docs(plan): 14.1 —— 缺陷在已发布生态里也复现,而不只是在夹具里

Author: speak-agent

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

@@ -1081,6 +1081,21 @@ and reading the byte count the program prints:
 
 The build succeeds and the artifact is the old one.
 
+**And against the published ecosystem, not only against a fixture.** The same
+edit, in an `xlings subos --sandbox`, on a project whose only reference to any
+of this is `plugins = { version = "0.3.0", features = ["rules-spirv"] }`, with
+the engine addressed by its store path so no development checkout can carry the
+result:
+
+```
+store path: .../xim-x-mcpp/2026.9.7.1/bin/mcpp
+before: bytes=1480
+after:  bytes=1480          and `Finished dev in 0.06s` -- nothing rebuilt at all
+```
+
+That is what a consumer of the published packages gets today, which is why the
+fixture measurement alone was not treated as sufficient.
+
 **This is not a discovery, and that is the more useful half of it.** Section 4.3
 stated the requirement -- "the payload must be declared as an action input
 explicitly" -- before any of this was written. What it added was five more