feat(build): L3 build.mcpp — native imperative build program (v0.0.78)

A project-local build.mcpp (C++, Zig build.zig / Cargo build.rs model) compiled
with the host toolchain and run before the main build. It emits stdout mcpp:
directives that augment the build; a declared-input cache re-runs it only when
its source/inputs/env change (the documented fix for the .mcpp_ok blind spot).

- src/build/build_program.cppm: new module — compile+run, directive parser
  (cxxflag/cflag/link-lib/link-search/cfg/generated/rerun-if-changed/
  rerun-if-env-changed), declared-input cache, apply to buildConfig
- prepare.cppm: invoke after toolchain detect + L1 cfg merge, before modgraph
  scan; skipped under cross --target (host build/run). -x c++ so the .mcpp
  extension compiles as C++
- tests/e2e/89_build_mcpp.sh: define+generated source reach the build; cache
  short-circuits unchanged re-run; changed env forces re-run
- docs/07-build-mcpp.md (+zh): user guide; design doc + version bump 0.0.78

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Author: Sunrisepeak

.agents/docs/2026-06-30-l3-build-mcpp-implementation-design.md

@@ -0,0 +1,115 @@
+# L3 `build.mcpp` — native imperative build program (implementation design)
+
+Companion to `2026-06-29-manifest-environment-and-platform-design.md` (§L3). This
+doc nails down the concrete MVP shipped in mcpp 0.0.78.
+
+## What it is
+
+A project-local `build.mcpp` (a C++ source file, Zig's `build.zig` / Cargo's
+`build.rs` model — but in the project's own language, so no second language and it
+dogfoods mcpp). mcpp compiles it with the **host** toolchain and runs it **before**
+the main build; the program emits stdout directives that augment the main build.
+
+```cpp
+// build.mcpp
+#include <cstdio>
+int main() {
+    std::puts("mcpp:cxxflag=-DHAVE_FEATURE=1");
+    std::puts("mcpp:link-lib=m");
+    std::puts("mcpp:rerun-if-env-changed=USE_FAST");
+}
+```
+
+## Directive protocol (Discipline 1 — structured output, not global mutation)
+
+The program communicates **only** via stdout lines; everything else is ignored
+(so the program may freely log to stderr/stdout). Recognized directives:
+
+| Directive | Effect |
+|---|---|
+| `mcpp:cxxflag=<flag>`         | append `<flag>` to `buildConfig.cxxflags` |
+| `mcpp:cflag=<flag>`           | append `<flag>` to `buildConfig.cflags` |
+| `mcpp:link-lib=<name>`        | append `-l<name>` to `buildConfig.ldflags` |
+| `mcpp:link-search=<dir>`      | append `-L<abs dir>` to `buildConfig.ldflags` (dir resolved against the project root) |
+| `mcpp:cfg=<name>`             | append `-D<name>` to **both** cflags and cxxflags |
+| `mcpp:generated=<path>`       | add `<path>` (relative to project root) to `buildConfig.sources` so the modgraph scanner picks it up |
+| `mcpp:rerun-if-changed=<path>`| declare a file input (re-run gate, see Discipline 2) |
+| `mcpp:rerun-if-env-changed=<VAR>` | declare an env input (re-run gate) |
+
+It *requests* graph edges (flags/libs/sources); it never silently mutates build state.
+Unknown `mcpp:` directives are ignored with a one-line warning (forward-compat).
+
+## Declared-I/O re-run contract (Discipline 2 — fixes the `.mcpp_ok` blind spot)
+
+The program is **not** re-run every build. Its parsed directives + declared inputs
+are cached at `<proj>/.mcpp/build.mcpp.cache`. On each build we re-run iff:
+
+- the cache is missing, **or**
+- the `build.mcpp` source content hash changed, **or**
+- the host compiler identity changed, **or**
+- any declared `rerun-if-changed` file's content hash changed (or the file vanished), **or**
+- any declared `rerun-if-env-changed` variable's current value changed, **or**
+- any `generated=` output path no longer exists.
+
+Otherwise the cached directives are reused without recompiling/running. This is the
+documented replacement for the bare `.mcpp_ok` success marker ("process exited 0 ≠
+outputs correct"): a **declared-input / declared-output contract**. Hashing reuses
+the existing FNV-1a helpers (`mcpp::toolchain::hash_file` / `hash_string`).
+
+Because the applied directives land in `buildConfig.{cflags,cxxflags,ldflags}` —
+which already feed `canonical_compile_flags` → the fingerprint — and generated
+sources feed the modgraph, the **main** build is automatically sensitive to a
+changed `build.mcpp` output. The cache only avoids needless re-execution / file
+regeneration (which would otherwise bump mtimes and force spurious rebuilds).
+
+## Constraints (à la carte + supply-chain)
+
+- **Leaf only.** `build.mcpp` chooses flags/sources/codegen and emits link
+  requirements; it must **not** gate the top-level dependency graph (that stays in
+  the applicative L1 `[target.'cfg(...)']` tables). The directive set deliberately
+  excludes "add a registry dependency".
+- **Host build, target cfg.** It compiles+runs on the **host**. The MVP therefore
+  runs it only for **native** builds; under an explicit cross `--target` it is
+  **skipped with a warning** (compiling it with the cross frontend would yield a
+  binary that can't run on the host). Host-toolchain-for-cross is a follow-up.
+- **Isolation.** Executed as a build action: child-only env (no calling-process
+  mutation, via `capture_exec`), declared inputs/outputs. Extending the same
+  declared-I/O contract to recipe `install()` is future work.
+
+## Integration (src/build/prepare.cppm)
+
+New module `src/build/build_program.cppm` exports
+`run_build_program(Manifest&, root, hostCompiler, cppStandard)`. Called from
+`prepare.cppm` right after toolchain detection (`tc`), i.e. **after** target
+resolution + the L1 cfg-flag merge (buildConfig flags final) and **before** the
+modgraph scanner (so `generated=` sources are scanned). Compile line:
+
+```
+<hostCompiler> -std=<cppStandard> -O0 -o <proj>/.mcpp/build.mcpp.bin <proj>/build.mcpp
+```
+
+Compile/run failures are hard errors surfaced with captured output.
+
+## Tests
+
+- `tests/e2e/89_build_mcpp.sh` — a `build.mcpp` emitting a `cxxflag` define + a
+  `generated` source; assert the define reaches the TU (a `#ifdef` gate) and the
+  generated source links. Second build asserts the cache short-circuits re-run;
+  touching a declared `rerun-if-changed` input forces re-run.
+
+## Forward note — `.mcpp` as a first-class C++ extension
+
+The compiler doesn't know the `.mcpp` extension, so we compile build.mcpp with an
+explicit `-x c++` (otherwise the driver hands it to the linker as a "linker
+script"). This is a special case of a broader convention worth adopting: **inside
+an mcpp project, `.mcpp` is just C++.** A natural next step is to add `.mcpp` to the
+main build's source glob (`src/**/*.{cppm,cpp,cc,c}` → `+ .mcpp`) with the same
+`-x c++` treatment, so a project may use `.mcpp` for ordinary sources/modules — the
+extension becomes a marker of "an mcpp-native C++ file" rather than a separate
+language. `build.mcpp` is the first instance; the `-x c++` handling here is the
+seed. Deferred (out of MVP scope) but the direction is intentional.
+
+## mcpp-index dual perspective
+
+A new workspace member `tests/examples/build-mcpp` whose `build.mcpp` emits a
+define consumed by `main.cpp`, exercising the feature through the real pipeline.

docs/07-build-mcpp.md

@@ -0,0 +1,88 @@
+# `build.mcpp` — a native build program
+
+**English** | [简体中文](zh/07-build-mcpp.md)
+
+Most projects need nothing more than `mcpp.toml`. When you need build-time logic —
+probe the host, generate a source, decide a flag from the environment — put a
+`build.mcpp` in your project root. It is the mcpp analog of Zig's `build.zig` and
+Cargo's `build.rs`, but written in **C++**: no second language, and it dogfoods
+mcpp itself.
+
+mcpp compiles `build.mcpp` with your toolchain and runs it **before** the main
+build. The program talks to mcpp by printing `mcpp:` directives to stdout; those
+directives augment the build.
+
+## Quick example
+
+```cpp
+// build.mcpp
+#include <cstdio>
+#include <fstream>
+
+int main() {
+    // Generate a source the main build will compile + link.
+    std::ofstream("src/generated.cpp") << "const char* banner() { return \"hi\"; }\n";
+
+    std::puts("mcpp:generated=src/generated.cpp");   // add it to the build
+    std::puts("mcpp:cxxflag=-DHAVE_BANNER=1");        // define a macro for all C++ TUs
+
+    if (std::getenv("USE_FAST")) std::puts("mcpp:cxxflag=-DFAST_PATH=1");
+    std::puts("mcpp:rerun-if-env-changed=USE_FAST");  // re-run me when USE_FAST changes
+    return 0;
+}
+```
+
+```bash
+mcpp build      # compiles + runs build.mcpp, then builds the project
+```
+
+## Directives
+
+Print these to stdout (one per line). Any line that does not start with `mcpp:`
+is ignored, so you can freely log diagnostics.
+
+| Directive | Effect |
+|---|---|
+| `mcpp:cxxflag=<flag>`              | add `<flag>` to the C++ compile flags |
+| `mcpp:cflag=<flag>`                | add `<flag>` to the C compile flags |
+| `mcpp:link-lib=<name>`             | link `-l<name>` |
+| `mcpp:link-search=<dir>`           | add a library search dir (`-L`; relative dirs resolve against the project root) |
+| `mcpp:cfg=<name>`                  | define `-D<name>` for both C and C++ |
+| `mcpp:generated=<path>`            | add a generated source (relative to the project root) to the build |
+| `mcpp:rerun-if-changed=<path>`     | re-run `build.mcpp` when this file changes |
+| `mcpp:rerun-if-env-changed=<VAR>`  | re-run `build.mcpp` when this env var changes |
+
+The program **requests** build edges (flags, libraries, sources). It cannot add a
+registry dependency — keep your dependency graph declarative in `mcpp.toml`
+(including platform-conditional `[target.'cfg(...)'.dependencies]`). `build.mcpp`
+is for *leaf* decisions: flags, codegen, link requirements.
+
+## Incremental: declared inputs (no needless re-runs)
+
+mcpp does **not** re-run `build.mcpp` on every build. It caches the program's
+directives and re-runs only when something it depends on changed:
+
+- the `build.mcpp` source itself,
+- the toolchain,
+- any file you declared with `rerun-if-changed`,
+- any env var you declared with `rerun-if-env-changed`,
+- (or a `generated` output went missing).
+
+So **declare your inputs**: if your program reads `config.h` or the `USE_FAST`
+variable, emit `mcpp:rerun-if-changed=config.h` / `mcpp:rerun-if-env-changed=USE_FAST`.
+This replaces the old "process exited 0, so assume it's fine" guesswork with an
+explicit input/output contract — incremental builds stay correct.
+
+When nothing changed you'll see `build.mcpp up to date (cached)`; otherwise
+`build.mcpp compiling` / `running`.
+
+## Notes & limits
+
+- **Runs on the host.** `build.mcpp` compiles and runs with the host toolchain.
+  Under a cross build (`mcpp build --target <triple>`) it is **skipped with a
+  warning** for now (host-toolchain-for-cross is a planned follow-up). Gate
+  *dependencies* on the target with `[target.'cfg(...)']` tables instead — those
+  evaluate on the resolved target. See [05 - mcpp.toml Manifest Guide](05-mcpp-toml.md).
+- **CWD is the project root**, so relative paths (`src/generated.cpp`) land where
+  you expect.
+- A non-zero exit from `build.mcpp` aborts the build and prints its output.

docs/README.md

@@ -9,3 +9,4 @@
 - [04 - Building from Source & Contributing](04-build-from-source.md)
 - [05 - mcpp.toml Manifest Guide](05-mcpp-toml.md)
 - [06 - Workspaces](06-workspace.md)
+- [07 - build.mcpp Build Program](07-build-mcpp.md)

docs/zh/07-build-mcpp.md

@@ -0,0 +1,81 @@
+# `build.mcpp` —— 原生构建程序
+
+[English](../07-build-mcpp.md) | **简体中文**
+
+绝大多数工程只需要 `mcpp.toml`。当你需要构建期逻辑——探测主机、生成源码、依据环境
+决定某个编译开关——就在工程根目录放一个 `build.mcpp`。它是 mcpp 版的 Zig `build.zig`
+/ Cargo `build.rs`,但用 **C++** 编写:不引入第二种语言,而且 mcpp 自己吃自己的狗粮。
+
+mcpp 用你的工具链编译 `build.mcpp`,并在主构建**之前**运行它。程序通过向 stdout 打印
+`mcpp:` 指令与 mcpp 通信,这些指令会增补本次构建。
+
+## 快速示例
+
+```cpp
+// build.mcpp
+#include <cstdio>
+#include <fstream>
+
+int main() {
+    // 生成一份源码,主构建会编译 + 链接它。
+    std::ofstream("src/generated.cpp") << "const char* banner() { return \"hi\"; }\n";
+
+    std::puts("mcpp:generated=src/generated.cpp");   // 加入构建
+    std::puts("mcpp:cxxflag=-DHAVE_BANNER=1");        // 为所有 C++ TU 定义宏
+
+    if (std::getenv("USE_FAST")) std::puts("mcpp:cxxflag=-DFAST_PATH=1");
+    std::puts("mcpp:rerun-if-env-changed=USE_FAST");  // USE_FAST 变化时重跑我
+    return 0;
+}
+```
+
+```bash
+mcpp build      # 编译 + 运行 build.mcpp,然后构建工程
+```
+
+## 指令
+
+把这些打印到 stdout(每行一条)。任何不以 `mcpp:` 开头的行都会被忽略,因此你可以
+自由打印诊断日志。
+
+| 指令 | 作用 |
+|---|---|
+| `mcpp:cxxflag=<flag>`              | 给 C++ 编译追加 `<flag>` |
+| `mcpp:cflag=<flag>`                | 给 C 编译追加 `<flag>` |
+| `mcpp:link-lib=<name>`             | 链接 `-l<name>` |
+| `mcpp:link-search=<dir>`           | 增加库搜索目录(`-L`;相对路径按工程根目录解析) |
+| `mcpp:cfg=<name>`                  | 为 C 与 C++ 同时定义 `-D<name>` |
+| `mcpp:generated=<path>`            | 把生成的源码(相对工程根目录)加入构建 |
+| `mcpp:rerun-if-changed=<path>`     | 该文件变化时重跑 `build.mcpp` |
+| `mcpp:rerun-if-env-changed=<VAR>`  | 该环境变量变化时重跑 `build.mcpp` |
+
+程序**请求**构建边(开关、库、源码),它**不能**新增注册表依赖——请把依赖图保持在
+`mcpp.toml` 里声明式管理(包括平台条件依赖 `[target.'cfg(...)'.dependencies]`)。
+`build.mcpp` 用于*叶子*决策:开关、代码生成、链接需求。
+
+## 增量:声明输入(避免无谓重跑)
+
+mcpp **不会**每次构建都重跑 `build.mcpp`。它会缓存程序产出的指令,只有当它依赖的东西
+变化时才重跑:
+
+- `build.mcpp` 源码本身,
+- 工具链,
+- 任何用 `rerun-if-changed` 声明的文件,
+- 任何用 `rerun-if-env-changed` 声明的环境变量,
+- (或某个 `generated` 产物丢失了)。
+
+所以请**声明你的输入**:如果程序读了 `config.h` 或 `USE_FAST` 变量,就分别 emit
+`mcpp:rerun-if-changed=config.h` / `mcpp:rerun-if-env-changed=USE_FAST`。这用一份明确的
+输入/输出契约取代了过去「进程退出码为 0 就当成功」的猜测——让增量构建保持正确。
+
+无变化时你会看到 `build.mcpp up to date (cached)`;否则是 `build.mcpp compiling` /
+`running`。
+
+## 说明与限制
+
+- **在主机上运行。** `build.mcpp` 用主机工具链编译并运行。在交叉构建
+  (`mcpp build --target <triple>`)下目前会**跳过并给出警告**(主机工具链交叉是计划中的
+  后续项)。要按目标平台门控*依赖*,请改用 `[target.'cfg(...)']` 表——它们按解析后的目标
+  求值。参见 [05 - mcpp.toml 工程文件指南](05-mcpp-toml.md)。
+- **当前工作目录是工程根目录**,因此相对路径(`src/generated.cpp`)会落在你预期的位置。
+- `build.mcpp` 非零退出会中止构建并打印其输出。

docs/zh/README.md

@@ -9,3 +9,4 @@
 - [04 - 从源码构建 & 参与贡献](04-build-from-source.md)
 - [05 - mcpp.toml 工程文件指南](05-mcpp-toml.md)
 - [06 - 工作空间](06-workspace.md)
+- [07 - build.mcpp 构建程序](07-build-mcpp.md)

mcpp.toml

@@ -1,6 +1,6 @@
 [package]
 name        = "mcpp"
-version     = "0.0.77"
+version     = "0.0.78"
 description = "Modern C++ build & package management tool"
 license     = "Apache-2.0"
 authors     = ["mcpp-community"]