release: 10.12.0 - the first date-based version (DeepFilterNet length fix)

.github/workflows/release.yml

@@ -1,8 +1,15 @@
 name: Release
 
-# Push a version tag (v0.2.0, v0.3.1, ...) and this job builds the .deb on
-# Ubuntu 24.04 and publishes it as a GitHub Release asset with sha256sums -
-# that release is what tools/install-deb.sh downloads.
+# Push a version tag and this job builds the .deb on Ubuntu 24.04 and publishes
+# it as a GitHub Release asset with sha256sums - that release is what
+# tools/install-deb.sh downloads.
+#
+# Tags are the version with a v prefix. Since 10.12.0 the versions are
+# date-based MM.DD.N (first release of a day = .0), so v10.12.0, v10.12.1, ...;
+# the older v0.x tags were semantic. Whatever the shape, the tag must equal the
+# top debian/changelog version (plus the -1 revision) or the mismatch gate below
+# refuses to publish - that check is what makes a bumped meson.build and a
+# forgotten changelog impossible to ship together.
 
 "on":
   push:

CHANGELOG.md

@@ -1,11 +1,17 @@
 # Changelog
 
 All notable changes to Nextless are documented here. The format follows
-[Keep a Changelog](https://keepachangelog.com/en/1.1.0/), the versions follow
-[Semantic Versioning](https://semver.org/spec/v2.0.0.html), and this project keeps one line per
+[Keep a Changelog](https://keepachangelog.com/en/1.1.0/), and this project keeps one line per
 user-visible change rather than one per commit.
 
-## [Unreleased]
+**Version scheme.** Since `10.12.0` the versions are date-based: `MM.DD.N`, where `N` counts
+releases made that day starting at 0 - so `10.12.0` is "October 12th, first cut", and a second
+release the same day would be `10.12.1`. The tag is the version with a `v` prefix (`v10.12.0`),
+which is what `.github/workflows/release.yml` matches and what must equal `debian/changelog`
+(`10.12.0-1`) or the publish step refuses to run. The earlier `0.x` line was semantic versioning;
+it stays as history and `dpkg` still orders it correctly (`10.12.0` upgrades `0.3.1`).
+
+## [10.12.0] - 2026-10-12
 
 ### Added
 - **The optional DeepFilterNet denoiser is now measured, not assumed**

FINDINGS.md

@@ -836,3 +836,23 @@ Roadmap 上 "Silence should be a no-op, not an `ASR error: empty result`" 的实
 #### 未处理风险
 - VAD 跑在 EBU R128 归一化**之后**，所以安静房间的底噪会被 +107 dB 的增益抬成"有语音"，静音仍然白花一次进程调用（约 1.4s）。正确修法是对归一化前的信号做 VAD，或给增益设上限（例如 30 dB）。这条与降噪顺序纠缠，单独处理。
 - `Nextless: no audio captured` / `microphone read failed` 这类状态同样没有清理时机，只是它们本来就该被看见。
+
+### 2026-10-12 — 版本号改成日期制（用户定的规则）
+
+#### 决策
+- 从 **`10.12.0`** 起不再用 semver，改成 **`MM.DD.N`**：当天第一个版本 `N=0`，同一天再发就是 `10.12.1`。tag 是版本号加 `v` 前缀（`v10.12.0`）。已有的 `0.2.0/0.3.0/0.3.1` 保留为历史，不回溯改写。
+- 规则写在三处，因为读的人不同：`CHANGELOG.md` 顶部（发布口径权威）、`release.yml` 头部注释（打 tag 的人先看到）、`debian/changelog` 条目里（打包者）。
+
+#### 发一个版本要动的地方（这次的全集）
+- `meson.build` 的 `version:` —— addon 版本是从它注入 `adapter/src/nextless-addon.conf.in` 的 `@PROJECT_VERSION@`，所以 **`fcitx5-diagnose` 显示的就是这一个值**。本机升之前显示 0.2.0、仓库已经是 0.3.1，说明那次之后没重装 addon；这就是「版本对不上到底哪来的」的答案。
+- `debian/changelog` 新条目 `fcitx5-nextless (<ver>-1)`：星期的英文缩写必须是**真实星期**（dpkg 校验；用 `LC_ALL=C date -d <day> +"%a, %d %b %Y %H:%M:%S %z"`），条目之间恰好一个空行。本地先 `dpkg-parsechangelog -l debian/changelog` 验，别等 CI。
+- `docs/nextless.1` 的 `.TH` 日期与版本两个参数。
+- `CHANGELOG.md`：`## [Unreleased]` 改名成 `## [<ver>] - <YYYY-MM-DD>`；release.yml 的发布说明就是按 `## [<ver>]` 抓这一段，抓不到只发警告不发正文。
+- tag 与 `debian/changelog` 不一致时 mismatch 门会**拒绝发布**，所以顺序固定为：bump PR → 合 main → 打 tag → push tag。
+
+#### 风险与判断
+- 排序没问题：dpkg 里 `10.12.0 > 0.3.1`（先比第一段），GitHub `releases/latest` 也跟新版本；`install-deb.sh` 只按 `${tag}-1_amd64.deb` 拼确定性文件名，对版本形状没有假设。
+- 代价要说清楚：日期版本号看不出「补丁还是破坏性变更」。缓解是发布说明必须有实质内容（现在强制从 CHANGELOG 取），以及 CHANGELOG 一行只写一条用户可见变化。
+
+#### 验证
+- `dpkg-parsechangelog` 解析新条目通过，时间戳 `Mon, 12 Oct 2026 08:50:00 +0800`；`meson test -C build` 21/21。

README.md

@@ -358,7 +358,7 @@ only way to get the 1.8 GB back, and `nextless-get-models` will re-download what
 **Check what is actually loaded**, because a stale install is otherwise invisible:
 
 ```bash
-fcitx5-diagnose | grep -i nextless     # e.g. "Nextless 0.3.1"
+fcitx5-diagnose | grep -i nextless     # e.g. "Nextless 10.12.0"
 ```
 
 If that prints an older version than the release you installed, the running addon is the old one -

README.zh.md

@@ -321,7 +321,7 @@ grep -v '^#' build/meson-logs/install-log.txt | sudo xargs -d '\n' rm -f --
 **确认到底加载的是哪个版本**，否则装了旧版你看不出来：
 
 ```bash
-fcitx5-diagnose | grep -i nextless     # 例如 "Nextless 0.3.1"
+fcitx5-diagnose | grep -i nextless     # 例如 "Nextless 10.12.0"
 ```
 
 如果这里显示的数字比你装的 release 旧，说明在跑的还是老 addon：要么安装那步没执行，

SECURITY.md

@@ -44,7 +44,9 @@ Use [GitHub Security Advisories](https://github.com/RevolutionLA/nextless/securi
 — private, and the right channel. Please do not open a public issue.
 
 I aim to reply within 7 days and ship a fix or a mitigation within 30. Reportedly affected
-versions: `0.3.1` and everything before it, including upstream `vinput`.
+versions: every release, `10.12.0` included (and upstream `vinput`) - assume your version is
+affected until an advisory says otherwise. Versions before `10.12.0` were numbered semver
+(`0.3.1` was the last of them).
 
 Not valid on its own: an open port, a missing HTTP header, a dependency advisory without an
 exploitable path through this add-on, or "the cloud backend sends my audio to a cloud backend".

debian/changelog

@@ -1,3 +1,26 @@
+fcitx5-nextless (10.12.0-1) unstable; urgency=medium
+
+  * Capture-path behaviour fix (issue #61): with denoise=deepfilter, both legs
+    of the 16k -> 48k -> 16k resample are now drained (soxr returns its
+    group-delay tail only when end-of-input is signalled with in == NULL, not
+    with ilen == 0), and deep-filter's own constant 30 ms is appended back from
+    the raw input. A denoised utterance therefore keeps its full length instead
+    of silently losing 69-97 ms off the end, which is where the last syllable
+    lives. A response more than 160 ms short is treated as a malfunction and
+    falls back to speexdsp rather than being padded into a half-denoised buffer.
+  * Benchmark harness: bench_asr --denoise (and BACKENDS="deepfilter" in
+    run_bench.sh) measures the add-on's own dfDenoise(). docs/benchmarks.md
+    carries the first real-binary DeepFilterNet row - RTF 0.13-0.16, ~1.0-1.3 s
+    per 5-10 s utterance, 60 MB - with the machine header and raw JSONL
+    committed next to it.
+  * README: the measured DeepFilterNet cost, and a new "Rolling back" section
+    (apt remove vs purge, downgrading with dpkg -i, the meson install-log
+    manifest, and fcitx5-diagnose as the which-version-is-loaded check).
+  * Versioning moves to date-based MM.DD.N (first release of a day is .0);
+    this is the first under that scheme and supersedes 0.3.1.
+
+ -- RevolutionLA <[email protected]>  Mon, 12 Oct 2026 08:50:00 +0800
+
 fcitx5-nextless (0.3.1-1) unstable; urgency=medium
 
   * Bug-fix release, no new configuration surface:

docs/nextless.1

@@ -1,5 +1,5 @@
 .\" Man page for fcitx5-nextless
-.TH nextless 1 "2026-10-09" "0.3.1" "fcitx5-nextless"
+.TH nextless 1 "2026-10-12" "10.12.0" "fcitx5-nextless"
 .SH NAME
 fcitx5\-nextless \- push-to-talk voice input addon for fcitx5
 .SH SYNOPSIS

meson.build

@@ -1,5 +1,5 @@
 project('fcitx5-nextless', 'cpp',
-  version: '0.3.1',
+  version: '10.12.0',
   default_options: ['cpp_std=c++20', 'warning_level=2', 'werror=true'])
 
 # warning_level=2 is -Wall -Wextra (plus -Werror); 3 would add -Wpedantic, and that one is not