ci: check OpenVMM's own CI instead of running OpenVMM's tests

.github/actions/build-openvmm/action.yml

@@ -52,6 +52,7 @@ runs:
         echo "path=${cache_tools}" >> "${GITHUB_OUTPUT}"
 
     - name: Restore OpenVMM inputs on Linux
+      id: restore-openvmm-inputs-linux
       if: runner.os != 'Windows'
       uses: actions/cache/restore@v5
       env:
@@ -65,6 +66,7 @@ runs:
           openvmm-inputs-v1-${{ runner.os }}-${{ runner.arch }}-
 
     - name: Restore OpenVMM inputs on Windows
+      id: restore-openvmm-inputs-windows
       if: runner.os == 'Windows'
       uses: actions/cache/restore@v5
       with:
@@ -192,3 +194,70 @@ runs:
           build/openvmm.provenance.json
           ${{ inputs.binary-path }}
         key: openvmm-binary-v6-${{ inputs.backend }}-${{ runner.os }}-${{ runner.arch }}-${{ hashFiles('rust-toolchain.toml') }}-${{ steps.openvmm.outputs.sha }}
+
+    # No other job builds OpenVMM, since OpenVMM's own CI runs its tests
+    # (doc/ci.md#openvmm-ci), so a dev push whose build missed the binary
+    # cache saves the inputs and the Cargo cache that the builds restore.
+    - name: Save OpenVMM inputs on Linux
+      if: >-
+        success() &&
+        github.event_name == 'push' &&
+        github.ref == 'refs/heads/dev' &&
+        runner.os != 'Windows' &&
+        steps.restore-openvmm-binary-linux.outputs.cache-hit != 'true' &&
+        steps.restore-openvmm-inputs-linux.outputs.cache-hit != 'true'
+      uses: actions/cache/save@v5
+      env:
+        PATH: ${{ steps.cache-tools-linux.outputs.path }}
+      with:
+        path: |
+          openvmm/flowey-persist/flowey_lib_common__download_gh_release
+          openvmm/flowey-persist/flowey_lib_common__cache
+        key: openvmm-inputs-v1-${{ runner.os }}-${{ runner.arch }}-${{ steps.openvmm.outputs.sha }}
+
+    - name: Save OpenVMM inputs on Windows
+      if: >-
+        success() &&
+        github.event_name == 'push' &&
+        github.ref == 'refs/heads/dev' &&
+        runner.os == 'Windows' &&
+        steps.restore-openvmm-binary-windows.outputs.cache-hit != 'true' &&
+        steps.restore-openvmm-inputs-windows.outputs.cache-hit != 'true'
+      uses: actions/cache/save@v5
+      with:
+        path: |
+          openvmm/flowey-persist/flowey_lib_common__download_gh_release
+          openvmm/flowey-persist/flowey_lib_common__cache
+        key: openvmm-inputs-v1-${{ runner.os }}-${{ runner.arch }}-${{ steps.openvmm.outputs.sha }}
+
+    - name: Save shared Cargo cache on Linux
+      if: >-
+        success() &&
+        github.event_name == 'push' &&
+        github.ref == 'refs/heads/dev' &&
+        runner.os != 'Windows' &&
+        steps.restore-openvmm-binary-linux.outputs.cache-hit != 'true' &&
+        steps.restore-cargo-linux.outputs.cache-hit != 'true'
+      uses: actions/cache/save@v5
+      env:
+        PATH: ${{ steps.cache-tools-linux.outputs.path }}
+      with:
+        path: |
+          ${{ format('{0}/cargo-home/registry', runner.temp) }}
+          ${{ format('{0}/cargo-home/git', runner.temp) }}
+        key: openvmm-cargo-v1-${{ runner.os }}-${{ runner.arch }}-${{ hashFiles('openvmm/Cargo.lock') }}
+
+    - name: Save shared Cargo cache on Windows
+      if: >-
+        success() &&
+        github.event_name == 'push' &&
+        github.ref == 'refs/heads/dev' &&
+        runner.os == 'Windows' &&
+        steps.restore-openvmm-binary-windows.outputs.cache-hit != 'true' &&
+        steps.restore-cargo-windows.outputs.cache-hit != 'true'
+      uses: actions/cache/save@v5
+      with:
+        path: |
+          ${{ format('{0}/cargo-home/registry', runner.temp) }}
+          ${{ format('{0}/cargo-home/git', runner.temp) }}
+        key: openvmm-cargo-v1-${{ runner.os }}-${{ runner.arch }}-${{ hashFiles('openvmm/Cargo.lock') }}

.github/copilot-instructions.md

@@ -73,8 +73,9 @@
 - The session ends 59 minutes after it starts, setup included, on 4 vCPUs. Run the
   narrowest affected check first. Rebuild OpenVMM, which takes about 5 minutes and
   4 GB, only when the `openvmm` gitlink changed or the cached binary is missing,
-  and guest artifacts only when their inputs changed. Leave `test-openvmm-unit`,
-  `test-openvmm`, `build-guest --guest all`, `verify-guest-determinism`, and
+  and guest artifacts only when their inputs changed. Leave `test-openvmm-unit`
+  and `test-openvmm` to OpenVMM's own CI, whose results NVX CI checks for the
+  pin, and `build-guest --guest all`, `verify-guest-determinism`, and
   `benchmark` to CI.
 - `nvx-rs/target/release/nvx-rs run` and `nvx-rs/target/release/nvx-rs sandbox`
   run microVMs and sandboxes. Build `nvx-rs` there, in about a minute, with

.github/skills/nvx-openvmm-promote/SKILL.md

@@ -291,6 +291,24 @@ unit-test, and formatting gates required by the nested instructions before relyi
 NVX CI. For a pin-only promotion of an already reviewed source, record existing exact-
 head OpenVMM evidence and do not claim unexecuted local checks.
 
+NVX CI runs none of OpenVMM's tests. Its `openvmm-ci` job requires OpenVMM's checkin
+gates to have passed the x86-64 Linux and Windows unit tests and the KVM, MSHV, and
+Windows VMM tests for the pinned SHA, in an OpenVMM PR or OpenVMM CI run whose head is
+that SHA in `microsoft/openvmm`; see [OpenVMM CI](../../../doc/ci.md#openvmm-ci). The
+OpenVMM pull request gets those runs only once it is ready for review, from a branch
+of `microsoft/openvmm`. Record that evidence with the platform launcher:
+
+```bash
+python3 scripts/nvx.py check-openvmm-ci --revision <revision>
+```
+
+```powershell
+python scripts\nvx.py check-openvmm-ci --revision <revision>
+```
+
+A pending job means that OpenVMM's CI still runs; a failed one needs a rerun or a fix
+in OpenVMM before the promotion can pass.
+
 Choose additional NVX runtime checks from the changed behavior and its dependencies.
 Use `nvx-run` for a smoke test, `nvx-debug` for a failure, and `nvx-benchmark` only
 when performance evidence is required. Report unavailable KVM, MSHV, or WHP coverage
@@ -319,7 +337,10 @@ Ask for Draft or Ready state before creation. Ready pull requests start the comp
 [NVX matrix](../../workflows/ci.yml); changing an already active pull request to Draft
 does not cancel its current run. After each push, follow only the workflow run whose
 head SHA matches the current pull-request head. Treat older runs as superseded and use
-`nvx-ci-investigation` for a failed or unexpectedly cancelled current-head run.
+`nvx-ci-investigation` for a failed or unexpectedly cancelled current-head run. The
+run's `openvmm-ci` job waits up to 300 minutes while OpenVMM's CI of the pin runs; when
+it fails because a required OpenVMM job failed, rerun that job in OpenVMM, then the
+`openvmm-ci` job.
 
 `.gitmodules` tracks `feature/microvm/dev`, so the pinned revision must be on that branch
 before the NVX pull request merges. `microsoft/openvmm` allows only squash merges, which