSON keeps the internal `SPEAKER_N` label. Downstream parsers of the old bracket form must update. (#54)
 
-All notable changes to bestASR are documented here. The format follows
-[Keep a Changelog](https://keepachangelog.com/en/1.1.0/); versions follow SemVer.
-
 ## [0.10.0] - 2026-07-06
 
 ### Added

diff --git a/changelog/20260710_v0.12.0-async-gui-watchdog.md b/changelog/20260710_v0.12.0-async-gui-watchdog.md
@@ -0,0 +1,47 @@
+# 2026-07-10 — v0.12.0：async MCP 工具 + 首顆 stapled bestASR.app + watchdog 修正
+
+## 出貨（GitHub Release v0.12.0，四個 assets）
+
+- **MCP async job mode（#86，PR #89）**：`transcribe` 增加 opt-in `async` 參數
+  → 立即回 `job_id`；新增唯讀工具 `transcribe_status` / `transcribe_result`
+  （bounded long-poll，25 s cap）。`JobRegistry` actor：TTL 逐出 + 每次 start
+  全域 sweep（registry 有界）、與同步路徑共用 SingleFlight 序列化。
+- **macOS GUI dual-track bundle（#87，PR #93）**：SwiftUI `bestasr-gui`
+  （`BestASRGUICore` 可測 session 狀態機 + thin views；拖放/檔案選擇、
+  語言/effort/格式選項持久化、stage+elapsed 誠實進度、結果預覽 + Finder 顯示）。
+  `scripts/release-app.sh`：universal build → bundle 組裝 → Developer ID
+  nested-first 簽章 → notarize → **staple + validate** → zip。CLI 在 bundle 內
+  改名 `bestasr-cli`（預設 APFS 不分大小寫，`bestasr` 會覆寫 GUI `bestASR`——
+  smoke test 首跑即抓到，issue 原佈局不可行）。
+- **LibriSpeech 英文標準語料（#88，PR #90）**：test-clean + dev-clean 各 4 組
+  ×6 utterances（8 corpora / 48 utterances），來源 tarball 與轉檔產物全程
+  SHA-256 pin；`references/asr-benchmark-landscape.md` 收錄 cited survey
+  （license/access 矩陣、WER/CER 方法論）。
+- **ExternalProcessEngine watchdog 修正（#91，PR #95）**：`Process.isRunning`
+  啟動期 race 會讓 watchdog 零迭代跳過 → 無界 `waitUntilExit()`（CI 上掛滿
+  stub 的 3600 s）。改用 `terminationHandler` 驅動的 `ExitLatch`（run() 前掛好）
+  作為迴圈條件——迴圈只能經「真的退出」或 deadline kill 離開；hung-adapter
+  測試 stub 3600 s → 15 s（regress 成本從 1 小時壓到秒級）。
+
+## 程式檔變更（今日）
+
+- `Sources/BestASRKit/Engines/ExternalProcessEngine.swift` — ExitLatch watchdog
+- `Tests/BestASRKitTests/ExternalEngineTests.swift` — stub 降 15 s
+- `Sources/BestASRKit/Models/DataModels.swift` — `BestASRVersion.current` → 0.12.0
+- `scripts/release-app.sh` / `scripts/release-mcp.sh` / `scripts/install.sh` —
+  bash 3.2 空陣列 `set -u` guard（`${arr[@]+...}`；verify #87 HIGH，三處同款潛伏 bug 一次掃除）
+  與 enum-scoped + semver-asserted 版本 parse
+
+## 發布動作
+
+- `release-mcp.sh 0.12.0`：notarize Accepted → Release v0.12.0（`bestasr-mcp` + sha256）
+- `release-app.sh`：notarize Accepted → **staple 驗證通過**（首顆離線可驗證 artifact）
+  → `bestASR-0.12.0.zip` + sha256 上傳同一 release
+- `/plugin-update bestasr`：marketplace + plugin 0.11.0 → 0.12.0
+
+## 維運
+
+- Spectra changes `mcp-surface` / `mcp-transcribe-async` / `add-librispeech-corpus` /
+  `gui-app-dual-track` 歸檔（PR #92、#94），living specs 同步（新 `mcp-surface`、
+  `gui-app` capability spec；corpora 增 LibriSpeech requirement）。
+- Issues #86 / #87 / #88 / #91 均以完整 closing summary 手動關閉；backlog 歸零。