docs(plan): 13.2.1 —— 这一批的跨仓波次,以及为什么它和上一批同形

约束没变(plugins 的 CI 钉的是已发布的 mcpp),后果也没变。写进文档而不是留在
PR 里,是因为这个形状现在是两个仓库之间的性质,而不是某一次改动的偶然。

附带一条这一批才看清的推论:引擎那一半能自己被测(e2e 633 不用任何规则包),
plugins 那一半不能——它的下界是 PR 打开时还没发生的那个发布。所以第二波的
绿色 CI 是**最后**一份可得的证据而不是第一份,顶上来的是本地对着本地构建的引擎
跑的那次。

Author: speak-agent

.agents/docs/2026-09-07-module-first-heterogeneous-surface.md

@@ -950,6 +950,41 @@ reads it. An engine that does not set it leaves the header surface in place,
 which is what every consumer had before, so an older engine keeps its behaviour
 and a newer one moves a project to the module surface with nothing declared.
 
+#### 13.2.1 The same structure, for the batch that closed 14.1 and 14.2
+
+The constraint has not changed and neither has its consequence, so the second
+batch has the same two waves and the same reason. It is recorded here rather
+than left to the pull requests because the shape is now a property of the two
+repositories rather than an accident of one change.
+
+```
+wave 1   mcpp 2026.9.8.1: `mcpp::recompile_if_changed` (protocol 9),
+         e2e 633, three ninja-backend unit tests, docs/07 both languages,
+         and this document's alignment with what shipped
+                 |
+                 +--- release mcpp ---+
+                                      |
+wave 2                                +--> mcpp:plugins 0.3.1: a depfile on
+                                           all six rules, the `.incbin`
+                                           declaration in `surface::emit`,
+                                           three CI steps whose criterion is
+                                           the artifact's bytes, and
+                                           MCPP_VERSION raised to 2026.9.8.1
+                                      |
+wave 3                                +--> mcpp-index: 0.3.1 in the three
+                                           platform tables, `latest` REPLACED
+                                      |
+wave 4                                +--> sandbox verification against the
+                                           published packages
+```
+
+The engine half is testable on its own -- e2e 633 uses no rule package -- and
+the plugins half is not, because its floor is the release that has not happened
+when the pull request opens. Nothing in wave 2 can be checked by its own CI
+until wave 1 is published, so a green wave-2 CI is the LAST evidence available
+rather than the first, and the local run against a locally built engine is what
+stands in for it.
+
 ### 13.3 Evaluation criteria, by the dimensions this work is judged on
 
 | Dimension | What decides it | Where it is checked |