docs(bench): the schedule column — the default-only table understated mcpp badly

你问得对:上一版表是 **`bmi_schedule` 关着**测的,而那正是专门优化冷构建与级联的
开关。补测同一棵树、同样五个场景的 `+schedule=on` 对照臂,**10 ok / 0 failed**:

| 场景 | schedule=on | 默认 | 对 cmake |
|---|---|---|---|
| cold | **37.56s** | 92.49s | 1.29x → **3.18x** |
| touch-hub | **1.04s** | 1.79s | 54.96x → **93.93x** |
| edit-body | **29.65s** | 88.38s | 1.11x → **3.32x** |
| edit-comment | **30.44s** | 93.81s | 1.04x → **3.22x** |
| noop | 0.72s | 0.74s | 0.49x → 0.50x(仍然输) |

**我先前对 `edit-body`/`edit-comment` 的解释只对默认配置成立。** 我写的是「级联是
欠的,所以 mcpp 没有优势」—— 级联确实是欠的,但开了调度之后 mcpp 把**同一份欠下的
工作**做快了 3.3 倍(BMI 一产出就发布,而不是等代码生成结束)。只放默认列,读者会
得出「mcpp 在真实编辑场景上和 cmake 一样」的结论,那是错的。

两列并排,因为它们回答不同的问题:默认列是用户今天拿到的,调度列是这个开关能给
什么。`noop` 两列都输,照写。

**§8b 的正确性缺陷在这里没有复现**(十个格子全 ok)。它只在生成的 fixture 上出现
—— 那条紧密的 unit_0→unit_1 链会撞进窗口。该键因此仍然默认关闭:只有一个工作负载
能暴露的缺陷,仍然是缺陷。

Author: speak-agent

bench/README.md

@@ -200,13 +200,32 @@ First measurement in which all three arms produce a running binary — the cmake
 and xmake columns below were `failed` cells until the arms were finished, and the
 table that stood here was mcpp-against-mcpp for that reason.
 
-| scenario | **mcpp** | cmake | xmake |
-|---|---|---|---|
-| `cold` | **92.49s** · 1.29x | 119.46s · 1.00x | 105.02s · 1.14x |
-| `noop` | **0.74s** · 0.49x | 0.36s · 1.00x | 0.40s · 0.90x |
-| `touch-hub` | **1.79s** · 54.96x | 98.16s · 1.00x | 98.16s · 1.00x |
-| `edit-body` | **88.38s** · 1.11x | 98.43s · 1.00x | 98.00s · 1.00x |
-| `edit-comment` | **93.81s** · 1.04x | 97.98s · 1.00x | 97.97s · 1.00x |
+| scenario | **mcpp** `bmi_schedule=on` | mcpp default | cmake | xmake |
+|---|---|---|---|---|
+| `cold` | **37.56s** · 3.18x | 92.49s · 1.29x | 119.46s · 1.00x | 105.02s · 1.14x |
+| `noop` | **0.72s** · 0.50x | 0.74s · 0.49x | 0.36s · 1.00x | 0.40s · 0.90x |
+| `touch-hub` | **1.04s** · 93.93x | 1.79s · 54.96x | 98.16s · 1.00x | 98.16s · 1.00x |
+| `edit-body` | **29.65s** · 3.32x | 88.38s · 1.11x | 98.43s · 1.00x | 98.00s · 1.00x |
+| `edit-comment` | **30.44s** · 3.22x | 93.81s · 1.04x | 97.98s · 1.00x | 97.97s · 1.00x |
+
+**Both mcpp columns are here because one of them was misleading on its own.**
+The default column is what a user gets today; `bmi_schedule=on` is the opt-in
+split schedule, and leaving it out understated mcpp badly — `edit-body` reads
+1.11x in the default column and 3.32x with the schedule on.
+
+* **`edit-body` and `edit-comment` are not "no advantage".** The cascade really
+  is owed in both (the perturbed function body lives in an interface unit, so
+  the BMI genuinely changes). The default column shows mcpp doing that owed work
+  at cmake's pace; the schedule column shows it doing the SAME work 3.3x faster,
+  by publishing each BMI as soon as it exists instead of after code generation.
+* **`noop` is the one mcpp loses outright**, in both columns: 0.72–0.74s against
+  cmake's 0.36s. That is per-invocation overhead, and it is the number a user
+  feels on every edit-build cycle.
+* **The `bmi_schedule` correctness bug (§8b) does NOT reproduce here.** All ten
+  cells are `ok`. It reproduces on the generated fixture, whose tight
+  unit_0→unit_1 chain hits the window; three real trees (mcpp's own and both
+  xlings styles) do not. That is why the key is still opt-in — a defect that
+  only one workload can show is still a defect.
 
 <sub>xlings `2026.8.11.2`, gcc 16.1.0 payload, Linux x86_64 · i9-13900K · n=1 ·
 `--baseline cmake`. Raw report: `bench/results/xlings-3way-20260814/`.</sub>

bench/README.zh-CN.md

@@ -164,13 +164,28 @@ job 一个测量都没有，这个状态持续了好几周。
 **第一次三条臂都能产出可运行二进制的测量** —— 在此之前 cmake 与 xmake 两列一直是
 `failed`,所以这里原本只有 mcpp 对 mcpp。
 
-| 场景 | **mcpp** | cmake | xmake |
-|---|---|---|---|
-| `cold` | **92.49s** · 1.29x | 119.46s · 1.00x | 105.02s · 1.14x |
-| `noop` | **0.74s** · 0.49x | 0.36s · 1.00x | 0.40s · 0.90x |
-| `touch-hub` | **1.79s** · 54.96x | 98.16s · 1.00x | 98.16s · 1.00x |
-| `edit-body` | **88.38s** · 1.11x | 98.43s · 1.00x | 98.00s · 1.00x |
-| `edit-comment` | **93.81s** · 1.04x | 97.98s · 1.00x | 97.97s · 1.00x |
+| 场景 | **mcpp** `bmi_schedule=on` | mcpp 默认 | cmake | xmake |
+|---|---|---|---|---|
+| `cold` | **37.56s** · 3.18x | 92.49s · 1.29x | 119.46s · 1.00x | 105.02s · 1.14x |
+| `noop` | **0.72s** · 0.50x | 0.74s · 0.49x | 0.36s · 1.00x | 0.40s · 0.90x |
+| `touch-hub` | **1.04s** · 93.93x | 1.79s · 54.96x | 98.16s · 1.00x | 98.16s · 1.00x |
+| `edit-body` | **29.65s** · 3.32x | 88.38s · 1.11x | 98.43s · 1.00x | 98.00s · 1.00x |
+| `edit-comment` | **30.44s** · 3.22x | 93.81s · 1.04x | 97.98s · 1.00x | 97.97s · 1.00x |
+
+**两列 mcpp 都在这里,是因为只放一列会误导。** 默认列是用户今天拿到的东西;
+`bmi_schedule=on` 是可选的分离调度,不放它会严重低估 mcpp —— `edit-body` 在默认
+列是 1.11x,开了调度是 3.32x。
+
+* **`edit-body` 与 `edit-comment` 不是「没有优势」。** 两者的级联都是**欠的**
+  (被扰动的函数体在接口单元里,BMI 真的变了)。默认列显示 mcpp 以 cmake 的速度
+  做完这份欠下的工作;调度列显示它把**同一份工作**做快了 3.3 倍 —— 靠的是 BMI
+  一产出就发布,而不是等代码生成结束。
+* **`noop` 是 mcpp 真正输的一项**,两列都输:0.72–0.74s 对 cmake 的 0.36s。这是
+  每次调用的固定开销,也是用户在每一次「改一行、构建一次」里都感受得到的数字。
+* **§8b 那个 `bmi_schedule` 正确性缺陷在这里没有复现**,十个格子全 `ok`。它只在
+  生成的 fixture 上复现 —— 那条紧密的 `unit_0→unit_1` 链会撞进窗口;三棵真实的树
+  (mcpp 自己和 xlings 两种风格)都不会。这也正是该键仍然默认关闭的原因:**只有
+  一个工作负载能暴露的缺陷,仍然是缺陷。**
 
 <sub>xlings `2026.8.11.2`,gcc 16.1.0 载荷,Linux x86_64 · i9-13900K · n=1 ·
 `--baseline cmake`。原始报告:`bench/results/xlings-3way-20260814/`。</sub>

bench/results/xlings-3way-20260814/xlings-schedule.json

@@ -0,0 +1,166 @@
+{
+  "protocol_version": 1,
+  "started_at": "2026-08-13T22:29:14Z",
+  "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",
+      "compiler": "gcc",
+      "profile": "release",
+      "scenario": "cold",
+      "fixture": "xlings-2026.8.11.2",
+      "variant": "modules",
+      "status": "ok",
+      "note": "mcpp 2026.8.13.1",
+      "runs": 1,
+      "median_s": 91.608,
+      "min_s": 91.608,
+      "max_s": 91.608,
+      "samples": [91.608]
+    },
+    {
+      "engine": "mcpp@2026.8.13.1",
+      "compiler": "gcc",
+      "profile": "release",
+      "scenario": "noop",
+      "fixture": "xlings-2026.8.11.2",
+      "variant": "modules",
+      "status": "ok",
+      "note": "mcpp 2026.8.13.1",
+      "runs": 1,
+      "median_s": 0.723,
+      "min_s": 0.723,
+      "max_s": 0.723,
+      "samples": [0.723]
+    },
+    {
+      "engine": "mcpp@2026.8.13.1",
+      "compiler": "gcc",
+      "profile": "release",
+      "scenario": "touch-hub",
+      "fixture": "xlings-2026.8.11.2",
+      "variant": "modules",
+      "status": "ok",
+      "note": "mcpp 2026.8.13.1",
+      "runs": 1,
+      "median_s": 1.786,
+      "min_s": 1.786,
+      "max_s": 1.786,
+      "samples": [1.786]
+    },
+    {
+      "engine": "mcpp@2026.8.13.1",
+      "compiler": "gcc",
+      "profile": "release",
+      "scenario": "edit-body",
+      "fixture": "xlings-2026.8.11.2",
+      "variant": "modules",
+      "status": "ok",
+      "note": "mcpp 2026.8.13.1 · perturbation: in-body",
+      "runs": 1,
+      "median_s": 89.457,
+      "min_s": 89.457,
+      "max_s": 89.457,
+      "samples": [89.457]
+    },
+    {
+      "engine": "mcpp@2026.8.13.1",
+      "compiler": "gcc",
+      "profile": "release",
+      "scenario": "edit-comment",
+      "fixture": "xlings-2026.8.11.2",
+      "variant": "modules",
+      "status": "ok",
+      "note": "mcpp 2026.8.13.1 · perturbation: in-body",
+      "runs": 1,
+      "median_s": 93.843,
+      "min_s": 93.843,
+      "max_s": 93.843,
+      "samples": [93.843]
+    },
+    {
+      "engine": "mcpp@2026.8.13.1+schedule=on",
+      "compiler": "gcc",
+      "profile": "release",
+      "scenario": "cold",
+      "fixture": "xlings-2026.8.11.2",
+      "variant": "modules",
+      "status": "ok",
+      "note": "mcpp 2026.8.13.1",
+      "runs": 1,
+      "median_s": 37.561,
+      "min_s": 37.561,
+      "max_s": 37.561,
+      "samples": [37.561]
+    },
+    {
+      "engine": "mcpp@2026.8.13.1+schedule=on",
+      "compiler": "gcc",
+      "profile": "release",
+      "scenario": "noop",
+      "fixture": "xlings-2026.8.11.2",
+      "variant": "modules",
+      "status": "ok",
+      "note": "mcpp 2026.8.13.1",
+      "runs": 1,
+      "median_s": 0.723,
+      "min_s": 0.723,
+      "max_s": 0.723,
+      "samples": [0.723]
+    },
+    {
+      "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": 1,
+      "median_s": 1.045,
+      "min_s": 1.045,
+      "max_s": 1.045,
+      "samples": [1.045]
+    },
+    {
+      "engine": "mcpp@2026.8.13.1+schedule=on",
+      "compiler": "gcc",
+      "profile": "release",
+      "scenario": "edit-body",
+      "fixture": "xlings-2026.8.11.2",
+      "variant": "modules",
+      "status": "ok",
+      "note": "mcpp 2026.8.13.1 · perturbation: in-body",
+      "runs": 1,
+      "median_s": 29.646,
+      "min_s": 29.646,
+      "max_s": 29.646,
+      "samples": [29.646]
+    },
+    {
+      "engine": "mcpp@2026.8.13.1+schedule=on",
+      "compiler": "gcc",
+      "profile": "release",
+      "scenario": "edit-comment",
+      "fixture": "xlings-2026.8.11.2",
+      "variant": "modules",
+      "status": "ok",
+      "note": "mcpp 2026.8.13.1 · perturbation: in-body",
+      "runs": 1,
+      "median_s": 30.444,
+      "min_s": 30.444,
+      "max_s": 30.444,
+      "samples": [30.444]
+    }
+  ]
+}