tests: a check that ran is recognised by either wrapper spelling; the action PATH's one re-run is stated

On the MSVC ABI every action now runs through the named __action wrapper,
which carries --path-prepend. e2e 780 recognised a check that ran by the
positional __action-stamp spelling and failed on the Windows row; 780 and 790
accept both. The comment on the wrapper stated that an action's command line
survives an upgrade; on the MSVC ABI it gains the toolset's runtime directory,
so each action re-runs once after the upgrade, which the CHANGELOG now lists.

The plan document gains its implementation record (section 8: departures from
section 7 and six findings).

Author: speak-agent

.agents/docs/2026-09-28-ecosystem-design-and-optimisation-plan.md

@@ -619,3 +619,72 @@ verify verify-published.sh against the new and the previous pair in fresh SubOS
        sandboxes with the CN mirror; GalTranslPP on Windows with the new mcpp and
        qt-base revision 1; then the issues are closed with their readings
 ```
+
+## 8. Implementation record (revision 4)
+
+The tasks of §7 were implemented on 2026-09-28 in one pull request per
+repository: openxlings/xlings#628 (X1 to X6), mcpp-community/mcpp#730 (M1 to
+M9), an xim-pkgindex pull request (I1 to I3) and an mcpp-index pull request (N1,
+N2), with the issue mcpplibs/mcpp-index#482 (N3). This section states what
+landed where the implementation departed from §7, what was found while
+implementing it, and the readings.
+
+### 8.1 Departures from §7
+
+- **X3.** `prevLines` is kept, deprecated and always 0, rather than dropped: a
+  minor protocol version only adds (interface specification 1.3).
+- **M4.** The edge-advice channel (SPEC-007 R4.5) is a rule for every build
+  edge, not a mechanism of `place-dlls`: an action writes
+  `.mcpp-advice/<its output>.advice` the same way. e2e 821 reads it through an
+  action on both build paths, because on Linux no engine edge writes it.
+- **M5.** `check_workflow_assertions.py` also accepts a `PIPESTATUS` read on the
+  line after the pipe (rule W2), which the workflows use.
+- **M8.** `verify-published.sh` takes `M` and `XS`, binaries to verify in place
+  of the published ones, so that it can be rehearsed before a release; a run
+  that uses either says so at its start and its end.
+- **I2.** `installed()` does not assert that the runtime files are absent. The
+  packaging revision is what replaces an installed payload (xlings
+  2026.9.27.1), and a line naming the runtime files would be the one the static
+  test of I1 refuses.
+
+### 8.2 Found while implementing
+
+- **F1. A Windows test renamed a directory that a scanner still held.** E2E-01
+  (`bootstrap_home_test.ps1`) renamed the portable home 0.2 s after `self init`
+  and failed with "You do not have sufficient access rights"; the same failure
+  had occurred on 2026-09-14 on a branch that did not touch init, and passed on
+  that branch's next run. The test now renames with `[IO.Directory]::Move`,
+  which either renames or leaves the tree intact, retries for at most 10 s, and
+  on failure names the processes running from the tree.
+- **F2. A second producer of terminal frames.** Off a terminal, the sub-index
+  build scripts (`xim-pkgindex-awesome`, `-scode`, `-d2x`) still write
+  `\r[i/n] <ns>::<file>\033[K`; they run in-process and write to stdout
+  directly. It is the class of #626 in a producer X4 did not cover, and it is
+  present in 2026.9.28.1 (openxlings/xlings#629, open). `verify-published.sh`
+  asserts the download lines and reports these frames as a reading.
+- **F3. The index's sweep alert could not open its issue.** The job checks
+  nothing out, so `gh` could not infer the repository ("not a git
+  repository"), and it watched only the `workspace` job. The two red sweeps of
+  2026-09-26 therefore opened nothing (mcpplibs/mcpp-index#482). N2 sets
+  `GH_REPO` and also runs the alert when `mirror-cn-reachable` fails.
+- **F4. A filtered unit run was reported as the unit suite.** A local run of
+  the mcpp unit binaries under a `GTEST_FILTER` naming the new suites printed
+  "134 passed", which counts binaries; the full suite had one stale expectation
+  (the `place-dlls` command now carries `--crt`), found by the macOS self-host
+  row. The unfiltered run passes.
+- **F5. The toolset on the Visual Studio runner is newer than the Qt payload's
+  runtime.** The measurement recorded MSVC 14.51.36231 (Visual Studio 2026) as
+  the toolset whose runtime directory the action `PATH` receives, against the
+  14.44 copy `xim:qt-base` carried: the case D1 and D2 describe is the ordinary
+  state of a current CI image.
+- **F6. The action PATH crossed the invariant that an action's command line
+  survives an upgrade.** An action that declares neither `env` nor `cwd` kept
+  the positional `__action-stamp` spelling so that its command line, and
+  ninja's command hash, stayed the one an earlier engine wrote. D3 gives every
+  action of an MSVC-ABI build the named wrapper with `--path-prepend`, so each
+  such action re-runs once on the first build after the upgrade. The crossing
+  had no test; e2e 780, which recognised a check that ran by the old spelling,
+  failed on the Windows row and found it. The detector accepts both spellings,
+  the comment in `cli.cppm` states the exception, and the CHANGELOG lists the
+  one re-run.
+

CHANGELOG.md

@@ -75,6 +75,8 @@
 - 若干条件表命中同一目标、且字典序与具体程度给出不同次序的 manifest,其标量取值与列表参数的次序
   随之改变。
 - Windows 上以 GNU 方言编译的工程,编译命令多出 depfile 参数,升级后第一次构建完整重建一次。
+- 面向 MSVC ABI 的构建中,每个 action 的命令行多出工具集运行时目录(`__action --path-prepend`),
+  升级后第一次构建中每个 action(包括 `check` 与 `prepare`)重新运行一次。
 
 ## [2026.9.28.1] - 2026-09-28
 

src/cli.cppm

@@ -1027,10 +1027,13 @@ int run(int argc, char** argv) {
     // `--path-prepend` puts that directory first on the command's PATH, so a
     // tool the action runs (Qt's moc.exe, a vcpkg port's generator) starts
     // with the toolset's runtime and not with whatever copy a library package
-    // happened to ship beside it (the 2026-09-28 design, §2.9). An action that declares neither keeps the
-    // positional `__action-stamp` spelling above, so its command line -- and
-    // ninja's command hash for its edge -- is the one an earlier engine wrote,
-    // and upgrading re-runs no check and no `prepare`.
+    // happened to ship beside it (the 2026-09-28 design, §2.9). Elsewhere, an
+    // action that declares neither keeps the positional `__action-stamp`
+    // spelling above, so its command line -- and ninja's command hash for its
+    // edge -- is the one an earlier engine wrote, and upgrading re-runs no
+    // check and no `prepare`. On the MSVC ABI the first build after upgrading
+    // to 2026.9.28.2 re-runs each action once, because its command line gained
+    // the directory.
     // `mcpp depfile-filter --raw <file> --out <file> -- <compiler argv>...`
     // (internal: written into build.ninja for GCC on a Windows host, WS2 of
     // the 2026-09-28 design). Runs the compile with inherited stdio and no

tests/e2e/780_a_passing_check_moves_its_stamp_past_a_changed_input.sh

@@ -50,7 +50,11 @@ EOF
 
 # `-v` prints each edge ninja runs; the check's is the one that goes through
 # the engine's stamp wrapper.
-ran() { grep -q '__action-stamp' "$1"; }
+# The check ran when its wrapper appears in the verbose log: the positional
+# `__action-stamp`, or the named `__action` that every action of an MSVC-ABI
+# build uses since 2026.9.28.2, because it puts the toolset's C++ runtime first
+# on the action's PATH.
+ran() { grep -qE '__action(-stamp)? ' "$1"; }
 
 "$MCPP" build -v > b1.log 2>&1 || { cat b1.log; echo "FAIL: build failed"; exit 1; }
 ran b1.log || { cat b1.log; echo "FAIL: the first build did not run the check"; exit 1; }

tests/e2e/790_a_prepare_action_populates_an_unknown_directory.sh

@@ -103,7 +103,7 @@ int main() {
 EOF
 
 MCPP="${MCPP:-mcpp}"
-ran_prepare() { grep -q 'PREPARE prep:install\|__action-stamp.*install\.sh' "$1"; }
+ran_prepare() { grep -qE 'PREPARE prep:install|__action(-stamp)? .*install\.sh' "$1"; }
 
 # ── 1. first build: header compiles, program runs through runtime_search_dir
 "$MCPP" build -v > b1.log 2>&1 || { cat b1.log; echo "FAIL: first build failed"; exit 1; }