mcpp 2026.9.26.1 as the measured client, and the graph's own Linux targets in the openkal measurement

Pins move, with no registration in this change (openkal-musl 0.19.2 and
openkal-llvm-runtime 0.15.2 were registered in #467 and #468 and resolved from
the published index before anything here named them):

- validate.yml and openkal-compat.yml install mcpp 2026.9.26.1, and
  index.toml's latest_mcpp names it. min_mcpp does not move: the release
  changes no descriptor grammar.
- tests/openkal/pins.toml: runtime 0.15.2 (which pins openkal-musl 0.19.2
  exactly), mcpp 2026.9.26.1, and two targets, x86_64-linux-musl and
  aarch64-linux-musl. From mcpp 2026.9.26.1 a link over the graph's C library
  searches no host library directory (mcpp-community/mcpp#696); before it, a
  member's `-lm` on x86_64-linux-musl linked the host's glibc objects without
  a diagnostic. Neither target was measured here, which is why neither showed.
- compat.py: a Linux image of the host's own architecture runs without a
  runner (`host_executes`), so x86_64-linux-musl is measured as run rather
  than `--no-run`; aarch64-linux-musl runs under qemu-aarch64, which the
  workflow now installs. Three selftest cases state the rule.
- compat.libaio, compat.libdrm and compat.libinput: the comments that blamed
  `c_standard = "gnu11"` itself now name the defect (a dependency's own C
  standard was dropped up to mcpp 2026.9.25.1, mcpp-community/mcpp#695) and
  keep the `-D` spellings, which work under every engine this index admits.

Author: speak-agent

.github/workflows/openkal-compat.yml

@@ -38,7 +38,7 @@ env:
   # run if this drifts from pins.toml. See the comment beside `mcpp` in
   # pins.toml for why this pin, not just validate.yml's, is gated by the
   # index floor.
-  MCPP_VERSION: "2026.9.21.3"
+  MCPP_VERSION: "2026.9.26.1"
   # CROSS-VALIDATION BEFORE A RELEASE, NOT AFTER IT.
   #
   # Empty in an ordinary run, so this measurement keeps using the RELEASED
@@ -140,16 +140,18 @@ jobs:
       # The Windows cross headers of the host are installed deliberately. A
       # graph that reached them would build differently with them present, so a
       # result that is the same here and on a runner without them is evidence
-      # that the graph is closed. Wine runs the Windows members' tests.
-      - name: Host headers the graph must not reach, and a runner for Windows
+      # that the graph is closed. Wine runs the Windows members' tests, and
+      # qemu-aarch64 the aarch64-linux-musl members' (tests/openkal/pins.toml).
+      - name: Host headers the graph must not reach, and runners for Windows and aarch64
         if: steps.select.outputs.members != ''
         shell: bash
         run: |
           sudo apt-get update -qq
-          sudo apt-get install -y -qq mingw-w64 wine64 > /dev/null
+          sudo apt-get install -y -qq mingw-w64 wine64 qemu-user > /dev/null
           ls /usr/x86_64-w64-mingw32/include/windows.h
           command -v wine || sudo ln -s "$(command -v wine64)" /usr/local/bin/wine
           wine --version
+          qemu-aarch64 --version | head -1
 
       - name: The pins agree with this workflow
         if: steps.select.outputs.members != ''

.github/workflows/validate.yml

@@ -313,7 +313,20 @@ env:
   # cache epoch moves from 3 to 4 in the same release, so the first run is cold.
   #
   # `index.toml` min_mcpp does not move: no descriptor grammar changed.
-  MCPP_VERSION: "2026.9.25.1"
+  # ── 2026-09-26: raised 2026.9.25.1 -> 2026.9.26.1 ─────────────────────────
+  #
+  # 2026.9.26.1 changes three things a member's build can see
+  # (mcpp-community/mcpp#693-#696): a package's `c_standard` reaches that
+  # package's own C units and no one else's; a link over a graph-supplied C
+  # library searches no host library directory; and paths are held as UTF-8,
+  # so a project in a directory whose name is not ASCII builds on Windows. The
+  # three descriptors that spell `-D_GNU_SOURCE` because a dependency's
+  # `gnu11` used to be dropped keep the define, which works under every engine
+  # this index admits. The raise is the measurement of that on every member
+  # and platform.
+  #
+  # `index.toml` min_mcpp does not move: no descriptor grammar changed.
+  MCPP_VERSION: "2026.9.26.1"
 
 jobs:
   lint:

docs/openkal-compat.md

@@ -175,10 +175,15 @@ change can affect:
 Adding a member therefore measures that member, not the whole list. The
 workflow does not block a merge unless the comparison below is enabled.
 
-The targets are `x86_64-linux-gnu` and `x86_64-windows-musl`. The Windows
-target is named for the C environment the graph presents there, which is
-musl's, not MinGW's. Results measured before 2026-09-23 are recorded as
-`x86_64-windows-gnu`, the name that target had then.
+The targets are `x86_64-linux-gnu`, `x86_64-windows-musl`, `x86_64-linux-musl`
+and `aarch64-linux-musl`. The Windows target is named for the C environment the
+graph presents there, which is musl's, not MinGW's. Results measured before
+2026-09-23 are recorded as `x86_64-windows-gnu`, the name that target had then.
+The two `linux-musl` targets joined on 2026-09-26, with mcpp 2026.9.26.1: a link
+over the graph's C library searches no host library directory from that release
+(mcpp-community/mcpp#696), and these are the targets where that link is the
+graph's own. `x86_64-linux-musl` runs on the host as it is, and
+`aarch64-linux-musl` runs under qemu-aarch64.
 
 It installs the Windows cross toolchain's host headers on purpose. A build that
 reached the host's headers would change its result when they are present, so a

index.toml

@@ -103,7 +103,12 @@
 # The pull request that moves validate.yml's pin to 2026.9.25.1 runs every
 # member on every platform with that release, which is what "known good"
 # means for this hint. The release changes no descriptor grammar.
+# ── 2026-09-26: latest_mcpp -> 2026.9.26.1, min_mcpp unchanged ─────────────
+#
+# Moved in the pull request that moves validate.yml's pin and the openkal
+# measurement's pin to 2026.9.26.1, which run every member on every platform
+# with it. The release changes no descriptor grammar.
 [index]
 spec        = "1"
 min_mcpp    = "2026.9.18.3"
-latest_mcpp = "2026.9.25.1"
+latest_mcpp = "2026.9.26.1"

pkgs/c/compat.libaio.lua

@@ -22,11 +22,13 @@
 -- `_body_io_syscall` macro), and `sigset_t` never arrives, so even the PUBLIC
 -- header fails to parse at `io_pgetevents(…, sigset_t *sigmask)`. Upstream
 -- never hits this because its Makefile compiles in the compiler's default gnu
--- mode. Declaring `c_standard = "gnu11"` looks like the tidier fix and is a
--- trap: mcpp 2026.8.27.2 accepts the string and still emits `-std=c11`
--- (verified in the emitted compile_commands.json), so the build fails exactly
--- as if nothing had been declared. The define is the spelling that actually
--- takes effect.
+-- mode. `c_standard = "gnu11"` is the declaration upstream's build implies,
+-- and mcpp up to 2026.9.25.1 did not apply it to a package in the dependency
+-- position: the root's value reached every C unit of the graph and a
+-- dependency's own was dropped, so the build failed exactly as if nothing had
+-- been declared (verified in the emitted compile_commands.json under
+-- 2026.8.27.2). mcpp 2026.9.26.1 applies it (mcpp-community/mcpp#695); the
+-- define stays because it works under every engine this index admits.
 --
 -- Two further GNU-isms ride along and are fine under plain `-std=c11` because
 -- gcc/clang only diagnose them under -pedantic: `syscall.h`'s named-variadic

pkgs/c/compat.libdrm.lua

@@ -110,9 +110,11 @@ package = {
         import_std = false,
         c_standard = "c11",
 
-        -- `c_standard = "gnu11"` is accepted and silently emits -std=c11, so
-        -- _GNU_SOURCE has to be spelled out. Without it O_CLOEXEC, `asprintf`
-        -- and `major`/`minor` are hidden and xf86drm.c does not compile.
+        -- _GNU_SOURCE is spelled out because mcpp up to 2026.9.25.1 dropped a
+        -- dependency's own `c_standard = "gnu11"` (mcpp-community/mcpp#695;
+        -- applied from 2026.9.26.1), and the define works under every engine
+        -- this index admits. Without it O_CLOEXEC, `asprintf` and
+        -- `major`/`minor` are hidden and xf86drm.c does not compile.
         -- -fPIC because this becomes a .so; -fvisibility=hidden to match
         -- upstream's `gnu_symbol_visibility : 'hidden'`, so only the symbols
         -- libdrm_macros.h marks drm_public are exported.

pkgs/c/compat.libinput.lua

@@ -75,9 +75,11 @@ package = {
         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.
+        -- note in cflags. `c_standard = "gnu11"` was tried here first, when
+        -- mcpp (up to 2026.9.25.1) applied only the root package's value and
+        -- dropped a dependency's own (mcpp-community/mcpp#695, fixed in
+        -- 2026.9.26.1); the defines below work under every engine this index
+        -- admits, so they stay.
         c_standard = "c11",
 
         -- `"*"` is the package ROOT, and it is here for exactly one file:
@@ -240,10 +242,11 @@ package = {
             -- 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.
+            -- This is a `-D` rather than `c_standard = "gnu11"` because a
+            -- dependency's own C standard did not reach the compiler before
+            -- mcpp 2026.9.26.1 (mcpp-community/mcpp#695): gnu11 was set here
+            -- and these exact errors survived it. The `-D` works under every
+            -- engine this index admits.
             --
             -- Safe because `__typeof__` IS `typeof`, always available in any
             -- dialect, and nothing can be named `typeof` in code that expects

tests/openkal/compat.py

@@ -101,6 +101,20 @@ def host_triple() -> str:
     return f"{arch}-{system}"
 
 
+def host_executes(target: str, host: str | None = None) -> bool:
+    """Whether this host runs the target's programs itself, with no runner.
+
+    A Linux host executes a Linux image of its own architecture whatever C
+    library the image carries: an `x86_64-linux-musl` program is a static
+    image, and mcpp runs it on an `x86_64-linux-gnu` machine as it is (`mcpp
+    test --target x86_64-linux-musl` needs no runner). Measured with `--no-run`
+    instead, such a cell would record `builds` for a member that runs."""
+    h = (host or host_triple()).split("-")
+    t = target.split("-")
+    return len(h) >= 2 and len(t) >= 2 and h[1] == "linux" and t[1] == "linux" \
+        and h[0] == t[0]
+
+
 def packages_of(manifest: dict) -> list[str]:
     """The index packages a member depends upon, as `namespace.name`."""
     found: list[str] = []
@@ -241,7 +255,8 @@ def measure(member: str, target: str, pins: dict) -> dict:
     toolchain = pins["toolchain"]
     native = target == host_triple()
     runner = (pins.get("runners") or {}).get(target)
-    can_run = bool(native or (runner and shutil.which(runner[0])))
+    can_run = bool(native or host_executes(target)
+                   or (runner and shutil.which(runner[0])))
     cmd = command_for(target, toolchain, native, can_run)
     proc = subprocess.run(cmd, cwd=work, capture_output=True, text=True,
                           timeout=pins.get("timeout", 3600))
@@ -533,6 +548,14 @@ def cmd_selftest(_args: argparse.Namespace) -> int:
         ("the host names no target and runs them",
          command_for("x86_64-linux-gnu", "llvm@22.1.8", True, True),
          ["mcpp", "test", "--toolchain", "llvm@22.1.8"]),
+        # A static musl image of the host's own architecture is run as it
+        # is; a runner is for what the host cannot execute.
+        ("a Linux image of the host's architecture runs without a runner",
+         host_executes("x86_64-linux-musl", "x86_64-linux-gnu"), True),
+        ("another architecture needs a runner",
+         host_executes("aarch64-linux-musl", "x86_64-linux-gnu"), False),
+        ("and so does another system",
+         host_executes("x86_64-windows-musl", "x86_64-linux-gnu"), False),
     ]
 
     # `select` decides how much of a pull request's four-hour matrix runs. It

tests/openkal/pins.toml

@@ -4,7 +4,7 @@
 # `runtime` is the C++ runtime for openkal; it pins openkal-musl exactly, and
 # openkal-musl selects the openkal implementation for each target, so this one
 # version names the whole graph.
-runtime   = "0.15.1"
+runtime   = "0.15.2"
 toolchain = "llvm@22.1.8"
 # `mcpp` here moves together with index.toml's `min_mcpp` and
 # .github/workflows/openkal-compat.yml's own MCPP_VERSION (that workflow
@@ -18,6 +18,28 @@ toolchain = "llvm@22.1.8"
 # index requires mcpp >= 2026.9.18.1 but this is mcpp 2026.9.17.3 [E0006]"
 # before a single line of the member's own source is read.
 #
+# 2026-09-26 (fourth move): `runtime` -> 0.15.2 and `mcpp` -> 2026.9.26.1,
+# AND THE GRAPH'S OWN LINUX TARGETS JOIN THE MEASUREMENT.
+#
+# mcpp 2026.9.26.1 links a program over a graph-supplied C library with no
+# host library directory in its search (mcpp-community/mcpp#696). Before it,
+# clang derived `/usr/lib/x86_64-linux-gnu` and its siblings for such a link,
+# so a member's `-lm` on x86_64-linux-musl linked the HOST's glibc objects into
+# a musl image without a diagnostic, and on aarch64-linux-musl failed on the
+# host's x86_64 linker script. Neither target was measured here, which is why
+# neither showed: the two rows measure exactly that link.
+#
+# 0.15.2 pins openkal-musl 0.19.2, which ships musl's eight empty archives
+# (`libm.a`, `libpthread.a`, ...), so those names are answered by the graph.
+# The runtime pins openkal-musl exactly, so `runtime` alone names the whole
+# graph, as the header above says. mcpp moves in the same change because the
+# rows mean nothing under an engine that still searches the host.
+#
+# x86_64-linux-musl runs natively; aarch64-linux-musl runs under
+# qemu-aarch64, which openkal-compat.yml installs. Every link was registered,
+# published and verified resolvable from the PUBLISHED index before this pin
+# moved (#467, #468).
+
 # 2026-09-21 (third move): `runtime` -> 0.15.1, AND THIS ONE IS A WINDOWS FIX
 # REACHING THE MEASUREMENT THROUGH FOUR PACKAGES.
 #
@@ -114,16 +136,19 @@ toolchain = "llvm@22.1.8"
 # openkal-musl 0.16.0 and openkal-llvm-runtime 0.12.0), because a pin that
 # moves onto an unreachable or divergent asset fails every member at once and
 # reports it as a compatibility result.
-mcpp      = "2026.9.21.3"
+mcpp      = "2026.9.26.1"
 
 # The Windows target is named for the C environment the graph presents there,
 # which is musl's (openkal-musl declares `[c-abi] presents = "posix"`), not
 # MinGW's: `x86_64-windows-gnu` read as a mingw-w64 build, which this graph
 # has never been. Results measured before 2026-09-23 are recorded under that
 # older name.
-targets = ["x86_64-linux-gnu", "x86_64-windows-musl"]
+targets = ["x86_64-linux-gnu", "x86_64-windows-musl", "x86_64-linux-musl", "aarch64-linux-musl"]
 
 # A target that is not the host is tested through its runner when the runner is
 # on PATH, and built otherwise.
 [runners]
 x86_64-windows-musl = ["wine"]
+# x86_64-linux-musl needs none: a Linux host runs a Linux image of its own
+# architecture as it is (compat.py `host_executes`).
+aarch64-linux-musl = ["qemu-aarch64"]