feat(toolchain): aarch64 Linux supports musl-gcc, and the rest is deferred in the open

`available_toolchain_indexes()` branched per OS and never per architecture, so
an aarch64 Linux host was told llvm could be installed. Measured against the
index and upstream on 2026-08-26:

    xlings-res/llvm 20.1.7 / 22.1.8   no linux-aarch64 asset
    llvm/llvm-project 20.1.7, 21.1.0  no linux-aarch64 asset
    llvm/llvm-project 19.1.7          has one, too old for `import std`

so `mcpp toolchain install llvm 22.1.8` there is a 404 the list promised would
work. `mingw-cross-gcc` is x86_64-only for the same reason and gets the same
gate. What remains on that host is `musl-gcc`, which mcpp already routes gcc to
on non-x86_64 Linux, and `ninja`.

⭐ This is a POLICY statement — which families mcpp supports on this host, the
same kind of statement `tier` makes for a target row — not a copy of the index's
coverage. The plan that retires it is
.agents/docs/2026-08-26-aarch64-linux-ecosystem-closure.md §P1.

⚠️ AND ITS PREMISE IS RECHECKED EVERY RUN. `check_aarch64_llvm_deferral.sh`
fails when its reason stops holding — the day an aarch64 llvm is published — and
names the gate to remove. A deferral nobody rechecks is indistinguishable from a
defect. Network trouble is not read as "it appeared": an unreadable asset list
leaves the premise alone and says so.

Author: speak-agent

.agents/docs/2026-08-26-aarch64-linux-ecosystem-closure.md

@@ -198,7 +198,26 @@ P0 ──┬── P1(llvm 构建)──┬── P3(索引接线)── P4(mcpp
 - **判据**:`xlings install llvm@22.1.8` 在 aarch64 上成功;
   失败时的消息点名架构。
 
-### P4 — mcpp 对 aarch64 诚实
+### P4 — mcpp 对 aarch64 诚实(✅ 已落地)
+
+⭐ **决定:aarch64 上先只支持 `musl-gcc`,其余显式标记延缓。**
+
+- **P4.1 ✅** `available_toolchain_indexes()` 在非 x86_64 Linux 上不再列出
+  `llvm` 与 `mingw-cross-gcc`。⚠️ 这是一句**政策陈述**(mcpp 在这台宿主上支持
+  哪些族),不是索引数据的抄本 —— 与目标行的 `tier` 同类。
+- **P4.2 ✅ 延缓的前提每轮重测**:`.github/tools/check_aarch64_llvm_deferral.sh`
+  查 `xlings-res/llvm` 的 20.1.7 / 22.1.8 是否出现了 `linux-aarch64` 资产。
+  ⚠️ **它在理由不再成立时变红**,与一般的检查方向相反。
+  ⚠️ 网络故障不得被读成「出现了」:读不到资产表就说读不到,保持前提不动。
+- **P4.3 ✅ 生态 e2e 的豁免按理由给**:298 在 aarch64 上跳过,理由是
+  `llvm is not installed here`。按理由给而不是按宿主给,llvm 落地那天它自动
+  从「跳过」变回「断言」,workflow 一行都不用改。
+
+⚠️ **`host_can_serve()` 对裸机仍无条件 `true`** —— 那条理由(clang/lld 按构造
+就是交叉编译器)预设了 clang 在这台机器上存在。P4.1 的门让四个裸机行不再被
+列出,所以症状已经消失;这一处的**根因**留到 P1 之后再处理,因为届时它自然成立。
+
+### P4-原文 — mcpp 对 aarch64 诚实
 
 - **P4.1** `available_toolchain_indexes()` 目前按 OS 分支不问架构。
   ⚠️ **不要把索引数据抄进 mcpp** —— 那正是会漂移的形状。

.github/tools/check_aarch64_llvm_deferral.sh

@@ -0,0 +1,39 @@
+#!/usr/bin/env bash
+# ⭐⭐ A DEFERRAL'S PREMISE, RECHECKED.
+#
+# `available_toolchain_indexes()` omits llvm on non-x86_64 Linux because no
+# linux-aarch64 llvm exists — not in xlings-res, and not upstream since 19.x.
+# That is a deferral, and a deferral nobody rechecks is indistinguishable from
+# a defect.
+#
+# ⚠️ THIS FAILS WHEN THE REASON STOPS HOLDING, which is the opposite of what a
+# check usually does. The day an aarch64 llvm is published, it goes red and
+# names the gate to remove — see
+# `.agents/docs/2026-08-26-aarch64-linux-ecosystem-closure.md` §P1.
+#
+# ⚠️ Network trouble must not be read as "it appeared". An unreadable asset list
+# leaves the premise alone and says so: a check that turns a flaky API into a
+# claim about the world is worse than no check.
+set -uo pipefail
+
+fail=0
+for tag in 22.1.8 20.1.7; do
+    names="$(curl -sSL --retry 3 --retry-all-errors --max-time 60 \
+               "https://api.github.com/repos/xlings-res/llvm/releases/tags/$tag" 2>/dev/null \
+             | grep -oE '"name"[[:space:]]*:[[:space:]]*"[^"]+"' \
+             | sed -E 's/.*"([^"]+)"$/\1/')"
+    if [ -z "$names" ]; then
+        echo "  ?   $tag: could not read the asset list — premise left alone"
+        continue
+    fi
+    if printf '%s\n' "$names" | grep -q 'linux-aarch64'; then
+        echo "::error::xlings-res/llvm $tag now publishes a linux-aarch64 asset"
+        echo "        the deferral in available_toolchain_indexes() has outlived its reason"
+        echo "        see .agents/docs/2026-08-26-aarch64-linux-ecosystem-closure.md §P1"
+        fail=1
+        continue
+    fi
+    echo "  ok  $tag: still no linux-aarch64 asset"
+done
+[ "$fail" = 0 ] || exit 1
+echo "OK: the aarch64 llvm deferral still has its reason"

.github/workflows/ci-target-matrix.yml

@@ -88,6 +88,13 @@ jobs:
           done
           [ "$fail" = 0 ] || exit 1
 
+      # ⚠️ A DEFERRAL NOBODY RECHECKS IS INDISTINGUISHABLE FROM A DEFECT. This
+      # step fails when its reason STOPS holding — the day an aarch64 llvm is
+      # published — which is the opposite of what a check usually does.
+      - name: The aarch64 llvm deferral still has its reason
+        if: matrix.host == 'linux-aarch64'
+        run: bash .github/tools/check_aarch64_llvm_deferral.sh
+
       - name: Each invariant RAN
         run: |
           set -euo pipefail

src/toolchain/registry.cppm

@@ -640,11 +640,39 @@ bool host_can_serve(const triple::Triple& target) {
 }
 
 std::vector<AvailableIndex> available_toolchain_indexes() {
+    // ⚠️⚠️ NOT EVERY FAMILY EXISTS FOR EVERY (OS, ARCH), AND THIS LIST USED TO
+    // SAY OTHERWISE.
+    //
+    // The branches below are per-OS and there were none per-ARCH, so an aarch64
+    // Linux host was told llvm could be installed. Measured 2026-08-26 against
+    // the index and upstream:
+    //
+    //   xlings-res/llvm 20.1.7 / 22.1.8   no linux-aarch64 asset
+    //   llvm/llvm-project 20.1.7, 21.1.0  no linux-aarch64 asset
+    //   llvm/llvm-project 19.1.7          has one — too old for `import std`
+    //
+    // so `mcpp toolchain install llvm 22.1.8` there is a 404 that this list
+    // promised would work. Same family as the rest of this release: a table
+    // that answers a narrower question than the one it is asked.
+    //
+    // ⭐ THIS IS A POLICY STATEMENT, NOT A COPY OF THE INDEX. It says which
+    // families mcpp SUPPORTS on this host — the same kind of statement `tier`
+    // makes for a target row — and the plan that retires it is
+    // `.agents/docs/2026-08-26-aarch64-linux-ecosystem-closure.md` §P1.
+    //
+    // ⚠️ AND ITS PREMISE IS ASSERTED IN CI, so it cannot outlive its reason.
+    // `ci-target-matrix.yml`'s aarch64 job checks that no linux-aarch64 llvm
+    // asset has appeared; the day one does, that step reds and names this
+    // gate. A deferral nobody rechecks is indistinguishable from a defect.
+    const bool linuxNonX86 =
+        mcpp::platform::is_linux && mcpp::platform::host_arch != "x86_64";
+
     std::vector<AvailableIndex> out{
         { "gcc",      Family::Gcc },
         { "musl-gcc", Family::Gcc },
-        { mcpp::toolchain::llvm::package_name(), Family::Llvm },
     };
+    if (!linuxNonX86)
+        out.push_back({ mcpp::toolchain::llvm::package_name(), Family::Llvm });
     // The Windows-PE gcc payload is host-split at the distribution layer
     // (§4.3); each host lists the package it would actually install.
     if constexpr (mcpp::platform::is_windows) {
@@ -660,7 +688,8 @@ std::vector<AvailableIndex> available_toolchain_indexes() {
         out.push_back({ std::string(mcpp::platform::host_arch) + "-linux-musl-gcc",
                         Family::Gcc });
     } else if constexpr (mcpp::platform::is_linux) {
-        out.push_back({ "mingw-cross-gcc", Family::Gcc });
+        // Same gate: `mingw-cross-gcc` publishes x86_64 only.
+        if (!linuxNonX86) out.push_back({ "mingw-cross-gcc", Family::Gcc });
     }
     return out;
 }