feat(libinput): 让输入链的顶端第一次真的被编译 (#298)

* feat(libinput): 让输入链的顶端第一次真的被编译

compat.libinput 随 #294 进了索引,而它从未被编译过——没有测试成员消费它,
而没人编译的包不会编译失败。加上 tests/examples/libinput 之后一次构建暴露出
四类问题,都是"上游用 meson,我们不用"才会有的:

1. 缺 libinput-version.h。meson 从 .h.in 生成,libinput-private.h:45 无条件
   include 它,所以 40 个源文件全挂。补成 generated_files。

2. config.h 只写了一半。HTTP_DOC_LINK / LIBINPUT_QUIRKS_OVERRIDE_FILE /
   LIBINPUT_PLUGIN_{LIB,ETC}DIR / HAVE_VERSIONSORT / HAVE_MTDEV 都缺。
   HAVE_VERSIONSORT 尤其关键:不定义它,libinput-versionsort.h 会给出自己的
   static strverscmp,而 glibc 已经 extern 声明过——是硬错误不是遮蔽。

3. include_dirs 少了包根。全树只有 libinput-plugin-mouse-wheel-lowres.c:31
   写 `#include "src/evdev-frame.h"`,上游因为从项目根编译而免费拿到这个拼法。

4. typeof。util-mem.h:180 的 (typeof(*ptr_))_steal(ptr_) 是 GNU 关键字,
   -std=c11 下被当成未声明函数调用,cast 塌成 int,伤害落在十几个根本没提
   typeof 的文件里,报成 -Wint-conversion。

   c_standard = "gnu11" 不是解法:mcpp 收下这个字符串,仍然发 -std=c11。
   这次是实测的——设过 gnu11,上面这些错误原样还在。所以走 -Dtypeof=__typeof__。

测试成员断言的是四包交接处(libudev 起 context、udev 枚举 seat、dispatch、
可 poll 的 fd),不断言能打开 /dev/input/event*——那需要权限,设
MCPP_RUN_INPUT_DEVICES=1 才要求。6 项全过。

顺带把 quirks 那段注释改准:LIBINPUT_QUIRKS_DIR 是有环境变量出口的
(libinput.c:1911),和 GBM_BACKENDS_PATH 同一形状,缺的是提供方而不是机制。

* docs: 记录 xkeyboard-config 跨索引闭环与 libinput 的四个 bug

§10.9.1 xkeyboard-config 已落地(xim-pkgindex#732),含实现中改的两处 discovery
形状:XKB_CONFIG_ROOT 是标量所以用 op=set(表里第一个非列表项),以及
declare_xkb —— 声明变量和放置内容是两件事,只声明的那一版让变量指向不存在的目录。

§10.9.2 新增:libinput quirks 是同一形状的第三处。之前注释里写的「编译期路径」
不准确,libinput.c:1911 是 getenv 优先。所以不是死路,是缺提供方。

§10.10 G5 从「待接上」改为已做,并把测试成员抓出的四个 bug 列出来——它们全都是
「包进了索引但从没被编译过」造成的。另加一张发现变量总表,标出哪两行仍是空的。

* docs(libinput): 对齐两处与改动后不符的注释

c_standard 从 gnu11 改回 c11 之后,config.h 里 HAVE_C23_AUTO 那段还写着
「builds as gnu11」。

以及 config.h 顶部对 LIBINPUT_QUIRKS_DIR 的说明还停在「编译期路径」,
与文件头已改准的版本不一致——libinput.c:1911 是 getenv 优先、编译期值兜底。

* docs: 补记从已发布索引的复验,以及它抓到的 store 撞名陷阱

合并 xim#732 后又从「已发布索引 + 全新 subos + 沙箱」复验一遍——上一次用的是
--add-xpkg 的本地副本,证明力不同。

第一次复验数据没放进去,差点得出「包坏了」。真因是 store 里还留着本地验证时的
local:xkeyboard-config@2.48,而 store 查找忽略 namespace,(name, version) 撞上
就把 install() 静默跳过、payload 为空,config() 照常跑。

顺带记下:declare_xkb 的 os.isdir 守卫确实触发了,但那条 log.warn 在 install
输出里没出现(同一次输出里 xlings 自己的 [warn] 是打出来的)。所以 declare_dri /
declare_gbm / declare_xkb 的警告都不能当诊断依赖。

Author: Sunrisepeak

.agents/docs/2026-08-30-graphics-stack-coverage-design.md

@@ -600,17 +600,141 @@ share/{bash-completion,man}` —— **不带 xkb 数据**。所以 `XKB_CONFIG_R
 ⚠ 顺带一提:一个**只需要 `xkb_keymap_new_from_string`** 的合成器不需要这份数据 ——
 那正是它从客户端收到 keymap 的路径,也是本索引测试跑通 parser 的方式。
 
+#### 10.9.1 已落地(openxlings/xim-pkgindex#732)
+
+上面的结论已实现,并在实现过程中改了 discovery 层的两处形状:
+
+**`XKB_CONFIG_ROOT` 是表里第一个非列表项。** DISCOVERY 的另外四行都是冒号分隔的搜索
+路径,`prepend` 在那里既正确又无损(而且必须是 prepend:一个提供方 `set` 会盖掉另一
+个,NVIDIA vendor 目录消失就是这么来的)。`XKB_CONFIG_ROOT` 是**标量** —— libxkbcommon
+读一次当一个目录用,`prepend` 上第二个提供方就变成 `dirA:dirB`,一个不存在的路径,而
+报错只会说「keymap 编译失败」。所以给 DISCOVERY 行加了可选的 `op`,默认仍是 `prepend`,
+这一行写 `set`。
+
+> 这与记忆里那条 `xvm envs 是 PATH 式合并`(标量 env 会被冒号拼接)是同一件事的第二次
+> 撞见。第一次的答案是「生成 launcher」;这次因为发现点在 discovery 层内部,所以修在
+> 了机制里。
+
+**声明变量 ≠ 放置内容。** 第一版只加了 DISCOVERY 行和 `declare_subos_env` 调用,装完之后:
+
+```
+XKB_CONFIG_ROOT=[.../subos/eco-2026-8-30-1/share/X11/xkb]
+ls: cannot access '.../share/X11/xkb': No such file or directory
+```
+
+变量在 shell 里读得好好的,指向不存在的目录。补 `graphics.declare_xkb`,与
+`declare_dri` / `declare_gbm` 同形(`xvm.files` 把树放进 subos)。
+
+**跨索引闭环已实测** —— mcpp 侧的 `freedesktop.libxkbcommon` 消费 xim 侧的数据集:
+
+```
+XKB_CONFIG_ROOT = .../subos/eco-2026-8-30-1/share/X11/xkb
+xkb_keymap_new_from_names(evdev/pc105/us) against that root ok
+   real layout: keycode 24 -> "q"
+…and the real us layout maps keycode 24 to "q"           ok
+0 check(s) failed
+```
+
+消费侧断言写在 `tests/examples/libxkbcommon`,且是**条件式**的:`XKB_CONFIG_ROOT` 未
+设置时只报告,设置了才要求编得出来。没有数据集的机器不是这个包的缺陷。
+
+**合并后又从「已发布索引 + 全新 subos + 沙箱」复验了一遍**,因为上面那次用的是
+`xlings config --add-xpkg` 的本地副本:
+
+```
+xim:xkeyboard-config@2.48 installed          ← 来自已发布索引,不是 local:
+ROOT=[.../subos/eco-2026-8-30-2/share/X11/xkb]
+compat geometry keycodes rules symbols types
+  rules/evdev 行数: 512
+  布局数: 151
+```
+
+**这次复验抓到一个真实陷阱,而且差点得出"包坏了"的结论**:第一次从已发布索引装,
+数据**没有**放进去。原因不在包 —— store 里还留着本地验证时的
+`local:xkeyboard-config@2.48`,而 xlings 的 store 查找**忽略 namespace**,
+`(name, version)` 撞上就把 `install()` 静默跳过,payload 是空的,而 `config()` 照常跑。
+清掉 store 重装即正常。
+
+**两条要记住的**:
+
+1. 用 `--add-xpkg` 本地验证过的包,发布后必须**先清 store 再验一遍**,否则验的
+   仍然是本地那份。
+2. `declare_xkb` 的 `os.isdir` 守卫确实触发了,但那条 `log.warn`
+   **在 install 输出里根本没出现**(同一次输出里 xlings 自己的 `[warn]` 是打出来的)。
+   所以 `declare_dri` / `declare_gbm` / `declare_xkb` 的警告都不能当诊断依赖 ——
+   判断内容有没有真放进去只能直接 `ls`。
+
+#### 10.9.2 同一形状的第三处:libinput quirks
+
+写 `compat.libinput` 时把 `LIBINPUT_QUIRKS_DIR` 也编译成空,注释里当时写的是「编译期
+路径,空即用内置默认」。**这句话不准确**:`libinput.c:1911` 是
+`getenv("LIBINPUT_QUIRKS_DIR")` 优先、编译期值兜底 —— 和 `GBM_BACKENDS_PATH`
+一模一样的形状。
+
+所以这不是死路,是**缺提供方**:`.quirks` 文件就在 libinput 自己的 tarball 里,但
+mcpp-index 没有发布数据目录的机制(与 §10.9 同因)。在有东西填上之前,libinput 打印
+
+```
+failed to find data files ... will negatively affect device behavior
+```
+
+并跑在内置默认上。这是**优雅降级**:枚举、事件、手势都正常,丢的是逐机型调校(比如
+某块触摸板的压力区间)。`tests/examples/libinput` 就是在这条消息存在的情况下全绿的。
+
+补法与 xkeyboard-config 完全对称(xim 数据包 + 一行 DISCOVERY),优先级低于布局数据:
+布局缺了是**硬失败**,quirks 缺了只是不够贴合。
+
 ### 10.10 仍未做
 
 - ~~G4b `libxkbcommon`~~ **已做**(fork,bison parser 预生成)。它需要的数据集见 §10.9:
   属于生态,不属于这里。
-- ~~G5 `libinput`~~ **描述符已做**;测试成员待本 PR 合并、`freedesktop.libevdev` 发布
-  后接上 —— 它同时需要 `compat.*` 和 `freedesktop.*` 两个 namespace 指向同一个
-  checkout,而成员级 `[indices]` 是替换而非合并。
+- ~~G5 `libinput`~~ **已做,并且测试成员一接上就抓出了四个 bug**(#298)。
+  形态问题解决得很朴素:这个成员**不声明 `[indices]`**,继承根的
+  `compat = { path = "." }` —— compat 来自 checkout,`freedesktop.libevdev` 来自已发布
+  索引(它确实已发布)。声明 `freedesktop` 反而会让 `compat.libinput` 去查已发布索引。
+
+  四个 bug 值得单列,因为它们**全都是"包进了索引但从没被编译过"造成的** ——
+  没有测试成员消费它,而没人编译的包不会编译失败(记忆里那条「绿 CI 不等于包被编译」
+  的教科书案例):
+
+  | 问题 | 症状落点 |
+  |------|---------|
+  | 缺 `libinput-version.h`(meson 从 `.h.in` 生成) | `libinput-private.h:45` 无条件 include,40 个源文件全挂 |
+  | `config.h` 只写了一半 | `HTTP_DOC_LINK` / `LIBINPUT_QUIRKS_OVERRIDE_FILE` / `LIBINPUT_PLUGIN_{LIB,ETC}DIR` / `HAVE_VERSIONSORT` / `HAVE_MTDEV` |
+  | `include_dirs` 少了包根 | 全树只有一个文件写 `#include "src/evdev-frame.h"` |
+  | `typeof` 是 GNU 关键字 | `-std=c11` 下 cast 塌成 `int`,错误落在十几个没提 typeof 的文件里 |
+
+  其中两条有普遍意义:
+
+  **`HAVE_VERSIONSORT`** —— 不定义它,`libinput-versionsort.h` 会给出自己的
+  `static strverscmp`,而 glibc 已经 `extern` 声明过。这是硬错误,不是遮蔽。
+  「少定义一个 HAVE_ 宏最多退化」的直觉在这里是错的。
+
+  **`c_standard = "gnu11"` 不生效,这次是实测的** —— mcpp 收下这个字符串,仍然发
+  `-std=c11`,那些 typeof 错误原样还在。所以走 `-Dtypeof=__typeof__`。这与记忆里
+  「c_standard 的 gnu 模式被静默忽略」一致,但那条记的是 `_GNU_SOURCE` 类的**库特性
+  宏**;`typeof` 是**语言方言**,`-D_GNU_SOURCE` 对它无效。两者要分开记。
 - ~~G6 `libudev` / `libseat`~~ **已做**,而且没有碰 systemd:libudev 用
   **libudev-zero**(三个实现里唯一既活着又可独立分发的),libseat 只开 seatd 与
   builtin 后端。两者的代价都在描述符里点名了。
-- **xkeyboard-config**:见 §10.9,应进 xim-pkgindex。
-
-前两条是普通工作量;G6 是需要决策的。合成器可以在「已有 DRM master」的前提下开发
-(从 TTY 直接启动、或 `SEATD_SOCK`),把 session 管理留到最后。
+- ~~**xkeyboard-config**~~ **已做**(xim-pkgindex#732),见 §10.9.1。跨索引闭环已实测。
+- **libinput quirks 数据**:见 §10.9.2。形状与 xkeyboard-config 对称,优先级更低 ——
+  布局缺了是硬失败,quirks 缺了只是不够贴合。**未做**。
+
+至此渲染链与输入链都不再有「静默落到宿主」的边:
+
+| 子系统 | 发现变量 | 提供方 | 状态 |
+|--------|---------|--------|------|
+| DRI 驱动 | `LIBGL_DRIVERS_PATH` | `xim:mesa` | ✅ |
+| EGL vendor | `__EGL_VENDOR_LIBRARY_DIRS` | `xim:mesa` + host-link 哨兵 | ✅ |
+| Vulkan ICD | `XDG_DATA_DIRS` / 共享 vendor 目录 | 同上 | ✅ #731 待合 |
+| GBM 后端 | `GBM_BACKENDS_PATH` | `xim:mesa` | ✅ |
+| 键盘布局 | `XKB_CONFIG_ROOT` | `xim:xkeyboard-config` | ✅ #732 待合 |
+| 输入 quirks | `LIBINPUT_QUIRKS_DIR` | *(无)* | ⬜ 优雅降级中 |
+| USB 名字库 | `USB_IDS_PATH` | *(无)* | ⬜ 优雅降级中 |
+
+最后两行是同一类:机制在、提供方缺、缺了只丢锦上添花的东西。**它们缺失时会说出来**
+—— 这正是当初把编译期默认值一律留空要换的东西。
+
+G6 仍是需要决策的:合成器可以在「已有 DRM master」的前提下开发(从 TTY 直接启动、或
+`SEATD_SOCK`),把 session 管理留到最后。

mcpp.toml

@@ -89,6 +89,7 @@ members = [
     "tests/examples/egl",
     "tests/examples/gl",
     "tests/examples/libevdev",
+    "tests/examples/libinput",
     "tests/examples/libseat",
     "tests/examples/libudev",
     "tests/examples/libxkbcommon",

pkgs/c/compat.libinput.lua

@@ -31,13 +31,25 @@
 -- way. Nothing in this index provides it, and without it libinput falls back
 -- to generic tablet handling rather than failing.
 --
--- The quirks database (`quirks.c`) reads `.quirks` files at runtime from a
--- directory compiled in as `LIBINPUT_QUIRKS_DIR`. That path is EMPTY here for
--- the same reason libgbm's backend path and libxkbcommon's config root are:
--- upstream's default points into the build prefix, which after relocation is
--- the HOST's dataset. Empty means devices get libinput's built-in defaults
--- rather than a host machine's model quirks — correct behaviour, one fewer
--- silent host edge.
+-- The quirks database (`quirks.c`) reads `.quirks` files at runtime — model
+-- specific tuning like a touchpad's pressure range. `libinput.c:1911` takes the
+-- directory from `getenv("LIBINPUT_QUIRKS_DIR")` and falls back to a
+-- compiled-in path, and the compiled-in path is EMPTY here for the same reason
+-- libgbm's backend path and libxkbcommon's config root are: upstream's default
+-- points into the build prefix, which after relocation is the HOST's dataset.
+--
+-- So this is a WIRING question, not a dead end — the same shape as
+-- GBM_BACKENDS_PATH, down to the environment variable. What is missing is a
+-- provider: the `.quirks` files ship inside libinput's own tarball, but
+-- mcpp-index has no way for a package to publish a data DIRECTORY, so nothing
+-- currently fills the variable. Until something does, libinput logs
+--
+--     failed to find data files ... will negatively affect device behavior
+--
+-- and runs on its built-in defaults. That is a real degradation and it is
+-- graceful: enumeration, events and gestures all work — what is lost is
+-- per-model tuning. Verified in tests/examples/libinput, which passes with the
+-- message present.
 package = {
     spec        = "1",
     namespace   = "compat",
@@ -62,30 +74,103 @@ package = {
     mcpp = {
         language   = "c++23",
         import_std = false,
+        -- Upstream asks for `c_std=gnu99` and means it — see the `typeof`
+        -- note in cflags. `c_standard = "gnu11"` was tried here first and is
+        -- NOT the fix: mcpp accepts the string and still emits `-std=c11`, so
+        -- the descriptor would claim a dialect the compiler never sees.
         c_standard = "c11",
 
-        include_dirs = { "*/src", "*/include", "mcpp_generated" },
+        -- `"*"` is the package ROOT, and it is here for exactly one file:
+        -- `libinput-plugin-mouse-wheel-lowres.c:31` writes
+        -- `#include "src/evdev-frame.h"` while every other source in the tree
+        -- writes `#include "evdev-frame.h"`. Upstream gets both spellings for
+        -- free because meson compiles from the project root; a package that
+        -- only puts `*/src` on the search path resolves 39 files and fails the
+        -- fortieth.
+        include_dirs = { "*", "*/src", "*/include", "mcpp_generated" },
 
         generated_files = {
             ["mcpp_generated/config.h"] = [==[
 #ifndef MCPP_LIBINPUT_CONFIG_H
 #define MCPP_LIBINPUT_CONFIG_H
 /* meson's config.h, written out for Linux with a GCC-compatible toolchain.
 
-   LIBINPUT_QUIRKS_DIR is EMPTY on purpose. quirks.c reads model-specific
-   `.quirks` files from it at runtime; upstream's default points into the
-   build prefix, which after relocation is the HOST's dataset. Empty means a
-   device gets libinput's built-in defaults instead of another machine's
-   quirks — the same stance compat.libgbm takes with GBM_BACKENDS_PATH.
+   LIBINPUT_QUIRKS_DIR is EMPTY on purpose, and it is a FALLBACK rather than
+   the only way in: `libinput.c:1911` reads `getenv("LIBINPUT_QUIRKS_DIR")`
+   first. Upstream's compiled-in default points into the build prefix, which
+   after relocation is the HOST's dataset — so empty means a device gets
+   libinput's built-in defaults instead of another machine's quirks, and an
+   environment that has a dataset can still name it. Exactly the shape
+   compat.libgbm has with GBM_BACKENDS_PATH, down to the variable.
 
    HAVE_LUA and HAVE_LIBWACOM are absent: both are optional upstream and both
    would add a dependency this package does not need to do its job. */
 #define LIBINPUT_QUIRKS_DIR ""
 #define LIBINPUT_QUIRKS_SRCDIR ""
+/* The user's quirks override, `<sysconfdir>/libinput/local-overrides.quirks`
+   upstream. Empty for the reason above: relocated, it names the HOST's file. */
+#define LIBINPUT_QUIRKS_OVERRIDE_FILE ""
+
+/* Plugin search paths, both EMPTY and both required to exist:
+   `libinput_plugin_system_append_default_paths` (libinput-plugin.c:387) names
+   them unconditionally, outside any HAVE_PLUGINS guard. Empty paths append
+   nothing, which is the behaviour this package wants — see the header comment
+   on why the Lua plugin system is off. */
+#define LIBINPUT_PLUGIN_LIBDIR ""
+#define LIBINPUT_PLUGIN_ETCDIR ""
+
+/* Printed into log messages that point a user at the documentation for the
+   behaviour being reported. meson builds it from the version: micro < 90 means
+   a release, so the URL names this release rather than `latest`. */
+#define HTTP_DOC_LINK "https://wayland.freedesktop.org/libinput/doc/1.31.3"
+
 #define HAVE_LIBEVDEV_DISABLE_PROPERTY 1
 #define HAVE_MEMFD_CREATE 1
 #define HAVE_LOCALE_H 1
 #define HAVE_STRERRORNAME_NP 1
+#define HAVE_SIGABBREV_NP 1
+#define HAVE_PIDFD_OPEN 1
+
+/* glibc has versionsort(3), and saying so is load-bearing rather than
+   cosmetic: `libinput-versionsort.h` provides its OWN static fallback when
+   this is unset, and a static definition of a name glibc already declared
+   extern is a hard error, not a shadow. */
+#define HAVE_VERSIONSORT 1
+
+/* mtdev is a declared dependency of this package, so the plugin that uses it
+   is compiled unconditionally (see `libinput-plugin-mtdev.c` in sources). */
+#define HAVE_MTDEV 1
+
+/* HAVE_C23_AUTO is deliberately ABSENT. meson probes for it by compiling
+   `auto foo = gmtime(NULL);`, which needs C23; this package builds as c11, so
+   the probe would fail here too and upstream's non-auto path is correct. */
+#endif
+]==],
+
+            -- meson's `src/libinput-version.h.in`, substituted for 1.31.3.
+            --
+            -- `libinput-private.h:45` includes it unconditionally, so every one
+            -- of the 40 sources below needs it — this is not an optional
+            -- convenience header. It went missing in the first version of this
+            -- descriptor and NOTHING CAUGHT IT: the package validated, entered
+            -- the index and shipped, because no test member consumed it and a
+            -- package nobody compiles cannot fail to compile. The
+            -- `tests/examples/libinput` member exists so that stays true only
+            -- until someone builds it — which is what found this.
+            --
+            -- The three numbers are the version in `xpm` above, split. They are
+            -- the package's public version macros, so a consumer doing
+            -- `#if LIBINPUT_VERSION_MAJOR >= 1` gets a real answer; leaving
+            -- them at 0 would compile just as well and lie.
+            ["mcpp_generated/libinput-version.h"] = [==[
+#ifndef LIBINPUT_VERSION_H
+#define LIBINPUT_VERSION_H
+
+#define LIBINPUT_VERSION_MAJOR 1
+#define LIBINPUT_VERSION_MINOR 31
+#define LIBINPUT_VERSION_MICRO 3
+#define LIBINPUT_VERSION "1.31.3"
+
 #endif
 ]==],
         },
@@ -144,6 +229,27 @@ package = {
         cflags = {
             "-D_GNU_SOURCE",
             "-DHAVE_CONFIG_H",
+
+            -- The GNU `typeof` keyword, supplied by hand.
+            --
+            -- `util-mem.h:180` is `(typeof(*ptr_))_steal(ptr_)` — the cast that
+            -- makes every `steal()` in the tree type-safe. Bare `typeof` is a
+            -- GNU extension (and C23's spelling); ISO C11 has only
+            -- `__typeof__`. Under `-std=c11` GCC parses `typeof(...)` as a call
+            -- to an undeclared function, the cast collapses to `int`, and the
+            -- damage lands as `-Wint-conversion` errors in a dozen unrelated
+            -- files that never mention typeof.
+            --
+            -- This is a `-D` rather than `c_standard = "gnu11"` because the
+            -- gnu dialects do not reach the compiler: mcpp accepts the string
+            -- and still emits `-std=c11`. Measured, not assumed — gnu11 was
+            -- set here and these exact errors survived it.
+            --
+            -- Safe because `__typeof__` IS `typeof`, always available in any
+            -- dialect, and nothing can be named `typeof` in code that expects
+            -- gnu99 anyway.
+            "-Dtypeof=__typeof__",
+
             "-fPIC",
         },
 

tests/examples/libinput/mcpp.toml

@@ -0,0 +1,22 @@
+# libinput test project — the top of the input chain.
+#
+# NO `[indices]` OF ITS OWN, and that is the interesting part.
+#
+# This member needs packages from TWO namespaces: `compat.libinput` (and the
+# `compat.libudev` / `compat.mtdev` it depends on) from this checkout, and
+# `freedesktop.libevdev` which libinput also links. A member-level `[indices]`
+# REPLACES the root's rather than merging, and one path can only be registered
+# under one namespace — so declaring `freedesktop = { path = "../../.." }` here
+# would make `compat.libinput` resolve against the published index instead.
+#
+# Inheriting the root's `compat = { path = "." }` is what works: the compat
+# packages come from the checkout, and `freedesktop.libevdev` resolves from the
+# PUBLISHED index — which is correct, because it is published. Same arrangement
+# tests/examples/asio-ssl relies on for the opposite reason.
+[package]
+name     = "libinput-tests"
+version  = "0.1.0"
+standard = "c++23"
+
+[target.'cfg(linux)'.dependencies.compat]
+libinput = "1.31.3"