fix(bench): tell xmake where libc++ is, and drop to three repetitions

**1. Windows 的 `import std` 不是缺口,是没告诉 xmake 去哪找。** 日志把解法写在
脸上,而我先前把它当成平台缺口豁免了:

    warning: std and std.compat modules not found!
             maybe try to add --sdk=<PATH/TO/LLVM> or install libc++
    error: <mcpp> missing std dependency for module mcpp.build.provisions

载荷里**本来就带着** `share/libc++/v1/std.cppm`。报错点名的是被测工程的模块,
所以读起来像「这个工程坏了」,而不是「引擎不知道自己的标准库在哪」——
windows/clang 格子每个场景都这么红。SDK 根目录由解析后的驱动路径推出
(`…/xim-x-llvm/<ver>/bin/clang++` → `…/<ver>`),和工具链 pin 同源,不需要手工同步。

⚠️ **第一版补丁放在了控制流到不了的地方**:`--sdk` 加在 `own_description` 分支
**之后**,而那个分支提前 return —— 偏偏 Windows 上失败的 `bench/projects/mcpp`
正是 own_description。这是这个仓库记过的老形状(「修补放在控制流到不了的地方」),
已移到分支之前,两条路径都覆盖。

**2. 每个场景 5 轮改成 3 轮。** 真实工程上「增量」并不便宜 —— `edit-comment` 在
xlings 上要重建 45 个导入者,五轮就是五次近乎完整的重建,一个 windows/clang 格子
为此花掉半小时以上。多出来的样本买不到相称的精度:钉住的工作负载上跨轮离散度
低于 2%,而每张发布的表都取中位数。**没人愿意等的矩阵,就是没人会看的矩阵。**

Author: speak-agent

bench/src/engines/xmake.cppm

@@ -26,6 +26,21 @@ inline std::string payload_toolchain(std::string_view compiler) {
     return clang ? "mcpp-clang" : "mcpp-gcc";
 }
 
+// The LLVM root a payload driver lives under: `…/xim-x-llvm/<ver>/bin/clang++`
+// → `…/xim-x-llvm/<ver>`. Empty for anything that is not a payload clang.
+inline std::string payload_sdk_root(std::string_view compiler) {
+    if (compiler.find("/registry/data/xpkgs/") == std::string_view::npos &&
+        compiler.find("\\registry\\data\\xpkgs\\") == std::string_view::npos)
+        return {};
+    if (compiler.find("clang") == std::string_view::npos) return {};
+    const auto slash = compiler.find_last_of("/\\");
+    if (slash == std::string_view::npos) return {};
+    const auto bin = compiler.substr(0, slash);          // …/<ver>/bin
+    const auto up  = bin.find_last_of("/\\");
+    if (up == std::string_view::npos) return {};
+    return std::string(bin.substr(0, up));               // …/<ver>
+}
+
 class XmakeEngine : public Engine {
 public:
     std::string_view name() const override { return "xmake"; }
@@ -63,6 +78,24 @@ public:
             // it cannot be turned off without also discarding the toolchain.
             "--ccache=n",
         };
+        // ── Tell xmake where libc++ lives, or `import std;` has no provider ──
+        //
+        // The payload SHIPS the std module (share/libc++/v1/std.cppm), but xmake
+        // looks for it under its own notion of an LLVM SDK and otherwise says
+        //     warning: std and std.compat modules not found!
+        //              maybe try to add --sdk=<PATH/TO/LLVM> or install libc++
+        //     error: <mcpp> missing std dependency for module mcpp.build.provisions
+        // — a message naming a module of the project under test, so it reads as
+        // "this project is broken" rather than "the engine was not told where
+        // its standard library is". Every scenario in the windows/clang cell
+        // failed that way.
+        //
+        // Derived from the resolved driver (…/bin/clang++), which is the same
+        // path the toolchain pin already produced, so there is nothing to keep
+        // in step by hand.
+        if (const auto sdk = payload_sdk_root(job.compiler); !sdk.empty())
+            argv.push_back("--sdk=" + sdk);
+
         // ── How the payload driver is pinned, and why it is not always CXX ──
         //
         // A real project's description (bench/projects/*/xmake.lua) DEFINES the

bench/src/spec.cppm

@@ -88,8 +88,15 @@ struct Job {
 // this module stays free of any assumption about the project being measured.
 
 // Cold builds are expensive and their variance is low; incremental scenarios are
-// cheap and noisier, so they get more repetitions. Encoded here rather than in
-// the runner so the policy is visible next to the scenario it applies to.
-constexpr int default_runs(Scenario s) { return s == Scenario::Cold ? 3 : 5; }
+// cheap and noisier, so they would justify more repetitions. Encoded here rather
+// than in the runner so the policy is visible next to the scenario it applies to.
+//
+// THREE, not five. On a real project an "incremental" scenario is not cheap —
+// `edit-comment` on xlings rebuilds 45 importers, so five repetitions is five
+// near-full rebuilds and a windows/clang cell spent over half an hour on one
+// engine. The extra samples were not buying accuracy worth that: the spread
+// across runs on the pinned workloads is under 2%, and every published table is
+// a median. A matrix nobody waits for is one nobody reads.
+constexpr int default_runs(Scenario) { return 3; }
 
 }  // namespace bench