MapleEve · MapleEve · Jun 9, 2026 · Jun 9, 2026 · Jun 9, 2026
diff --git a/.github/workflows/claude-code-review.yml b/.github/workflows/claude-code-review.yml
@@ -7,6 +7,7 @@ on:
 
 permissions:
   contents: read
+  id-token: write
   pull-requests: write
   issues: write
 

diff --git a/app/Dockerfile b/app/Dockerfile
@@ -1,8 +1,8 @@
-# TODO: pin digest — 运行：
+# Reproducible-build note: operators may pin this mutable base tag by digest:
 #   docker pull pytorch/pytorch:2.6.0-cuda12.4-cudnn9-runtime
 #   docker inspect pytorch/pytorch:2.6.0-cuda12.4-cudnn9-runtime --format='{{index .RepoDigests 0}}'
-# 然后改为：FROM pytorch/pytorch:2.6.0-cuda12.4-cudnn9-runtime@sha256:<digest>
-# 不 pin digest 的话 mutable tag 可能导致构建不可重现（对应评审 CD-H6 / BP-M6）。
+# Then replace the FROM line with:
+#   FROM pytorch/pytorch:2.6.0-cuda12.4-cudnn9-runtime@sha256:<digest>
 FROM pytorch/pytorch:2.6.0-cuda12.4-cudnn9-runtime
 
 ENV DEBIAN_FRONTEND=noninteractive

diff --git a/app/config.py b/app/config.py
@@ -9,7 +9,7 @@
 from pathlib import Path
 
 
-APP_VERSION = "0.7.6"
+APP_VERSION = "0.8.4"
 
 
 def _env_float(name: str, default: float) -> float:

diff --git a/doc/changelog.en.md b/doc/changelog.en.md
@@ -4,6 +4,10 @@
 
 ## Unreleased
 
+No unreleased changes.
+
+## 0.8.4 — Rust kernel foundation and release gates (2026-06-10)
+
 ### Features
 
 - Added an optional `artifacts` manifest to completed results. The manifest

diff --git a/doc/changelog.zh.md b/doc/changelog.zh.md
@@ -4,6 +4,10 @@
 
 ## Unreleased
 
+暂无未发布变更。
+
+## 0.8.4 — Rust kernel 基础能力与发布门禁 (2026-06-10)
+
 ### 功能
 
 - 新增完成态结果的可选 `artifacts` manifest。该字段只列出当前结果相关 artifact

diff --git a/doc/configuration.en.md b/doc/configuration.en.md
@@ -2,7 +2,7 @@
 
 [简体中文](./configuration.zh.md) | **English**
 
-This is the public configuration index for VoScript v0.7.6. It covers the
+This is the public configuration index for VoScript v0.8.4. It covers the
 environment variables that the current code reads, the per-request override
 semantics of `POST /api/transcribe`, and internal defaults that are documented
 for operators but are not stable public knobs yet. Do not assume a Whisper,
@@ -40,7 +40,7 @@ parameters yet.
 | `MODEL_IDLE_TIMEOUT_SEC` | `180` | GPU model idle-unload timeout, defaulting to 180 seconds (3 minutes). Set `0` to disable idle unload and keep models resident. When enabled, loaded models are released only after the serialized GPU runtime has been idle for this many seconds; on the next reload, ASR, diarization, and embedding each choose the visible CUDA device with the most free memory during their own lazy load. |
 | `RUST_KERNEL_MODE` | `off` | Optional Rust-backed provider/kernel mode. `off` keeps Python implementations; `required` makes selected Rust-backed paths import and run successfully or fail closed. The current selected paths are voiceprint scoring, result post-processing, and artifact manifest helper contracts; CI / Docker packaging still validates the Rust extension directly when the runtime default is off. |
 
-`MODELS_DIR` and `LANGUAGE` are defined in the config module, but v0.7.6's main
+`MODELS_DIR` and `LANGUAGE` are defined in the config module, but v0.8.4's main
 HTTP transcription path does not use them as stable public tuning knobs:
 Whisper local checkpoint lookup still expects `/models/faster-whisper-<WHISPER_MODEL>`,
 and default language should be controlled with the request `language` field or
@@ -97,7 +97,7 @@ cache is incomplete.
 
 Current internal ASR defaults are `beam_size=5`, `vad_filter=True`,
 `vad_parameters.min_silence_duration_ms=500`, and `condition_on_previous_text=False`.
-These do not have env or API fields in v0.7.6. Do not configure nonexistent
+These do not have env or API fields in v0.8.4. Do not configure nonexistent
 variables such as `WHISPER_BEAM_SIZE`, `WHISPER_COMPUTE_TYPE`, or `WHISPER_VAD_*`.
 
 ## Denoising
@@ -109,7 +109,7 @@ variables such as `WHISPER_BEAM_SIZE`, `WHISPER_COMPUTE_TYPE`, or `WHISPER_VAD_*
 | API `denoise_model` | omitted | Omitted means inherit `DENOISE_MODEL`; explicit `none` disables denoising for this job only. |
 | API `snr_threshold` | omitted | Omitted means inherit `DENOISE_SNR_THRESHOLD`; explicit values override the DeepFilterNet SNR gate for this job only. |
 
-v0.7.6 defaults to `DENOISE_MODEL=none` for clean meeting-recorder audio. Enable
+v0.8.4 defaults to `DENOISE_MODEL=none` for clean meeting-recorder audio. Enable
 `deepfilternet` or `noisereduce` only for noisy environments, either per job or
 as a service default. If you need clean recordings to be skipped automatically,
 use `deepfilternet`; `noisereduce` runs whenever it is selected.
@@ -171,7 +171,7 @@ Cohort lifecycle:
   files to build and save a cohort.
 - After each enroll / update, the background `cohort-rebuild` thread wakes every
   60 seconds and rebuilds after the latest enrollment is at least 30 seconds old.
-- v0.7.6 protects larger loaded or persisted cohorts during automatic rebuilds:
+- v0.8.4 protects larger loaded or persisted cohorts during automatic rebuilds:
   clearing transcription results, having only a few embeddings, or having fewer
   source embeddings than the current cohort will not shrink the cohort automatically.
 - `POST /api/voiceprints/rebuild-cohort` is an explicit manual rebuild and uses
@@ -206,6 +206,14 @@ New fields are added under the optional-field principle. Clients should ignore
 unknown fields and tolerate missing `words`, `alignment`, `artifacts`, and
 `warning`.
 
+## v0.8.4 Validation Wording
+
+v0.8.4 has internal live validation covering the optional Rust kernel foundation,
+selected voiceprint scoring, result post-processing, artifact/status helper
+contracts, Docker packaging smoke, and public release-scan gates. Public
+documentation records only these behavior categories, not real task names,
+sample names, job IDs, speaker IDs, hosts, logs, or paths.
+
 ## v0.7.6 Validation Wording
 
 v0.7.6 has internal live validation covering `/healthz` availability during GPU

diff --git a/doc/configuration.zh.md b/doc/configuration.zh.md
@@ -2,7 +2,7 @@
 
 **简体中文** | [English](./configuration.en.md)
 
-本文是 VoScript v0.7.6 的公开配置索引，覆盖当前代码已经读取并生效的
+本文是 VoScript v0.8.4 的公开配置索引，覆盖当前代码已经读取并生效的
 环境变量、`POST /api/transcribe` 的请求级覆盖语义，以及还没有暴露为稳定
 配置项的内部默认值。没有在本文列出的 Whisper / diarization / AS-norm 变量，
 不要假定已经可用。
@@ -38,7 +38,7 @@
 | `MODEL_IDLE_TIMEOUT_SEC` | `180` | GPU 模型空闲卸载超时，默认 180 秒（3 分钟）。设为 `0` 可关闭空闲卸载并保持模型常驻。开启后，只有串行 GPU 运行时空闲达到该秒数才释放已加载模型；下一次 reload 时 ASR、diarization 和 embedding 会在各自 lazy load 时分别选择当前可见 CUDA 中空闲显存最多的设备。 |
 | `RUST_KERNEL_MODE` | `off` | 可选 Rust-backed provider/kernel 路径开关。`off` 保持 Python 实现；`required` 要求被选择的 Rust-backed 路径可导入并执行，缺失或调用失败时 fail closed。当前被选择的路径是声纹计分、结果后处理和 artifact manifest helper contract；默认关闭时，CI / Docker packaging 仍会直接验证 Rust 扩展。 |
 
-`MODELS_DIR` 和 `LANGUAGE` 在配置模块里有定义，但 v0.7.6 的主 HTTP 转写路径
+`MODELS_DIR` 和 `LANGUAGE` 在配置模块里有定义，但 v0.8.4 的主 HTTP 转写路径
 没有把它们作为稳定公开调参入口使用：Whisper 本地 checkpoint 查找仍使用
 `/models/faster-whisper-<WHISPER_MODEL>`，语言默认请通过请求字段 `language`
 控制或留空自动检测。
@@ -90,7 +90,7 @@ Hugging Face snapshot，缓存不完整时再走 Hub。
 
 当前内部 ASR 默认值：`beam_size=5`、`vad_filter=True`、
 `vad_parameters.min_silence_duration_ms=500`、`condition_on_previous_text=False`。
-这些值在 v0.7.6 还没有对应 env 或 API 字段；不要写 `WHISPER_BEAM_SIZE`、
+这些值在 v0.8.4 还没有对应 env 或 API 字段；不要写 `WHISPER_BEAM_SIZE`、
 `WHISPER_COMPUTE_TYPE`、`WHISPER_VAD_*` 之类未实现配置。
 
 ## 降噪
@@ -102,7 +102,7 @@ Hugging Face snapshot，缓存不完整时再走 Hub。
 | API `denoise_model` | 省略 | 省略表示继承 `DENOISE_MODEL`；显式传 `none` 表示只对本次任务关闭降噪。 |
 | API `snr_threshold` | 省略 | 省略表示继承 `DENOISE_SNR_THRESHOLD`；显式传值只覆盖本次任务的 DeepFilterNet SNR gate。 |
 
-v0.7.6 默认面向干净会议录音，因此 `DENOISE_MODEL=none`。只有噪声环境才建议按任务
+v0.8.4 默认面向干净会议录音，因此 `DENOISE_MODEL=none`。只有噪声环境才建议按任务
 或服务级启用 `deepfilternet` / `noisereduce`。如需“干净录音自动跳过”，请选择
 `deepfilternet`；`noisereduce` 一旦被选择就会运行。
 
@@ -160,7 +160,7 @@ cohort 生命周期：
 - 否则扫描持久化转写结果和 `emb_*.npy` 构建并保存 cohort。
 - 每次 enroll / update 后，后台 `cohort-rebuild` 线程每 60 秒检查一次，在最近一次
   enroll 至少过去 30 秒后自动重建。
-- v0.7.6 的后台自动重建会保护更大的已加载或已持久化 cohort：清空转写结果、
+- v0.8.4 的后台自动重建会保护更大的已加载或已持久化 cohort：清空转写结果、
   只有少量 embedding，或源数量少于现有 cohort 时，不会自动缩小 cohort。
 - `POST /api/voiceprints/rebuild-cohort` 是显式手动重建，仍按当前可用 embedding
   立即生成新 cohort。
@@ -188,6 +188,13 @@ cohort 生命周期：
 新增字段按可选字段原则扩展；客户端应忽略不认识的字段，并容忍 `words` /
 `alignment` / `artifacts` / `warning` 缺失。
 
+## v0.8.4 验证口径
+
+v0.8.4 已用 internal live validation 覆盖：可选 Rust kernel 基础能力、已选择的
+声纹计分、结果后处理、artifact/status helper contract、Docker packaging smoke
+以及 public release scan gate。公开文档只记录行为类别，不发布真实任务名、
+样本名、job id、speaker id、主机、日志或路径。
+
 ## v0.7.6 验证口径
 
 v0.7.6 已用 internal live validation 覆盖：GPU cleanup 期间 `/healthz` 仍保持可用、

diff --git a/doc/security.en.md b/doc/security.en.md
@@ -22,7 +22,7 @@ Treat the service as if it were an internal database.
 
 ## Built-in hardening (on by default)
 
-As of 0.7.6 the following protections are in place out of the box:
+As of 0.8.4 the following protections are in place out of the box:
 
 1. **Container runs as a non-root user.** The Dockerfile creates an
    `app` user (uid/gid 1000 by default, overridable via `APP_UID`/

diff --git a/doc/security.zh.md b/doc/security.zh.md
@@ -19,7 +19,7 @@
 
 ## 内置的硬化（默认启用）
 
-当前版本（0.7.6）默认开启以下保护：
+当前版本（0.8.4）默认开启以下保护：
 
 1. **容器以非 root 用户运行**。Dockerfile 创建 `app` 用户（uid/gid 1000，
    可通过 `APP_UID`/`APP_GID` 覆盖），`USER app`。即使服务代码被 RCE，

diff --git a/tests/unit/test_main_lifespan.py b/tests/unit/test_main_lifespan.py
@@ -54,8 +54,10 @@ def test_rebuild_thread_alive_during_lifespan(app_client):
     assert thread.daemon, "cohort-rebuild thread must be a daemon thread"
 
 
-def test_openapi_version_reports_076(app_client):
-    assert app_client.app.version == "0.7.6"
+def test_openapi_version_matches_config(app_client):
+    from app.config import APP_VERSION
+
+    assert app_client.app.version == APP_VERSION
 
 
 def test_rebuild_thread_survives_tick_exception(app_client, monkeypatch):