Fix legacy llama standard headers on macOS without replacing archives

Author: Sunrisepeak

.agents/docs/2026-10-25-unified-mwrap-execution.md

@@ -367,3 +367,18 @@ configs, 19 updater regressions, four tag regressions and skill validation pass.
 The ad0f019-based audit preserves all existing versions, with zero remaining
 manual commit snapshots. Final remote index CI, ecosystem review/squash, actual
 index publication and artifact/Git installed consumers remain required.
+
+
+Final b0e9be4 index CI found a legacy macOS regression in workspace default
+shard 1/2 (job 121081914195): ggml-org.llamacpp b10069.2 implementation
+llama-mmap.cpp lacks explicit cstdlib declarations for posix_memalign/free.
+The new mwrap llama consumer passed. Existing Linux installation adaptation
+already supplies cstdlib/cerrno; extended that same source-scoped manifest
+adaptation to native macOS. Raw historical archive URLs/hash/version remain
+unchanged; macOS packaging revision/floor 1 refreshes installed old manifests.
+The actual Lua install hook was executed against the public legacy manifest,
+and both native tables parse with the expected ordinary-source-only flags.
+This is not a skipped legacy test: complete final-head CI must run again.
+The previous strict auto-merge observer was stopped before preparing this fix.
+At b0e9be4, site, updater and OpenKAL workflows passed; validate did not.
+Failure log: /tmp/mwrap-final-index-macos-default-1-failed.log.

pkgs/g/ggml-org.llamacpp.lua

@@ -88,6 +88,9 @@ package = {
         },
         macosx = {
             ["b10069.2"] = {
+                -- Refresh older installs after the explicit standard-header fix.
+                revision = 1,
+                floor = 1,
                 url = {
                     GLOBAL = "https://github.com/mcpplibs/llama.cpp-m/archive/refs/tags/b10069.2.tar.gz",
                     CN     = "https://gitcode.com/mcpp-res/llamacpp/releases/download/b10069.2/llamacpp-b10069.2.tar.gz",
@@ -135,16 +138,19 @@ import("xim.libxpkg.pkginfo")
 
 function install()
     local wrap = "llama.cpp-m-" .. pkginfo.version()
-    -- libc++ 23 exposes missing standard headers in this Linux checkpoint.
-    -- Scope the explicit includes to ordinary Linux implementation sources;
+    -- Current libc++ exposes missing standard headers in this checkpoint.
+    -- Scope explicit includes to ordinary Linux/macOS implementation sources;
     -- the exported module and the downloaded source files remain untouched.
     if pkginfo.version() == "b10069.2" then
         local manifest = path.join(wrap, "mcpp.toml")
         local content = assert(io.readfile(manifest)):gsub("\r\n", "\n")
         local patched, count = content:gsub("\n%[target%.linux%.build%]\n",
             '\n[target.linux.build]\nflags = [{ glob = "third_party/llama.cpp/**/*.cpp", cxxflags = ["-include", "cstdlib", "-include", "cerrno"] }]\n', 1)
         assert(count == 1, "llamacpp: expected the native Linux build table")
-        io.writefile(manifest, patched)
+        local macos, macos_count = patched:gsub("\n%[target%.macos%.build%]\n",
+            '\n[target.macos.build]\nflags = [{ glob = "third_party/llama.cpp/**/*.cpp", cxxflags = ["-include", "cstdlib", "-include", "cerrno"] }]\n', 1)
+        assert(macos_count == 1, "llamacpp: expected the native macOS build table")
+        io.writefile(manifest, macos)
     end
     local prefix = pkginfo.install_dir()
     os.tryrm(prefix)