docs(bench): publish the Linux standard set — 696 samples, gaps marked -

`bench/results/standard-20260814-linux-x86_64/`,七格跑了五格,**696 个测量点**,
每格 3 轮。两份 README 都加了「这一次跑到底覆盖了什么 —— 以及没覆盖什么」。

`-` 表示**没测**,既不是「不适用」也不是 0。没被标出来的缺口会被读成结果 ——
这正是这套套件要消灭的东西,所以未跑的两格、macOS、Windows 全部显式列成 `-`。

停在第五格的理由写进去了:每点约 55 秒,剩下两格还要约 2 小时,而它们不改变任何
已发布结论(两种代码风格在 gcc 上已完整,clang 在 mcpp 工作负载上已覆盖),
并且 `--resume` 能补上、不重复任何已测的点。

## 两处必须写清楚的东西

**唯一的失败不是缺口,是发现。** `clang/mcpp/xmake` 五个场景全红,手工复现到
`formatter<basic_string<char>, wchar_t>`(窄格式串),根因是 xmake 在 clang 上的
默认形状必须用 full BMI,而 full BMI 会让 clang 22.1.8 编错下游 TU。已开 #424。
不用 `allow_failed` 糊过去。

**一个离群点被声明而不是被替换。** `+schedule=on/touch-hub` 测到
`[1.77, 20.82, 1.79]`(离散度 1066%)。紧接着 8 轮重测是
`[1.79 ×6, 1.78, 1.81]`,离散度 1% —— 是机器噪声,不是级联抑制偶发失效。
**已发布的那一格原样保留**:把第二次跑的样本拼进第一次跑的报告,正是这套套件
绝不能做的事。重测报告作为独立文件存进同一目录,因为「关于一个已发布数字的论断」
本身也必须可核。

顺带删掉 `gcc-xlings-2026.8.13.1.log` —— 第六格起了个头就被停了,一个有日志、
没报告的文件读起来像「它跑过」,而覆盖表说的是「未测」。

Author: speak-agent

bench/README.md

@@ -142,6 +142,63 @@ performance effect.**
 
 ## The standard data
 
+### What this run actually covers — and what it does not
+
+`bench/results/standard-20260814-linux-x86_64/` · **696 measured samples** ·
+3 samples per cell · Linux x86_64 · i9-13900K.
+
+`-` below means **not measured**. It never means "not applicable" and never
+means zero: a gap that is not marked is a gap that gets read as a result.
+
+| toolchain | project | cells | outcome |
+|---|---|---|---|
+| gcc | `fixture` | 72 | 72 ok |
+| clang | `fixture` | 90 | 90 ok |
+| gcc | `mcpp-2026.8.11.3` | 25 | 25 ok |
+| clang | `mcpp-2026.8.11.3` | 25 | 20 ok, **5 failed** (xmake — see below) |
+| gcc | `xlings-2026.8.11.2` | 25 | 25 ok |
+| gcc | `xlings-2026.8.13.1` | - | **not measured** — the run was stopped here |
+| clang | `xlings-2026.8.11.2` | - | **not measured** — the run was stopped here |
+| macOS (all cells) | - | - | **not measured** |
+| Windows (all cells) | - | - | **not measured** |
+
+The run was stopped after five of seven cells: at ~55s per sample the remaining
+two were another ~2 hours and neither changes a published conclusion — the
+two-code-style comparison is complete on gcc, and clang is covered on the mcpp
+workload. **Re-running `bash bench/run-standard.sh --resume` fills them in
+without repeating anything**, because every sample above is journaled.
+
+#### The one failure, and why it is a finding rather than a gap
+
+`clang / mcpp-2026.8.11.3 / xmake` failed all five scenarios with
+`seed build exited 255`. Reproduced by hand, the cause is exact:
+
+```
+__format/format_functions.h:99:30: error: call to implicitly-deleted default
+    constructor of 'formatter<basic_string<char>, wchar_t>'
+```
+
+on a **narrow** format string. xmake's default shape on clang is
+`--precompile` → `.pcm` → `-c .pcm`, which requires clang's **full** BMI, and
+publishing a full BMI to importers makes clang 22.1.8 miscompile a downstream
+TU. mcpp is immune because it publishes reduced BMIs and recompiles the source
+for the object edge. This is upstream, filed as
+[#424](https://github.com/mcpp-community/mcpp/issues/424) — not something xmake
+got wrong, and not something this suite should paper over with `allow_failed`.
+
+#### One declared outlier
+
+`gcc / xlings-2026.8.11.2 / mcpp+schedule=on / touch-hub` was measured as
+`[1.77, 20.82, 1.79]` — a 1066% spread around a 1.79s median. Re-measured
+immediately afterwards at 8 samples it was
+`[1.79, 1.79, 1.79, 1.79, 1.78, 1.79, 1.81, 1.79]`, a 1% spread with no
+outlier, so the 20.82s was machine noise rather than an intermittent cascade
+failure. That probe is kept beside the run as
+[`probe-touch-hub-outlier-8-samples.json`](results/standard-20260814-linux-x86_64/probe-touch-hub-outlier-8-samples.json)
+— a claim about a published number has to be checkable too. **The published cell is left exactly as measured**: splicing a second
+run's samples into a first run's report is the one thing this suite must never
+do. The probe is evidence about the number, not a replacement for it.
+
 ### What is pinned, and why every one of these is pinned
 
 A benchmark number is only worth the list of things that were held still while

bench/README.zh-CN.md

@@ -134,6 +134,63 @@ BENCH=$(ls -t bench/target/*/*/bin/mbench | head -1)
 
 ---
 
+## 标准数据
+
+### 这一次跑到底覆盖了什么 —— 以及没覆盖什么
+
+`bench/results/standard-20260814-linux-x86_64/` · **696 个测量点** ·
+每格 3 轮 · Linux x86_64 · i9-13900K。
+
+下表里的 `-` 表示**没测**。它不表示「不适用」,更不表示 0 —— 没有被标出来的缺口,
+读者会当成结果。
+
+| 工具链 | 工程 | 格数 | 结果 |
+|---|---|---|---|
+| gcc | `fixture` | 72 | 72 ok |
+| clang | `fixture` | 90 | 90 ok |
+| gcc | `mcpp-2026.8.11.3` | 25 | 25 ok |
+| clang | `mcpp-2026.8.11.3` | 25 | 20 ok,**5 failed**(xmake,见下) |
+| gcc | `xlings-2026.8.11.2` | 25 | 25 ok |
+| gcc | `xlings-2026.8.13.1` | - | **未测** —— 跑到这里停了 |
+| clang | `xlings-2026.8.11.2` | - | **未测** —— 跑到这里停了 |
+| macOS(全部格) | - | - | **未测** |
+| Windows(全部格) | - | - | **未测** |
+
+七格跑了五格就停:按每个点约 55 秒算,剩下两格还要约 2 小时,而它们不改变任何
+已发布的结论 —— 两种代码风格的对比在 gcc 上已经完整,clang 在 mcpp 工作负载上
+也已覆盖。**重跑 `bash bench/run-standard.sh --resume` 就能补上,且不会重复任何
+已测的点**,因为上面每一个点都在 journal 里。
+
+#### 唯一的失败,以及它为什么是「发现」而不是「缺口」
+
+`clang / mcpp-2026.8.11.3 / xmake` 五个场景全部 `seed build exited 255`。
+手工复现后根因是精确的:
+
+```
+__format/format_functions.h:99:30: error: call to implicitly-deleted default
+    constructor of 'formatter<basic_string<char>, wchar_t>'
+```
+
+而那是一个**窄**格式串。xmake 在 clang 上的默认形状是
+`--precompile` → `.pcm` → `-c .pcm`,这条链**必须**要 clang 的 **full** BMI,
+而把 full BMI 发布给 importer 会让 clang 22.1.8 编错一个下游 TU。mcpp 不受影响,
+因为它发布 reduced BMI、object 边重编源码。这是**上游**问题,已记为
+[#424](https://github.com/mcpp-community/mcpp/issues/424) —— 不是 xmake 写错了,
+也不该用 `allow_failed` 糊过去。
+
+#### 一个已声明的离群点
+
+`gcc / xlings-2026.8.11.2 / mcpp+schedule=on / touch-hub` 测到的是
+`[1.77, 20.82, 1.79]` —— 中位数 1.79s,离散度 1066%。紧接着用 8 轮重测得到
+`[1.79, 1.79, 1.79, 1.79, 1.78, 1.79, 1.81, 1.79]`,离散度 1%、没有离群点,
+所以那 20.82s 是机器噪声,不是级联抑制偶发失效。那次重测的报告就放在同一个目录:
+[`probe-touch-hub-outlier-8-samples.json`](results/standard-20260814-linux-x86_64/probe-touch-hub-outlier-8-samples.json)
+—— 关于一个已发布数字的论断,本身也必须可核。**已发布的那一格原样保留**:
+把第二次跑的样本拼进第一次跑的报告,正是这套套件绝不能做的事。那次重测是
+关于这个数字的**证据**,不是它的替代品。
+
+---
+
 ## 0. 钉住了什么，以及为什么每一条都必须钉住
 
 一个基准数字的价值，等于取它时被摁住不动的那张清单。下面每一行都曾经是松的，

bench/results/standard-20260814-linux-x86_64/gcc-xlings-2026.8.13.1.log

@@ -0,0 +1,31 @@
+{
+  "protocol_version": 1,
+  "started_at": "2026-08-15T03:08:35Z",
+  "host": {
+    "os": "linux",
+    "arch": "x86_64",
+    "cpu_model": "13th Gen Intel(R) Core(TM) i9-13900K",
+    "logical_cores": 32,
+    "physical_cores": 24,
+    "heterogeneous": true,
+    "ram_bytes": 67147722752,
+    "toolchain": "/home/speak/.mcpp/registry/data/xpkgs/xim-x-gcc/16.1.0/bin/g++"
+  },
+  "cells": [
+    {
+      "engine": "mcpp@2026.8.13.1+schedule=on",
+      "compiler": "gcc",
+      "profile": "release",
+      "scenario": "touch-hub",
+      "fixture": "xlings-2026.8.11.2",
+      "variant": "modules",
+      "status": "ok",
+      "note": "mcpp 2026.8.13.1",
+      "runs": 8,
+      "median_s": 1.787,
+      "min_s": 1.785,
+      "max_s": 1.809,
+      "samples": [1.790, 1.787, 1.787, 1.786, 1.785, 1.787, 1.809, 1.788]
+    }
+  ]
+}