Revert "build: turn bmi_schedule ON for mcpp's own build" — CI found a real defect

一个 CI 周期就把它否掉了。`cross-build-test` 的
**windows 宿主 → x86_64-linux-musl** 那个 job 红在 BMI 边:

    failed: gcm.cache/mcpp.libs.json.gcm
    src/libs/json.cppm:3: fatal error: json.hpp: No such file or directory

即**分离出来的 BMI 边在那个宿主上拿不到 `include_dirs`**。已记为 #425,
含定位到规则的分析(detach 的 BMI 边把整条命令写进 `$out.cmd`,而 windows 上
普通编译边是把 `$local_includes` 放进另一个 rspfile 再 `@` 引用 —— 两条路径对
它的处理本来就不一样)。

⚠️ 这条恰好证明了那条规则的价值,也证明了我上一个提交的证据是不够的:

    linux 自举 + 83 个单测          绿
    macOS / windows 原生 e2e        绿(clang ⇒ two-phase)
    windows 宿主交叉到 linux-musl   红

我在提交信息里写了「本地已验证」,而本地验证覆盖不到这个组合。
「opt-in 直到在**所有平台**过完 CI」说的就是这件事,不是「直到它在我这儿能跑」。

对所有人的默认值 `auto`(= 关)本来就不受影响。mcpp.toml 里那一行换成了
说明为什么不开的注释,修好 #425 之后再开。

Author: speak-agent

mcpp.toml

@@ -15,27 +15,17 @@ default-profile = "release"
 # nlohmann/json.hpp lives in src/libs/json/; expose it to the global
 # module fragment `#include <json.hpp>` in src/libs/json.cppm.
 include_dirs = ["src/libs/json"]
-# Split the module edge: importers start when the BMI is published rather than
-# when the compiler exits. See docs/05-mcpp-toml.md.
+# NOT enabled here. `bmi_schedule = "on"` was turned on for mcpp's own build and
+# CI rejected it within one cycle: the windows-host cross-build to
+# x86_64-linux-musl failed on the BMI edge with
 #
-# ON HERE, `auto` (= off) EVERYWHERE ELSE — deliberately, and this is the point
-# of turning it on at all. The key stays opt-in until it has been through CI on
-# every platform, and nothing was exercising it, so it could never accumulate
-# the evidence needed to change that. mcpp's own build is where that evidence
-# is cheapest to get: whatever this schedule produces then has to pass 83 unit
-# tests and the full e2e suite, on Linux (detach-codegen, gcc) and on macOS and
-# Windows (two-phase, clang) — three platforms, every PR.
+#     failed: gcm.cache/mcpp.libs.json.gcm
+#     src/libs/json.cppm:3: fatal error: json.hpp: No such file or directory
 #
-# ⚠️ WHY THAT SAFETY NET MATTERS HERE. A scheduling bug does not fail loudly; it
-# stops rebuilding something. The one that already happened (§8b in
-# bench/README.md) left the object edge with the BMI as its only input, so
-# restat cleaned the object and ninja exited while the compiler was still
-# running — a binary that was silently short. Running the produced binary
-# against its own test suite is what turns that class of defect into a red job
-# instead of a bad release.
-#
-# Revert is this one line if anything here is ever unexplained.
-bmi_schedule = "on"
+# i.e. the split BMI edge loses `include_dirs` on that host. Linux was green and
+# the local self-host build was green (35.40s vs ~80s, 83 unit tests passing),
+# which is exactly why the rule is "opt-in until it has been through CI on EVERY
+# platform" rather than "until it works here". See issue #425.
 
 [toolchain]
 default = "gcc@16.1.0"