feat(bench): reproduce xlings' generated lua-stdlib module — 82/83, only the link left

你说得对,我上一版把边界划早了。两件看起来是墙的事都不是:

* **转依赖的头全都能描述** —— 每一个都在 registry 里,CMakeLists 现在全找到了
  (mbedtls 经 mcpplibs tinyhttps 进来,lua 经 capi.lua)。
* **生成的模块也能复现** —— `mcpplibs.xpkg.lua_stdlib` 不是签入文件,由 libxpkg 的
  `build.mcpp` 生成;但它做的事就是把 **11 个 `.lua` 嵌成字符串**,
  `embed_lua_stdlib.cmake` 照着做了一遍。「mcpp 跑了个构建程序」本身不构成边界。

现在 **83 条边里过了 82 条:每一个翻译单元都编译成功,只剩链接。**
剩下的是普通工作量而不是墙:ftxui / libarchive / lua / mbedtls 都是**源码**、
由 mcpp 编译,所以链接期缺符号(`undefined reference to archive_entry_pathname` …);
它们各自都带 CMakeLists,`add_subdirectory` 就能收尾。

⚠️ 路上踩到的三个真问题,都写进了注释:
* CMake 引号字符串里的 `\;` 会把**反斜杠原样**写进生成的 C++,
  每一行都 `error: stray '\' in program` —— 改用 bracket 语法。
* 抄来的模块表**已经漂过一次**:第一版正则只抓到 11 个里的 10 个,
  而失败出现在三个文件之外的消费者身上(`'base64_lua' is not a member of ...detail`)。
  生成器现在按**文件缺失**失败,而不是信任那张表。
* 文件集名不能含 `.` 与 `-`;ftxui/libarchive 的头在版本目录**再下一层**。

Author: speak-agent

bench/projects/xlings/CMakeLists.txt

@@ -21,22 +21,21 @@
 # same compiler binary, same language flags, same source set, same link output
 # kind, same standard library (`import std;`, not a header shim).
 #
-# ⚠️ STATUS: CONFIGURES AND COMPILES xlings' OWN 110 UNITS; DOES NOT LINK.
+# ⚠️ STATUS: 82 of 83 edges — EVERY TRANSLATION UNIT COMPILES; ONLY THE LINK FAILS.
 #
-# The dependencies themselves describe fine — every header one of them needs is
-# unpacked in mcpp's registry and the list below finds all of them, transitively
-# (mbedtls arrives via mcpplibs tinyhttps, lua via capi.lua).
+# Two things that looked like boundaries and were not:
 #
-# The blocker is one level past that, and it is specific:
+#   * Transitive headers. Every one is unpacked in mcpp's registry and the list
+#     below finds them all (mbedtls via mcpplibs tinyhttps, lua via capi.lua).
+#   * `mcpplibs.xpkg.lua_stdlib`, which is GENERATED by that package's
+#     `build.mcpp` rather than checked in. It embeds eleven `.lua` files as
+#     strings — small and fully specified, so `embed_lua_stdlib.cmake`
+#     reproduces it. "mcpp runs a build program" is not by itself a boundary.
 #
-#     xpkg-executor.cppm:5  fatal error: unknown compiled module interface:
-#                           no such module          [mcpplibs.xpkg.lua_stdlib]
-#
-# `mcpplibs.xpkg.lua_stdlib` is not a checked-in file. It is GENERATED at build
-# time by that package's `build.mcpp` program — mcpp's build-program protocol.
-# No foreign build system can produce it without implementing that protocol, so
-# this is not a gap in the description; it is the boundary of what a description
-# can be.
+# What remains is ordinary: ftxui / libarchive / lua / mbedtls arrive as SOURCE
+# and mcpp compiles them, so the link wants symbols nobody built here
+# (`undefined reference to archive_entry_pathname`, ...). They all ship their
+# own CMakeLists, so `add_subdirectory` finishes this — it is work, not a wall.
 #
 # So the xlings arm compares **mcpp against mcpp** (releases, schedules), which
 # is what a control target is for: it answers "does this engine change hold on a
@@ -158,6 +157,31 @@ endfunction()
 
 xlings_add_source_dep(mcpplibs-x-cmdline   0.0.2)
 xlings_add_source_dep(mcpplibs-x-xpkg      0.0.57)
+
+# `mcpplibs.xpkg.lua_stdlib` is generated, not checked in — libxpkg's build.mcpp
+# embeds ten .lua files as strings. Reproduced here so both arms compile the
+# same set of translation units; see embed_lua_stdlib.cmake for why a copied
+# module list is acceptable and how it fails when it drifts.
+file(GLOB xpkg_vers "${MCPP_XPKGS}/mcpplibs-x-xpkg/0.0.57/*")
+foreach(d IN LISTS xpkg_vers)
+  if(IS_DIRECTORY "${d}/src/lua-stdlib")
+    set(XPKG_PKG_ROOT "${d}")
+  endif()
+endforeach()
+if(XPKG_PKG_ROOT)
+  set(LUA_STDLIB_CPPM "${CMAKE_CURRENT_BINARY_DIR}/generated/xpkg-lua-stdlib.cppm")
+  file(GLOB_RECURSE LUA_STDLIB_SOURCES "${XPKG_PKG_ROOT}/src/lua-stdlib/*.lua")
+  add_custom_command(
+    OUTPUT "${LUA_STDLIB_CPPM}"
+    COMMAND "${CMAKE_COMMAND}" -DXPKG_ROOT=${XPKG_PKG_ROOT} -DOUT=${LUA_STDLIB_CPPM}
+            -P "${CMAKE_CURRENT_SOURCE_DIR}/embed_lua_stdlib.cmake"
+    DEPENDS ${LUA_STDLIB_SOURCES} "${CMAKE_CURRENT_SOURCE_DIR}/embed_lua_stdlib.cmake"
+    COMMENT "Embedding libxpkg's lua-stdlib")
+  target_sources(xlings PRIVATE
+    FILE_SET fs_lua_stdlib TYPE CXX_MODULES
+      BASE_DIRS "${CMAKE_CURRENT_BINARY_DIR}/generated"
+      FILES "${LUA_STDLIB_CPPM}")
+endif()
 xlings_add_source_dep(mcpplibs-x-tinyhttps 0.2.9)
 xlings_add_source_dep(mcpplibs.capi-x-lua  0.0.3)
 

bench/projects/xlings/README.md

@@ -55,21 +55,27 @@ units, and compiles them. **It does not link**, and the reason is worth having
 written down, because it is the honest limit of a hand-written foreign build
 description rather than a gap in effort:
 
-The dependencies themselves describe fine: every header they need is unpacked
-in mcpp's registry, and `CMakeLists.txt` finds all of them **transitively** —
-`mbedtls` arrives through mcpplibs `tinyhttps`, `lua` through `capi.lua`.
-
-The blocker is one level past that, and it is specific:
-
-```
-xpkg-executor.cppm:5  fatal error: unknown compiled module interface: no such module
-                      [mcpplibs.xpkg.lua_stdlib]
-```
-
-`mcpplibs.xpkg.lua_stdlib` **is not a checked-in file**. That package generates
-it at build time with a `build.mcpp` program — mcpp's build-program protocol. No
-foreign build system can produce it without implementing that protocol, so this
-is not a gap in the description; it is the boundary of what a description can be.
+**82 of 83 edges: every translation unit compiles; only the link fails.**
+
+Two things looked like boundaries and were not:
+
+* **Transitive headers.** All of them are unpacked in mcpp's registry and
+  `CMakeLists.txt` finds them — `mbedtls` via mcpplibs `tinyhttps`, `lua` via
+  `capi.lua`.
+* **A generated module.** `mcpplibs.xpkg.lua_stdlib` is not checked in; libxpkg's
+  `build.mcpp` produces it. But all it does is embed eleven `.lua` files as
+  strings, so [`embed_lua_stdlib.cmake`](embed_lua_stdlib.cmake) reproduces it.
+  *"mcpp runs a build program"* is not by itself a boundary.
+
+What is left is ordinary work rather than a wall: `ftxui`, `libarchive`, `lua`
+and `mbedtls` arrive as **source** and mcpp compiles them, so the link asks for
+symbols nobody built here (`undefined reference to archive_entry_pathname`, …).
+Each ships its own CMakeLists, so `add_subdirectory` finishes the arm.
+
+⚠️ The copied module list in the generator **already drifted once**: a first
+regex caught ten of eleven entries, and the failure surfaced three files away as
+`error: 'base64_lua' is not a member of ...detail`. The generator now fails on a
+missing `.lua` rather than trusting the list.
 
 **So the xlings arm compares mcpp against mcpp** (two releases, or two
 schedules). That is what a control target is for: it answers *"does this engine

bench/projects/xlings/embed_lua_stdlib.cmake

@@ -0,0 +1,81 @@
+# Reproduce `mcpplibs.xpkg.lua_stdlib` for the cmake arm of the benchmark.
+#
+# That module is not a checked-in file: the xpkg package generates it at build
+# time with a `build.mcpp` program. What the program does, though, is small and
+# fully specified — it embeds ten `.lua` files as strings — so a foreign build
+# system CAN reproduce it, and "mcpp runs a build program" is not by itself a
+# boundary. Reproducing it is what keeps the cross-engine comparison honest:
+# both arms then compile the same set of translation units.
+#
+# ⚠️ THE MODULE LIST IS COPIED, and copies drift — this one already did. A first
+# pass extracted ten of the eleven entries (a regex that missed `base64_lua`),
+# and the failure was not "list incomplete" but
+#
+#     xpkg-executor.cppm:585  error: 'base64_lua' is not a member of ...detail
+#
+# i.e. it surfaced in a consumer, three files away from the cause. The guard
+# below therefore fails on a MISSING FILE rather than trusting the list; the
+# alternative — quietly embedding ten of eleven — produces a binary that differs
+# from mcpp's while the benchmark reports a clean run.
+#
+# Regenerate with:
+#   grep -oE '\{ *"[A-Za-z0-9_]+" *, *"[^"]+\.lua" *\}' <pkg>/build.mcpp
+#
+# Usage (from add_custom_command):
+#   cmake -DXPKG_ROOT=<pkg> -DOUT=<file> -P embed_lua_stdlib.cmake
+
+if(NOT XPKG_ROOT OR NOT OUT)
+  message(FATAL_ERROR "embed_lua_stdlib.cmake needs -DXPKG_ROOT= and -DOUT=")
+endif()
+
+# (variable name, path relative to the package root) — mirrors MODULES in
+# libxpkg's build.mcpp.
+set(LUA_MODULES
+  "prelude_lua|src/lua-stdlib/prelude.lua"
+  "log_lua|src/lua-stdlib/xim/libxpkg/log.lua"
+  "pkginfo_lua|src/lua-stdlib/xim/libxpkg/pkginfo.lua"
+  "system_lua|src/lua-stdlib/xim/libxpkg/system.lua"
+  "subos_lua|src/lua-stdlib/xim/libxpkg/subos.lua"
+  "xvm_lua|src/lua-stdlib/xim/libxpkg/xvm.lua"
+  "utils_lua|src/lua-stdlib/xim/libxpkg/utils.lua"
+  "pkgmanager_lua|src/lua-stdlib/xim/libxpkg/pkgmanager.lua"
+  "elfpatch_lua|src/lua-stdlib/xim/libxpkg/elfpatch.lua"
+  "json_lua|src/lua-stdlib/xim/libxpkg/json.lua"
+  "base64_lua|src/lua-stdlib/xim/libxpkg/base64.lua")
+
+# Bracket syntax, not a quoted string: a quoted CMake string needs `\;` for a
+# literal semicolon, and that backslash reaches the generated C++ verbatim —
+# `error: stray '\' in program` on every line of the module preamble.
+set(text [[// Generated by bench/projects/xlings/embed_lua_stdlib.cmake — do not edit.
+// Mirrors what libxpkg's build.mcpp produces; edit the .lua sources.
+module;
+export module mcpplibs.xpkg.lua_stdlib;
+import std;
+
+export namespace mcpplibs::xpkg::detail {
+
+]])
+
+foreach(entry IN LISTS LUA_MODULES)
+  string(REPLACE "|" ";" parts "${entry}")
+  list(GET parts 0 var)
+  list(GET parts 1 rel)
+  set(src "${XPKG_ROOT}/${rel}")
+  if(NOT EXISTS "${src}")
+    message(FATAL_ERROR
+      "lua-stdlib source missing: ${src}\n"
+      "the copied module list has drifted from libxpkg's build.mcpp")
+  endif()
+  file(READ "${src}" body)
+  # A C++ raw string literal, so nothing in the Lua needs escaping. The
+  # delimiter is one no Lua file contains; if that ever stops being true the
+  # generated file will not compile, which is the loud failure we want.
+  string(APPEND text "inline const std::string_view ${var} = R\"XLUA(${body})XLUA\"")
+  string(APPEND text ";\n\n")
+endforeach()
+
+string(APPEND text "} // namespace mcpplibs::xpkg::detail\n")
+
+get_filename_component(outdir "${OUT}" DIRECTORY)
+file(MAKE_DIRECTORY "${outdir}")
+file(WRITE "${OUT}" "${text}")