feat(glx-runtime): GL 来自生态而非 /usr/lib,这次带上缺失的那个测试

.github/workflows/validate.yml

@@ -127,7 +127,19 @@ env:
   # index.toml 的 min_mcpp **不动**。两个会撞的 compat 邻居都是 Form B,所以旧客户端
   # 撞上时是响亮报错而不是静默用错包;下限是一道让整个索引对旧客户端失效的闸门
   # (mcpp#349),只该在描述符真的读不动时抬。这里读得动,差的是解析得对。
-  MCPP_VERSION: "2026.8.6.2"
+  #
+  # 2026.8.8.2 是本 PR 的**前置**,不是顺手升级。
+  #
+  # `compat.glx-runtime` 依赖 mesa,mesa 声明 `xim:glibc@>=2.38`;下限被任何更高
+  # 版本满足,于是安装图形栈会在既有 glibc 旁边**再装一个**。2026.8.8.2 之前的
+  # mcpp 用 `readdir` 的第一项来解析「那个 glibc payload」,编译侧与产物的
+  # interpreter 因此可以指向不同版本 —— 这正是 #179 落地后 `asio-module` 和
+  # `core` 变红、并导致整份改动被 #180 revert 的原因。
+  #
+  # 用早于 2026.8.8.2 的 mcpp 重新落地这份改动,就是在仍会犯这个错的引擎上复现
+  # 事故条件。tests/check_graphics_install_side_effects.sh 也需要它 —— 该测试在
+  # 更旧的 mcpp 上失败是**有意的**。
+  MCPP_VERSION: "2026.8.8.2"
 
 jobs:
   lint:
@@ -508,6 +520,48 @@ jobs:
           echo "plan=$(cat /tmp/plan.json)" >> "$GITHUB_OUTPUT"
           jq . /tmp/plan.json
 
+  # Installing the graphics stack must not change what UNRELATED members link
+  # against. This is the check #179 did not have: every test in this repo asks
+  # about its own package, so nothing asked whether installing one changes
+  # things for everyone else -- and `asio-module` / `core` went red on a
+  # change that never touched them.
+  #
+  # Its own job, on Linux, because it needs a real toolchain and a home it can
+  # watch payloads appear in. It reports INCONCLUSIVE (and fails) rather than
+  # green whenever its preconditions do not hold; see the script.
+  graphics-side-effects:
+    name: "graphics install: no side effects on unrelated members"
+    runs-on: ubuntu-latest
+    timeout-minutes: 60
+    steps:
+      - uses: actions/checkout@v4
+      - name: Download mcpp
+        shell: bash
+        env:
+          MCPP_ARCHIVE: mcpp-${{ env.MCPP_VERSION }}-linux-x86_64.tar.gz
+          MCPP_ROOT: mcpp-${{ env.MCPP_VERSION }}-linux-x86_64
+        run: |
+          curl -L -fsS -o "$MCPP_ARCHIVE" \
+            "https://github.com/mcpp-community/mcpp/releases/download/v${MCPP_VERSION}/${MCPP_ARCHIVE}"
+          tar -xzf "$MCPP_ARCHIVE"
+          root="$PWD/$MCPP_ROOT"
+          # MCPP_HOME is the tarball root, not ~/.mcpp.
+          #
+          # A released mcpp is self-contained: it resolves its registry from
+          # beside its own executable. Copying the payload tree into ~/.mcpp
+          # and pointing MCPP_HOME there therefore watches one home while the
+          # build uses another -- which this check detects and refuses to call
+          # green (it did, on the first run). Aligning them is the fix; the
+          # check was right.
+          echo "MCPP=$root/bin/mcpp" >> "$GITHUB_ENV"
+          echo "MCPP_HOME=$root" >> "$GITHUB_ENV"
+          echo "MCPP_VENDORED_XLINGS=$root/registry/bin/xlings" >> "$GITHUB_ENV"
+      - name: Side-effect check
+        shell: bash
+        run: |
+          "$MCPP" self config --mirror GLOBAL 2>/dev/null || true
+          bash tests/check_graphics_install_side_effects.sh
+
   workspace:
     # The shard suffix appears only when the platform is actually split.
     name: workspace (${{ matrix.platform }}${{ matrix.shards == 1 && '' || format(' {0}/{1}', matrix.shard, matrix.shards) }})
@@ -562,6 +616,32 @@ jobs:
           key: ${{ env.REGISTRY_CACHE_KEY }}
           restore-keys: |
             mcpp-registry-${{ runner.os }}-${{ env.MCPP_VERSION }}-
+      # Host tools built from source (protoc, grpc_cpp_plugin, …).
+      #
+      # Measured on the run that added grpc-codegen: 636s to build protoc and
+      # 660s to build grpc_cpp_plugin — 1296s of a 3363s member, repeated in
+      # every job of every run, because the registry cache above covers
+      # ~/.mcpp/registry and the tool store lives in ~/.mcpp/build-cache.
+      # `protobuf-protoc` and `grpc-module` build the SAME protoc, so a full
+      # run pays for it several times over.
+      #
+      # A coarse rolling key is safe here, which is why this is cheap. The
+      # store is content-addressed by `<pkg>@<version>/<hash>` and mcpp
+      # validates an entry FIELD BY FIELD against the recorded entry.json
+      # (epoch, target, host triple, compiler identity, profile, features,
+      # transitive dependency closure). A stale entry is not used; it is
+      # rebuilt. The worst case of a cache miss-match is the status quo.
+      #
+      # Deliberately NOT build-cache/v1/pkg: 6.0 GB measured locally against
+      # Actions' 10 GB per-repo budget, which would evict the registry cache
+      # this job needs more. The tool store is 116 MB.
+      - name: Cache host tool store
+        uses: actions/cache@v4
+        with:
+          path: ~/.mcpp/build-cache/v1/tool
+          key: mcpp-toolstore-${{ runner.os }}-${{ env.MCPP_VERSION }}-${{ github.run_id }}-${{ matrix.platform }}-${{ matrix.shard }}
+          restore-keys: |
+            mcpp-toolstore-${{ runner.os }}-${{ env.MCPP_VERSION }}-
       - name: Download mcpp
         shell: bash
         env:
@@ -577,6 +657,21 @@ jobs:
           root="$PWD/$MCPP_ROOT"
           mkdir -p "$HOME/.mcpp/registry"
           cp -a "$root/registry/." "$HOME/.mcpp/registry/"
+          # Point mcpp AT the home the caches above cover.
+          #
+          # Without this the copy above is one-way scenery: a released mcpp is
+          # self-contained and resolves its registry from beside its own
+          # executable, so every build used <tarball>/registry while the cache
+          # restored, and saved, ~/.mcpp/registry. Proof from the run that
+          # added grpc-codegen: abseil compiled from
+          # `<tarball>/registry/data/xpkgs/compat-x-abseil/...`, and
+          # xim:glibc@2.44 / xim:python@3.13.12 were downloaded again on a job
+          # that reported a registry cache HIT.
+          #
+          # With MCPP_HOME set, the restored registry is the one in play — so
+          # payloads stop being re-downloaded per job, and the tool store cache
+          # below lands where the build looks for it.
+          echo "MCPP_HOME=$HOME/.mcpp" >> "$GITHUB_ENV"
           if [[ "$RUNNER_OS" == "Windows" ]]; then
             echo "MCPP=$(cygpath -m "$root/${{ matrix.mcpp }}")" >> "$GITHUB_ENV"
             echo "MCPP_VENDORED_XLINGS=$(cygpath -m "$root/${{ matrix.xlings }}")" >> "$GITHUB_ENV"

mcpp.toml

@@ -19,6 +19,7 @@ members = [
     "tests/examples/catch2-v2",
     "tests/examples/catch2-v2-main",
     "tests/examples/cjson",
+    "tests/examples/cmdline",
     "tests/examples/cli11",
     "tests/examples/core",
     "tests/examples/curl",
@@ -61,6 +62,7 @@ members = [
     "tests/examples/libpng",
     "tests/examples/llamacpp",
     "tests/examples/llamacpp-metal",
+    "tests/examples/llmapi",
     "tests/examples/md4c",
     "tests/examples/spdlog-compiled",
     "tests/examples/tinyhttps",

pkgs/c/compat.glfw.lua

@@ -80,7 +80,7 @@ package = {
                 "src/posix_module.c",
             },
             deps = {
-                ["compat.glx-runtime"] = "2026.06.03",
+                ["compat.glx-runtime"] = "2026.08.08",
                 ["compat.x11"]       = "1.8.13",
                 ["compat.xcursor"]   = "1.2.3",
                 ["compat.xext"]      = "1.3.7",

pkgs/c/compat.glx-runtime.lua

@@ -2,13 +2,59 @@ package = {
     spec        = "1",
     namespace   = "compat",
     name        = "glx-runtime",
-    description = "Host GLVND/GLX/OpenGL runtime adapter for mcpp Linux window applications",
+    description = "GLVND/GLX/OpenGL runtime for mcpp Linux window applications, from the xlings graphics stack",
     licenses    = {"MIT"},
     repo        = "https://github.com/KhronosGroup/OpenGL-Registry",
     type        = "package",
 
+    -- WHERE THE GL RUNTIME COMES FROM, AND WHY IT CHANGED
+    --
+    -- Until 2026.08.08 this package symlinked the HOST's libGL/libEGL out of
+    -- /usr/lib*. That is the thing mcpp#352 is: the host's Mesa needs
+    -- GLIBC_2.43 and mcpp's payload glibc is 2.39, so the program linked
+    -- cleanly and exited 255 with no output. It is also the boundary the
+    -- xlings hermetic policy names first -- any .so under /usr/lib* or /lib*.
+    --
+    -- The runtime now comes from `xim:graphics`, the ecosystem's own stack:
+    -- 22 packages plus two sentinels that probe for a host-side userspace half
+    -- they do not own (the proprietary NVIDIA driver, WSL2's D3D12) and
+    -- succeed having linked nothing when it is absent. One dependency, every
+    -- host shape, no conditional in this file.
+    --
+    -- Measured on an NVIDIA host after the change: libEGL resolves to
+    -- xim-x-libglvnd/1.7.0/lib/libEGL.so.1 and GL_RENDERER is the GPU, not
+    -- llvmpipe. Both halves of that matter -- "a window appeared" is a false
+    -- pass, because llvmpipe renders one too.
     xpm = {
         linux = {
+            -- The whole hermetic graphics stack. A RUNTIME dep, not a build
+            -- one: nothing here compiles against it, the produced consumer
+            -- loads it.
+            --
+            -- PLATFORM level, beside the version entries rather than inside
+            -- one. Every other recipe in both indexes places it here, and the
+            -- first attempt at this change put it inside the 2026.08.08 entry:
+            -- the descriptor parsed, the stack was never installed, and the
+            -- install failed on the required-library check -- an error naming
+            -- libGL.so.1 rather than the misplaced key. Whether a per-version
+            -- `deps` is rejected or merely unread was not determined; what is
+            -- established is that it does not take effect.
+            --
+            -- It therefore also applies to the legacy 2026.06.03 entry below,
+            -- which does not use it. That costs a consumer still pinned there
+            -- a download it will not read, and the alternative -- deleting the
+            -- published version -- would break them outright.
+            deps = { runtime = { "xim:graphics" } },
+            ["2026.08.08"] = {
+                url    = {
+                    GLOBAL = "https://raw.githubusercontent.com/KhronosGroup/OpenGL-Registry/a30033d3e812c9bf10094f1010374a6b15e192eb/README.adoc",
+                    CN     = "https://gitcode.com/mcpp-res/glx-runtime/releases/download/2026.08.08/glx-runtime-2026.08.08.adoc",
+                },
+                sha256 = "ea68efce197e68413ebb62c51ab4bccfb2309a2fca776d31b49d972f59f3640e",
+            },
+            -- Kept so already-published consumers pinned to it keep resolving.
+            -- It sources libGL from the HOST and is the configuration behind
+            -- mcpp#352; new consumers must not pin it.
             ["2026.06.03"] = {
                 url    = {
                     GLOBAL = "https://raw.githubusercontent.com/KhronosGroup/OpenGL-Registry/a30033d3e812c9bf10094f1010374a6b15e192eb/README.adoc",
@@ -41,6 +87,7 @@ package = {
 }
 
 import("xim.libxpkg.pkginfo")
+import("xim.libxpkg.system")
 import("xim.libxpkg.log")
 
 local function sh_quote(value)
@@ -60,6 +107,26 @@ local function split_paths(value)
     return out
 end
 
+-- Where to take the GL libraries from.
+--
+-- The SUBOS VIEW (`<subos>/lib`), not a payload directory. A payload path pins
+-- a version, so a consumer's recorded RUNPATH would name mesa 25.0.7.1 forever
+-- and stop resolving the day it is upgraded; the view is the stable
+-- indirection -- the role /run/opengl-driver plays on NixOS. xlings repoints
+-- it as the active version changes and this package needs no new release.
+--
+-- The view also carries libc.so.6, crt1.o and the rest of the C runtime, and
+-- those must NEVER reach a consumer's RUNPATH: the consumer runs under mcpp's
+-- payload loader, and pairing one loader with another glibc's libc.so.6 faults
+-- inside the dynamic linker before main, with empty output. What keeps them
+-- out is the pattern list below -- so that list is a safety boundary, not a
+-- convenience, and nothing resembling `libc*` may ever be added to it.
+--
+-- MCPP_HOST_GL_LIBRARY_PATH still works and is now the ONLY door back to the
+-- host. Using it leaves the hermetic guarantee: the libraries it names were
+-- built against the host's glibc, and loading them under mcpp's payload glibc
+-- is exactly the configuration mcpp#352 reports. It exists for a machine whose
+-- GPU vendor the ecosystem does not cover yet.
 local function candidate_dirs()
     local out = {}
     local seen = {}
@@ -71,13 +138,12 @@ local function candidate_dirs()
     end
 
     for _, dir in ipairs(split_paths(os.getenv("MCPP_HOST_GL_LIBRARY_PATH"))) do
+        log.warn("MCPP_HOST_GL_LIBRARY_PATH names %s: GL will come from the "
+                 .. "host, which is the configuration behind mcpp#352", dir)
         add(dir)
     end
-    add("/lib/x86_64-linux-gnu")
-    add("/usr/lib/x86_64-linux-gnu")
-    add("/lib64")
-    add("/usr/lib64")
-    add("/usr/lib")
+
+    add(path.join(system.subos_sysrootdir(), "lib"))
     return out
 end
 
@@ -90,7 +156,11 @@ local host_gl_patterns = {
     "libEGL.so*",
     "libEGL_*.so*",
     "libGLES*.so*",
-    "libnvidia*.so*",
+    -- No libnvidia* here. The proprietary driver reaches the subos through
+    -- xim:nvidia-gl-host-link, which links it under the glvnd vendor names
+    -- already matched above; taking it by its own name would be a second
+    -- route to the same libraries, and the two would disagree the day the
+    -- driver is upgraded under us.
     "libglapi.so*",
     "libdrm*.so*",
     "libexpat.so*",
@@ -104,90 +174,43 @@ local required = {
     ["libGL.so.1"] = false,
 }
 
--- Is FILE a 64-bit ELF? e_ident[EI_CLASS] == ELFCLASS64.
---
--- Five bytes read directly. `file`/`readelf`/`patchelf` would each answer this
--- and each may be absent when a hook runs, and a probe that answers "cannot
--- tell" by assuming "fine" is the bug below.
-local function is_elf64(file)
-    local f = io.open(file, "rb")
-    if not f then return false end
-    local head = f:read(5)
-    f:close()
-    return head ~= nil and #head == 5
-       and head:sub(1, 4) == "\127ELF" and head:byte(5) == 2
-end
-
--- Link the host's GL runtime into one directory, FIRST HIT WINS, 64-bit only.
---
--- openxlings/xlings' mcpp#352: on Fedora 44 this produced
---     libGLX.so.0 -> /usr/lib/libGLX.so.0
--- a 32-bit library, and the application died with
---     libGLX.so.0: wrong ELF class: ELFCLASS32
---
--- TWO BUGS, and the obvious diagnosis ("the candidate order assumes Debian") is
--- not either of them -- `/usr/lib64` is already ahead of `/usr/lib` in the list:
---
---   1. `ln -sf` OVERWRITES. The loop reached /usr/lib64 first and linked the
---      correct file, then reached /usr/lib and replaced it. Last-wins, not
---      first-wins. `libOpenGL.so.0` survived as 64-bit purely because that host's
---      32-bit glvnd does not ship it -- which is why exactly one link in the bug
---      report was right.
---   2. NO ABI CHECK ANYWHERE, including in `required` below, which asserted that
---      libGLX.so.0 and libGL.so.1 EXIST. Both existed. Both were 32-bit.
---
--- There is no directory layout to assume: the FHS biarch clause makes /usr/lib
--- 32-bit (Fedora/RHEL/SUSE), Debian explicitly declined that clause and uses
--- /usr/lib/<triplet> so its /usr/lib is 64-bit, and Arch is a third answer
--- again. So the fix cannot be a better ordering -- it has to be an ABI check,
--- which makes the order stop mattering.
 local function link_runtime_libs(outdir)
     os.mkdir(outdir)
-    local claimed = {}
     for _, dir in ipairs(candidate_dirs()) do
         for _, pattern in ipairs(host_gl_patterns) do
-            -- Enumerate, then decide per file, instead of letting the shell
-            -- link them: the decision needs the ELF class and "have I already
-            -- taken this name", neither of which a `ln -sf` loop can express.
-            local pipe = io.popen("ls -1 " .. sh_quote(dir) .. "/" .. pattern
-                                  .. " 2>/dev/null")
-            if pipe then
-                for line in pipe:lines() do
-                    local lib = line:gsub("[\r\n]+$", "")
-                    local name = lib:match("[^/]+$")
-                    if lib ~= "" and name and not claimed[name]
-                       and is_elf64(lib) then
-                        claimed[name] = lib
-                        os.exec("ln -sf " .. sh_quote(lib) .. " "
-                                .. sh_quote(path.join(outdir, name)))
-                    end
-                end
-                pipe:close()
-            end
+            os.exec(
+                "for lib in " .. sh_quote(dir) .. "/" .. pattern ..
+                "; do [ -e \"$lib\" ] || continue; " ..
+                "ln -sf \"$lib\" " .. sh_quote(outdir) .. "/\"$(basename \"$lib\")\"; " ..
+                "done"
+            )
         end
     end
 
     for name, _ in pairs(required) do
-        local link = path.join(outdir, name)
-        -- Existence AND ABI. Existence alone passed on the Fedora host with
-        -- both links 32-bit, which is how a broken package reported success and
-        -- the failure surfaced as a silent exit code 255 from the application.
-        if not os.isfile(link) then
-            log.error("required host GL runtime library not found: %s", name)
-            log.error("  searched: %s", table.concat(candidate_dirs(), " "))
-            log.error("  install your distro's GL runtime (mesa / libglvnd)")
+        if not os.isfile(path.join(outdir, name)) then
+            log.error("%s is not in this subos. The GL runtime comes from "
+                      .. "`xim:graphics`; if it is declared and this still "
+                      .. "fires, the stack did not finish installing", name)
             return false
         end
-        if not is_elf64(link) then
-            log.error("host %s is not 64-bit (%s)", name, claimed[name] or link)
-            log.error("  a 32-bit library here fails at dlopen with")
-            log.error("  `wrong ELF class: ELFCLASS32` and the application")
-            log.error("  exits without output. Install the 64-bit GL runtime.")
+    end
+
+    -- Nothing resembling a C runtime may have come along. Asserted rather
+    -- than trusted: the pattern list is what keeps it out, and a pattern is
+    -- one careless edit away from matching more than it meant to. The failure
+    -- it prevents has no diagnostic of its own -- the consumer dies inside
+    -- the dynamic linker before main, printing nothing.
+    for _, bad in ipairs({"libc.so.6", "libc.so", "ld-linux-x86-64.so.2",
+                          "libpthread.so.0", "libdl.so.2", "libm.so.6"}) do
+        if os.isfile(path.join(outdir, bad)) then
+            log.error("%s was linked into the GL runtime directory. It would "
+                      .. "land on every consumer's RUNPATH and pair a second "
+                      .. "libc with mcpp's loader, which faults before main "
+                      .. "with no output at all", bad)
             return false
         end
     end
-    log.info("glx-runtime: linked %d host GL libraries (64-bit)",
-             (function() local n = 0 for _ in pairs(claimed) do n = n + 1 end return n end)())
     return true
 end