Repository navigation
[Draft] arm64: extend shared cold-runtime validation - #1516
Gautam Ahuja (ahuja-gautam) wants to merge 9 commits into
Conversation
This pull request changes existing testsIt modifies or removes 218 existing test asset(s), listed below. Review each with care: a change to a test can hide a regression that the test caught. See doc/contribute.md#changes-to-existing-tests. Test code
... and 105 more, which the job's summary lists. End-to-end tests... and 8 more, which the job's summary lists. CI test wiring
... and 30 more, which the job's summary lists. Relocated unchanged: 0; added: 23. Compared the merge base |
There was a problem hiding this comment.
🔵 Needs a closer look
Unproved native ARM64 admission and host-time qualification, together with an unmerged OpenVMM pin and temporary CI trust exception, require maintainer verification of landing order and coordinated rollback.
No open issues 🎉
What changed in this PR
This third layer of the ARM64 stack extends NVX’s existing cold-runtime validation to native Windows/WHP. It shares diagnostics, sandbox tests, artifact handling, and CI infrastructure with x86 while keeping ARM snapshot support outside scope.
Changes:
- Add ARM counter-warp diagnostics and architecture-aware CPU, device, interrupt, and console checks.
- Extend native tooling, artifact validation, cold-only adapter selection, and repeated sandbox Start/Stop checks.
- Wire architecture-specific CI inputs and results, remove the shared Windows argument binder’s Bash dependency, and add the narrowly scoped temporary OpenVMM CI evidence exception.
| File | Description |
|---|---|
tests/runtime-adapter/README.md |
Document native ARM execution and cold-only test selection. |
scripts/test-layout-allowlist.json |
Remove obsolete test-hook allowances. |
scripts/test_tool_versions.py |
Cover ARM tool selection and explicit runner host types. |
scripts/test_guest_binaries.py |
Verify ARM inventory includes the shared warp probe. |
scripts/setup/tool-versions.conf |
Pin native Windows ARM tool artifacts. |
scripts/setup/setup-windows-whp.ps1 |
Select native tools, targets, artifacts, and runner labels. |
scripts/setup/setup-linux.sh |
Accept Windows ARM entries in the shared manifest. |
scripts/setup/README.md |
Explain ARM provisioning and runner admission boundaries. |
scripts/nvx_tools/time_abi.py |
Parse ARM profile identities and select counter validation. |
scripts/nvx_tools/sandbox_ram_upper_tests.py |
Repeat RAM-upper starts and verify state resets. |
scripts/nvx_tools/sandbox_lifecycle_tests.py |
Repeat scratch-backed starts and retain per-cycle evidence. |
scripts/nvx_tools/runtime_adapter_tests.py |
Validate native artifacts and select classified cold tests. |
scripts/nvx_tools/openvmm_ci.py |
Require ARM jobs and constrain the temporary fork-run exception. |
scripts/nvx_tools/microvm_test_scripts/virtio-net.sh |
Use shared architecture-aware network-device checks. |
scripts/nvx_tools/microvm_test_scripts/virtio-mmio.sh |
Add fixed MMIO resource and ARM interrupt-binding checks. |
scripts/nvx_tools/microvm_test_scripts/sandbox-blocks.sh |
Check fixed block-device windows across architectures. |
scripts/nvx_tools/microvm_test_scripts/lifecycle.sh |
Observe ARM time, early console, and raw-byte output. |
scripts/nvx_tools/microvm_test_scripts/interrupt-map.sh.in |
Report ARM device-tree bindings and interrupt errors. |
scripts/nvx_tools/microvm_test_scripts/filesystem-shares.sh |
Use shared fixed-device checks for filesystem shares. |
scripts/nvx_tools/microvm_shards.py |
Exclude ARM and cold-only reports from x86 shard estimates. |
scripts/nvx_tools/microvm_durations.py |
Record architecture, scope, and execution variants. |
scripts/nvx_tools/host_time_probe.rs |
Add ARM architected-counter reads and frequency reporting. |
scripts/nvx_tools/guest_binaries.py |
Include the time probe in ARM guest exports. |
scripts/nvx_tools/common.py |
Add host architecture detection and Windows PE checks. |
scripts/nvx_tools/ci.py |
Account for ARM producers, runtime checks, and repetitions. |
scripts/nvx_tools/benchmark.py |
Forward architecture into time monitoring and SMP probes. |
scripts/nvx_tools/benchmark_scripts/smp-probe.sh.in |
Check ARM CPU identities, timers, and interprocessor interrupts. |
nvx-rs/tests/openvmm_cpu_profiles.rs |
Compare ARM catalog identities with pinned OpenVMM profiles. |
nvx-rs/tests/guest_tools.rs |
Exercise the actual ARM warp-probe executable. |
nvx-rs/src/host/cli/cpu_profile.rs |
Keep ARM identities outside x86 CPU guidance. |
nvx-rs/src/host/cli/cpu_generations.json |
Add the registered ARM CPU profile identity. |
nvx-rs/src/guest/time_probe/warp/tests.rs |
Test ARM counter frequency and affinity restoration. |
nvx-rs/src/guest/time_probe/warp.rs |
Share warp algorithms with ARM counter reads. |
nvx-rs/src/guest/time_probe/tests.rs |
Cover ARM output and unsupported-command refusals. |
nvx-rs/src/guest/time_probe/support/tests.rs |
Keep x86-specific parsing tests architecture-gated. |
nvx-rs/src/guest/time_probe/support.rs |
Add ARM frequency metadata and isolate x86 helpers. |
nvx-rs/src/guest/time_probe/mod.rs |
Expose ARM warp commands with architecture-specific help. |
nvx-rs/src/guest/time_probe/format/tests.rs |
Gate tests for x86-only formatting helpers. |
nvx-rs/src/guest/time_probe/format.rs |
Restrict unused clock-formatting helpers to x86. |
nvx-rs/src/guest/sys/scheduler.rs |
Make affinity queries available to ARM production code. |
nvx-rs/src/guest/sys/mod.rs |
Export shared affinity support across architectures. |
nvx-rs/src/guest/sys/clock.rs |
Expose the coarse monotonic clock on ARM. |
nvx-rs/src/guest/status.rs |
Share diagnostic setup-failure statuses with ARM. |
nvx-rs/src/guest/parse.rs |
Enable shared integer parsing on ARM. |
nvx-rs/src/guest/mod.rs |
Enable the ARM time-probe module. |
nvx-rs/src/bin/nvx-time-probe.rs |
Enable the real probe entry point on ARM Linux. |
nvx-rs/README.md |
Document ARM warp commands and counter reporting. |
kernel/patches/arm64/0001-nvx-platform-mmio.patch |
Apply log-level options before early-console output. |
doc/usage.md |
Describe architecture selection, cold scope, and repetitions. |
doc/openvmm-upstream-roadmap.md |
Map the selected ARM foundation and verifier commits. |
doc/ci.md |
Document ARM CI coverage, limitations, and the temporary exception. |
doc/build.md |
Describe the expanded ARM guest-helper inventory. |
.github/workflows/run-nvx-microvm-tests.yml |
Bind shared inputs through Python and route architecture-specific evidence. |
.github/workflows/build-openvmm-binary.yml |
Forward architecture to validation and build actions. |
.github/actions/validate-runner/action.yml |
Validate native architecture and select applicable qualification checks. |
.github/actions/build-openvmm/action.yml |
Verify the Windows binary’s PE architecture. |
.github/actions/build-guest-artifacts/action.yml |
Produce fresh native ARM artifacts in separate output paths. |
Review effort: Balanced
There was a problem hiding this comment.
🔵 Needs a closer look
Maintainer review is required for the temporary fork-CI trust exception and rollback, OpenVMM landing and base compatibility, and unproved native ARM64/WHP NetworkService admission and host-time qualification.
No open issues 🎉
Review effort: Balanced
There was a problem hiding this comment.
🔵 Needs a closer look
Maintainers must verify ARM64/WHP runner and Network Service admission, resolve nonpassing H4/H7 qualification, and coordinate the conflicted OpenVMM pin’s landing and rollback of the temporary fork-CI trust exception.
No open issues 🎉
Review effort: Balanced
There was a problem hiding this comment.
🔵 Needs a closer look
Unverified native ARM64/WHP runner and service-account admission, runtime evidence from a dirty OpenVMM build, and unresolved pin compatibility require maintainer inspection before landing.
No open issues 🎉
Review effort: Balanced
44d5cca to
34aa63a
Compare
34aa63a to
810e881
Compare
There was a problem hiding this comment.
🔵 Needs a closer look
Cross-ISA kernel and timing changes require maintainer assessment because final-head native ARM64/WHP runtime, H4/H7 qualification, service-account admission, and exact-pin integration remain unverified.
2 open issues
Review effort: Balanced
810e881 to
6967db8
Compare
There was a problem hiding this comment.
🟡 Changes recommended
The PR description still names the old unchanged pin and six-commit replay, conflicting with the roadmap’s new baseline and leaving reviewers unable to distinguish current-head validation from historical results.
1 open issue
2 resolved since last review
Review effort: Balanced
Run the maintained host/PE probes, cold scenario inventory, bounded same-provision repetitions and cold runtime-adapter tests on native ARM64/WHP while retaining x86 defaults and full snapshot tails. Carry architecture through existing artifact producers and CI callers, require upstream Windows ARM unit/VMM evidence in addition to x86 jobs, and preserve the existing workflow/repository trust filter. Export the real ARM counter warp helper and correct ISA-specific device, SMP, layered-bootstrap and console observations. Pin the published OpenVMM verifier without rewriting the Phase 1/2 prefix. Native functional evidence covers 45 cold CLI cells, 115 adapter tests and 128 repetition CLI invocations. H4/H7, full UTC qualification and actual CI service admission remain unqualified; no ARM release-catalog or snapshot qualification is claimed. Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
Admit only the authorized upstream-hosted run 43109699910 for PR 5045 and pin 38016, checking the live PR tuple and audited workflow blob. Keep trusted upstream admission and every selected latest-job verdict unchanged. Revert this isolated commit only after the coordinator ends the trial. Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
Run the existing shared Python binder directly with the native Python custom shell. Preserve its Python AST, environment, guards, argument JSON and step outputs. Execute the maintained binder tests with an empty PATH so a Bash dependency cannot be hidden by the test harness. Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
GitHub cannot resolve runner context while validating a step shell. Use the existing backend input, whose WHP jobs run on Windows, to select native Python. Restrict the regression harness to the same available input context and use the Linux backend for the Linux refusal case. Retain the failed startup run as evidence; do not rewrite either isolated commit. Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
GitHub also rejects inputs context in the shell field. Use two mutually exclusive OS-guarded literal Python shells, sharing one anchored script and environment. Preserve the binder AST and JSON and coalesce the selected step output for its existing artifact consumer. Retain both failed startup receipts; do not rewrite history. Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
This reverts commit 19d29eb after the coordinator ended the bounded trial. Restore the exact pretrial evidence policy, tests and documentation while retaining all permanent native-Python correctness binder repairs. The immutable a0 workflow-dispatch evidence remains valid for that historical branch-only trial; do not cancel or redispatch it. Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
Keep the original x86 fixed-NIC command-line assertion, guest-file adapter admission, and one-VP guest-boot/managed dispatch even with explicit SMP processor selections. Apply added native-ISA/static-ELF admission and processor expansion only to ARM, retaining its MMIO/DT and negative admission coverage. Restore the strict unresolved-template assertion over its original template scope. Add focused regression cases for both architectures and document the separate contracts. Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
Download the Azure Linux initramfs for an explicit Azure Linux correctness selection while retaining the debug-kernel exclusion. Preserve both native-Python correctness binder steps unchanged. Match the maintained guest CNTVCT_EL0 read with leading and trailing ISB and a compiler memory barrier, leaving the x86 probe unchanged. Extend the existing artifact-binding and source-version probe regression tests. The exact native ARM probe passed H5 across all 66 CPU pairs at the unchanged 1000 ns bound. Full-schedule H4 rate measurements met their existing limits but H4 still failed because the host publishes a free-running UTC clock (disabled=1); retain that failure, without retries or clock/service changes. This is host-diagnostic evidence, not OpenVMM backend or full host qualification. Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
Recognize the explicit Azure Linux guest-download conjunct in the workflow selector map without weakening its strict expression check. Restore default-x86 progress and execution expectations: boot and managed lifecycle stay at one VP while console exit and SMP variants expand the selected counts. Retain the existing explicit ARM processor-expansion parameterization and exact report assertions. Production code, workflow policy and runtime inputs are unchanged. Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
6967db8 to
24bdced
Compare
There was a problem hiding this comment.
🔵 Needs a closer look
Exact-source ARM64/WHP runtime validation and runner/service-account admission remain unproven, leaving the early-console and counter-ordering changes dependent on maintainer review of trusted platform evidence.
No open issues 🎉
1 resolved since last review
Review effort: Balanced
doc/contribute.md asks every pull request to reference an issue on the first line of its description, but nothing checks it: pull requests such as #1675, whose only references sit in the template's HTML comment, and #1516, whose Refs #1509 names a pull request, reference no issue, and a reviewer has to notice. Add the Issue link workflow, which pull_request_target runs, from the default branch, when a pull request is opened, reopened, pushed to, edited, labelled, or unlabelled. Its script, check_issue_link.py, reads the pull request through gh api and passes it when GitHub links it to an issue of the repository (closingIssuesReferences) or when its description holds close, fix, resolve (in any form), Part of, or Refs, with an optional colon, then #N, microsoft/nvx#N, or the issue's URL, where N is an issue. A pull request's number, a missing number, a bare #N, another repository's issue, and references in HTML comments or Markdown code do not count, nor does a keyword that one of them or the end of a block separates from its number: the script reads quotes, lists, headings, code, and comments as CommonMark does. It looks up at most the first 10 numbers, so a description cannot make it call the API without bound. Pull requests by dependabot[bot] or github-actions[bot], or labelled no-issue, are exempt, and closed ones are left alone. An unlinked pull request gets the needs-issue label and one marker comment that says how to reference an issue, and the check fails; editing the description reruns it, which then removes the label and rewrites the comment. Only a comment of github-actions[bot] that starts with the marker is the script's. The script does every read before its first write, so an API error fails the check and changes nothing, and its summary names the verdict and the numbers, never the description. The job has no condition, since a skipped run would replace the previous verdict, runs one check a pull request at a time, keeps no checkout credentials, and gets only the pull request's number. Its token reads contents and writes pull requests; reading an issue of a public repository needs no issues permission, so the nightly report and issue labelling stay the only workflows that write issues. Evidence: dry runs against the live repository pass #1678, which GitHub links to #1677, flag #1675, #1516, and #1509, as reading their descriptions confirms, and leave merged pull requests such as #1672 alone; before they merged, they passed #1674, which GitHub linked to #1666, and flagged #1671. test_issue_link.py covers the parser, with indented code, fences that end with their container, and hidden text or a block end between a keyword and its number; the verdicts and the writes through a fake gh api, the bound on lookups, an API error at each read, a comment that quotes the marker, and the workflow's triggers, permissions, checkout, concurrency, and inputs. validate-nvx runs it on Linux and Windows. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
doc/contribute.md asks every pull request to reference an issue on the first line of its description, but nothing checks it: pull requests such as #1675, whose only references sit in the template's HTML comment, and #1516, whose Refs #1509 names a pull request, reference no issue, and a reviewer has to notice. Add the Issue link workflow, which pull_request_target runs, from the default branch, when a pull request is opened, reopened, pushed to, edited, labelled, or unlabelled. Its script, check_issue_link.py, reads the pull request through gh api and passes it when GitHub links it to an issue of the repository (closingIssuesReferences) or when its description holds close, fix, resolve (in any form), Part of, or Refs, with an optional colon, then #N, microsoft/nvx#N, or the issue's URL, where N is an issue. A pull request's number, a missing number, a bare #N, another repository's issue, and references in HTML comments or Markdown code do not count, nor does a keyword that one of them or the end of a block separates from its number: the script reads quotes, lists, headings, code, and comments as CommonMark does. It looks up at most the first 10 numbers, so a description cannot make it call the API without bound. Pull requests by dependabot[bot] or github-actions[bot], or labelled no-issue, are exempt, and closed ones are left alone. An unlinked pull request gets the needs-issue label and one marker comment that says how to reference an issue, and the check fails; editing the description reruns it, which then removes the label and rewrites the comment. Only a comment of github-actions[bot] that starts with the marker is the script's. The script does every read before its first write, so an API error fails the check and changes nothing, and its summary names the verdict and the numbers, never the description. The job has no condition, since a skipped run would replace the previous verdict, runs one check a pull request at a time, keeps no checkout credentials, and gets only the pull request's number. Its token reads contents and writes pull requests; reading an issue of a public repository needs no issues permission, so the nightly report and issue labelling stay the only workflows that write issues. Evidence: dry runs against the live repository pass #1678, which GitHub links to #1677, flag #1675, #1516, and #1509, as reading their descriptions confirms, and leave merged pull requests such as #1672 alone; before they merged, they passed #1674, which GitHub linked to #1666, and flagged #1671. test_issue_link.py covers the parser, with indented code, fences that end with their container, and hidden text or a block end between a keyword and its number; the verdicts and the writes through a fake gh api, the bound on lookups, an API error at each read, a comment that quotes the marker, and the workflow's triggers, permissions, checkout, concurrency, and inputs. validate-nvx runs it on Linux and Windows. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
doc/contribute.md asks every pull request to reference an issue on the first line of its description, but nothing checks it: pull requests such as #1675, whose only references sit in the template's HTML comment, and #1516, whose Refs #1509 names a pull request, reference no issue, and a reviewer has to notice. Add the Issue link workflow, which pull_request_target runs, from the default branch, when a pull request is opened, reopened, pushed to, edited, labelled, or unlabelled. Its script, check_issue_link.py, reads the pull request through gh api and passes it when GitHub links it to an issue of the repository (closingIssuesReferences) or when its description holds close, fix, resolve (in any form), Part of, or Refs, with an optional colon, then #N, microsoft/nvx#N, or the issue's URL, where N is an issue. A pull request's number, a missing number, a bare #N, another repository's issue, and references in HTML comments or Markdown code do not count, nor does a keyword that one of them or the end of a block separates from its number: the script reads quotes, lists, headings, code, and comments as CommonMark does, and lets a comment run across blocks, as a browser does, until raw HTML ends it. It looks up at most the first 10 numbers, so a description cannot make it call the API without bound. Pull requests by dependabot[bot] or github-actions[bot], or labelled no-issue, are exempt, and closed ones are left alone. An unlinked pull request gets the needs-issue label and one marker comment that says how to reference an issue, and the check fails; editing the description reruns it, which then removes the label and rewrites the comment. Only a comment of github-actions[bot] that starts with the marker is the script's. The script does every read before its first write, so an API error fails the check and changes nothing, and its summary names the verdict and the numbers, never the description. The job has no condition, since a skipped run would replace the previous verdict, runs one check a pull request at a time, keeps no checkout credentials, and gets only the pull request's number. Its token reads contents and writes pull requests; reading an issue of a public repository needs no issues permission, so the nightly report and issue labelling stay the only workflows that write issues. Evidence: dry runs against the live repository pass #1678, which GitHub links to #1677, flag #1675, #1516, and #1509, as reading their descriptions confirms, and leave merged pull requests such as #1672 alone; before they merged, they passed #1674, which GitHub linked to #1666, and flagged #1671. Each parser case also agrees with what GitHub's Markdown API renders, apart from a reference that a comment splits and a pull request's URL. test_issue_link.py covers the parser, with indented code, fences that end with their container, setext headings, sibling list items, comments that outlast their block or end as a browser ends them, and hidden text or a block end between a keyword and its number; the verdicts and the writes through a fake gh api, the bound on lookups, an API error at each read, a comment that quotes the marker or has no author, and the workflow's triggers, permissions, checkout, concurrency, and inputs. validate-nvx runs it on Linux and Windows. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
doc/contribute.md asks every pull request to reference an issue on the first line of its description, but nothing checks it: pull requests such as #1675, whose only references sit in the template's HTML comment, and #1516, whose Refs #1509 names a pull request, reference no issue, and a reviewer has to notice. Add the Issue link workflow, which pull_request_target runs, from the default branch, when a pull request is opened, reopened, pushed to, edited, labelled, or unlabelled. Its script, check_issue_link.py, reads the pull request through gh api and passes it when GitHub links it to an issue of the repository (closingIssuesReferences) or when its description holds close, fix, resolve (in any form), Part of, or Refs, with an optional colon, then #N, microsoft/nvx#N, or the issue's URL, where N is an issue. A pull request's number, a missing number, a bare #N, another repository's issue, and references in HTML comments or Markdown code do not count, nor does a keyword that one of them or the end of a block separates from its number: the script reads quotes, lists, headings, code, and comments as CommonMark does, and lets a comment run across blocks, as a browser does, until raw HTML ends it. It looks up at most the first 10 numbers, so a description cannot make it call the API without bound. Pull requests by dependabot[bot] or github-actions[bot], or labelled no-issue, are exempt, and closed ones are left alone. An unlinked pull request gets the needs-issue label and one marker comment that says how to reference an issue, and the check fails; editing the description reruns it, which then removes the label and rewrites the comment. Only a comment of github-actions[bot] that starts with the marker is the script's. The script does every read before its first write, so an API error fails the check and changes nothing, and its summary names the verdict and the numbers, never the description. The job has no condition, since a skipped run would replace the previous verdict, runs one check a pull request at a time, keeps no checkout credentials, and gets only the pull request's number. Its token reads contents and writes pull requests; reading an issue of a public repository needs no issues permission, so the nightly report and issue labelling stay the only workflows that write issues. Evidence: dry runs against the live repository pass #1678, which GitHub links to #1677, flag #1675, #1516, and #1509, as reading their descriptions confirms, and leave merged pull requests such as #1672 alone; before they merged, they passed #1674, which GitHub linked to #1666, and flagged #1671. Each parser case also agrees with what GitHub's Markdown API renders, apart from a reference that a comment splits and a pull request's URL. test_issue_link.py covers the parser, with indented code, fences that end with their container, setext headings, sibling list items, comments that outlast their block or end as a browser ends them, and hidden text or a block end between a keyword and its number; the verdicts and the writes through a fake gh api, the bound on lookups, an API error at each read, a comment that quotes the marker or has no author, and the workflow's triggers, permissions, checkout, concurrency, and inputs. validate-nvx runs it on Linux and Windows. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
doc/contribute.md asks every pull request to reference an issue on the first line of its description, but nothing checks it: pull requests such as #1675, whose only references sit in the template's HTML comment, and #1516, whose Refs #1509 names a pull request, reference no issue, and a reviewer has to notice. Add the Issue link workflow, which pull_request_target runs, from the default branch, when a pull request is opened, reopened, pushed to, edited, labelled, or unlabelled. Its script, check_issue_link.py, reads the pull request through gh api, in UTF-8 whatever the locale, and passes it when GitHub links it to an issue of the repository (closingIssuesReferences) or when its description holds close, fix, resolve (in any form), Part of, or Refs, with an optional colon, then #N, microsoft/nvx#N, or the issue's URL, where N is an issue. A pull request's number, a missing number, a bare #N, another repository's issue, and references in HTML comments or Markdown code do not count, nor does a keyword that one of them or the end of a block separates from its number: the script reads quotes, lists, headings, code, and HTML blocks as GitHub's CommonMark does, reads raw HTML as text but for its comments, and lets a comment run across blocks, as a browser does, until raw HTML ends it. It looks up at most the first 10 numbers, so a description cannot make it call the API without bound. Pull requests by dependabot[bot] or github-actions[bot], or labelled no-issue, are exempt, and closed ones are left alone. An unlinked pull request gets the needs-issue label and one marker comment that says how to reference an issue, and the check fails; editing the description reruns it, which then removes the label and rewrites the comment. Only a comment of github-actions[bot] that starts with the marker is the script's. The script does every read before its first write, so an API error fails the check and changes nothing, and its summary names the verdict and the numbers, never the description. The job has no condition, since a skipped run would replace the previous verdict, runs one check a pull request at a time, keeps no checkout credentials, and gets only the pull request's number. Its token reads contents and writes pull requests; reading an issue of a public repository needs no issues permission, so the nightly report and issue labelling stay the only workflows that write issues. Evidence: dry runs against the live repository pass #1678, which GitHub links to #1677, flag #1675, #1516, and #1509, as reading their descriptions confirms, and leave merged pull requests such as #1672 alone; before they merged, they passed #1674, which GitHub linked to #1666, and flagged #1671. Each parser case also agrees with what GitHub's Markdown API renders, apart from a reference that a comment splits, a pull request's URL, and the text of a declaration, which GitHub hides. test_issue_link.py covers the parser, with indented code, fences that end with their container, setext headings, sibling and empty list items, HTML blocks, comments that outlast their block or end as a browser ends them, and hidden text or a block end between a keyword and its number; the verdicts and the writes through a fake gh api, gh's input and output in UTF-8, the bound on lookups, an API error at each read, a comment that quotes the marker or has no author, and the workflow's triggers, permissions, checkout, concurrency, and inputs. validate-nvx runs it on Linux and Windows. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
doc/contribute.md asks every pull request to reference an issue on the first line of its description, but nothing checks it: pull requests such as #1675, whose only references sit in the template's HTML comment, and #1516, whose Refs #1509 names a pull request, reference no issue, and a reviewer has to notice. Add the Issue link workflow, which pull_request_target runs, from the default branch, when a pull request is opened, reopened, pushed to, edited, labelled, or unlabelled. Its script, check_issue_link.py, reads the pull request through gh api, in UTF-8 whatever the locale, and passes it when GitHub links it to an issue of the repository (closingIssuesReferences) or when its description holds close, fix, resolve (in any form), Part of, or Refs, with an optional colon, then #N, microsoft/nvx#N, or the issue's URL, where N is an issue. A pull request's number, a missing number, a bare #N, and another repository's issue do not count. The script reads the description as GitHub renders it, through the Markdown API in the repository's context, and only the text that GitHub shows counts: not an HTML comment, an attribute such as a link's title, or code, inline, fenced, indented, or in <pre>, nor a keyword that code or the end of a block, such as a list item or a table cell, separates from its number. GitHub's own renderer keeps the check in step with what the pull request's page shows, which a parser of CommonMark, GFM tables, and raw HTML written here kept missing. It looks up at most the first 10 numbers, so a description cannot make it call the API without bound. Pull requests by dependabot[bot] or github-actions[bot], or labelled no-issue, are exempt, and closed ones are left alone. An unlinked pull request gets the needs-issue label and one marker comment that says how to reference an issue, and the check fails; editing the description reruns it, which then removes the label and rewrites the comment. Only a comment of github-actions[bot] that starts with the marker is the script's. The script does every read before its first write, so an API error fails the check and changes nothing, and its summary names the verdict and the numbers, never the description. The job has no condition, since a skipped run would replace the previous verdict, runs one check a pull request at a time, keeps no checkout credentials, and gets only the pull request's number. Its token reads contents and writes pull requests; reading an issue of a public repository needs no issues permission, so the nightly report and issue labelling stay the only workflows that write issues. Evidence: dry runs against the live repository pass #1687, #1686, and #1682, which GitHub links to issues, and #1679, whose description references #1676, and flag #1675, #1516, #1510, and #1509, as reading their descriptions confirms. A 64 KiB description renders in about a second, and its text reads in at most 60 ms. test_issue_link.py covers the reading of GitHub's HTML, as gh api markdown renders it: keywords and targets, other repositories, code and <pre>, attributes, the ends of blocks and table cells, and inline elements; the verdicts and the writes through a fake gh api, the render request, gh's input and output in UTF-8, the bound on lookups, an API error at each read, a comment that quotes the marker or has no author, and the workflow's triggers, permissions, checkout, concurrency, and inputs. validate-nvx runs it on Linux and Windows. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
doc/contribute.md asks every pull request to reference an issue on the first line of its description, but nothing checks it: pull requests such as #1675, whose only references sit in the template's HTML comment, and #1516, whose Refs #1509 names a pull request, reference no issue, and a reviewer has to notice. Add the Issue link workflow, which pull_request_target runs, from the default branch, when a pull request is opened, reopened, pushed to, edited, labelled, or unlabelled. Its script, check_issue_link.py, reads the pull request through gh api, in UTF-8 whatever the locale, and passes it when GitHub links it to an issue of the repository (closingIssuesReferences) or when its description holds close, fix, resolve (in any form), Part of, or Refs, with an optional colon, then #N, microsoft/nvx#N, or the issue's URL, where N is an issue. A pull request's number, a missing number, a bare #N, and another repository's issue do not count. The script reads the description as GitHub renders it, through the Markdown API in the repository's context, and only the text that GitHub shows counts: not an HTML comment, an attribute such as a link's title, or code, inline, fenced, indented, or in <pre>, nor a keyword that code or the end of a block, such as a list item or a table cell, separates from its number. A link counts by where it goes, whatever its text: a link to an issue or a pull request as OWNER/REPO#N, and any other link as code does. GitHub's own renderer keeps the check in step with what the pull request's page shows, which a parser of CommonMark, GFM tables, and raw HTML written here kept missing. It looks up at most the first 10 numbers, so a description cannot make it call the API without bound. Pull requests by dependabot[bot] or github-actions[bot], or labelled no-issue, are exempt, and closed ones are left alone. An unlinked pull request gets the needs-issue label and one marker comment that says how to reference an issue, and the check fails; editing the description reruns it, which then removes the label and rewrites the comment. Only a comment of github-actions[bot] that starts with the marker is the script's. The script does every read before its first write, so an API error fails the check and changes nothing, and its summary names the verdict and the numbers, never the description. The job has no condition, since a skipped run would replace the previous verdict, runs one check a pull request at a time, keeps no checkout credentials, and gets only the pull request's number. Its token reads contents and writes pull requests; reading an issue of a public repository needs no issues permission, so the nightly report and issue labelling stay the only workflows that write issues. Evidence: dry runs against the live repository pass #1687, #1686, and #1682, which GitHub links to issues, and #1679, whose description references #1676, and flag #1675, #1516, #1510, and #1509, as reading their descriptions confirms. A 64 KiB description renders in about a second, and its text reads in at most 60 ms. test_issue_link.py covers the reading of GitHub's HTML, as gh api markdown renders it: keywords and targets, other repositories, code and <pre>, attributes, links, the ends of blocks and table cells, and inline elements; the verdicts and the writes through a fake gh api, the render request, gh's input and output in UTF-8, the bound on lookups, an API error at each read, a comment that quotes the marker or has no author, and the workflow's triggers, permissions, checkout, concurrency, and inputs. validate-nvx runs it on Linux and Windows. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
doc/contribute.md asks every pull request to reference an issue on the first line of its description, but nothing checks it: pull requests such as #1675, whose only references sit in the template's HTML comment, and #1516, whose Refs #1509 names a pull request, reference no issue, and a reviewer has to notice. Add the Issue link workflow, which pull_request_target runs, from the default branch, when a pull request is opened, reopened, pushed to, edited, labelled, or unlabelled. Its script, check_issue_link.py, reads the pull request through gh api, in UTF-8 whatever the locale, and passes it when GitHub links it to an issue of the repository (closingIssuesReferences) or when its description holds close, fix, resolve (in any form), Part of, or Refs, with an optional colon, then #N, microsoft/nvx#N, or the issue's URL, where N is an issue. A pull request's number, a missing number, a bare #N, and another repository's issue do not count. The script reads the description as GitHub renders it, through the Markdown API in the repository's context, and only the text that GitHub shows counts: not an HTML comment, an attribute such as a link's title, or code, inline, fenced, indented, or in <pre>, nor a keyword that code or the end of a block, such as a list item or a table cell, separates from its number. A link counts by where it goes, and of its text only a keyword that starts it: a link to an issue or a pull request, absolute or root-relative, with nothing after the number other than a slash, a query, or a fragment, as that keyword and OWNER/REPO#N, and any other link as code does. GitHub's own renderer keeps the check in step with what the pull request's page shows, which a parser of CommonMark, GFM tables, and raw HTML written here kept missing. It looks up at most the first 10 numbers, so a description cannot make it call the API without bound. Pull requests by dependabot[bot] or github-actions[bot], or labelled no-issue, are exempt, and closed ones are left alone. An unlinked pull request gets the needs-issue label and one marker comment that says how to reference an issue, and the check fails; editing the description reruns it, which then removes the label and rewrites the comment. Only a comment of github-actions[bot] that starts with the marker is the script's. The script does every read before its first write, so an API error fails the check and changes nothing, and its summary names the verdict and the numbers, never the description. The job has no condition, since a skipped run would replace the previous verdict, runs one check a pull request at a time, keeps no checkout credentials, and gets only the pull request's number. Its token reads contents and writes pull requests; reading an issue of a public repository needs no issues permission, so the nightly report and issue labelling stay the only workflows that write issues. Evidence: dry runs against the live repository pass #1687, #1686, and #1682, which GitHub links to issues, and #1679, whose description references #1676, and flag #1675, #1516, #1510, and #1509, as reading their descriptions confirms. A 64 KiB description renders in about a second, and its text reads in at most 60 ms. test_issue_link.py covers the reading of GitHub's HTML, as gh api markdown renders it: keywords and targets, other repositories, code and <pre>, attributes, links, the ends of blocks and table cells, and inline elements; the verdicts and the writes through a fake gh api, the render request, gh's input and output in UTF-8, the bound on lookups, an API error at each read, a comment that quotes the marker or has no author, and the workflow's triggers, permissions, checkout, concurrency, and inputs. validate-nvx runs it on Linux and Windows. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
doc/contribute.md asks every pull request to reference an issue on the first line of its description, but nothing checks it: pull requests such as #1675, whose only references sit in the template's HTML comment, and #1516, whose Refs #1509 names a pull request, reference no issue, and a reviewer has to notice. Add the Issue link workflow, which pull_request_target runs, from the default branch, when a pull request is opened, reopened, pushed to, edited, labelled, or unlabelled. Its script, check_issue_link.py, reads the pull request through gh api, in UTF-8 whatever the locale, and passes it when GitHub links it to an issue of the repository (closingIssuesReferences) or when its description holds close, fix, resolve (in any form), Part of, or Refs, with an optional colon, then #N, microsoft/nvx#N, or the issue's URL, where N is an issue. A pull request's number, a missing number, a bare #N, and another repository's issue do not count. The script reads the description as GitHub renders it, through the Markdown API in the repository's context, and only the text that GitHub shows counts: not an HTML comment, an attribute such as a link's title, or code, inline, fenced, indented, or in <pre>, nor a keyword that code or the end of a block, such as a list item or a table cell, separates from its number. A link counts by where it goes, and of its text only a keyword that starts it: a link to an issue or a pull request, absolute or root-relative, with nothing after the number other than a slash, a query, or a fragment, as that keyword and OWNER/REPO#N, which no text after the link extends, and any other link as code does. An issue's URL that GitHub shows as text follows the same rule, but for the punctuation that ends a sentence. GitHub's own renderer keeps the check in step with what the pull request's page shows, which a parser of CommonMark, GFM tables, and raw HTML written here kept missing. It looks up at most the first 10 numbers, so a description cannot make it call the API without bound. Pull requests by dependabot[bot] or github-actions[bot], or labelled no-issue, are exempt, and closed ones are left alone. An unlinked pull request gets the needs-issue label and one marker comment that says how to reference an issue, and the check fails; editing the description reruns it, which then removes the label and rewrites the comment. Only a comment of github-actions[bot] that starts with the marker is the script's. The script does every read before its first write, so an API error fails the check and changes nothing, and its summary names the verdict and the numbers, never the description. The job has no condition, since a skipped run would replace the previous verdict, runs one check a pull request at a time, keeps no checkout credentials, and gets only the pull request's number. Its token reads contents and writes pull requests; reading an issue of a public repository needs no issues permission, so the nightly report and issue labelling stay the only workflows that write issues. Evidence: dry runs against the live repository pass #1687, #1686, and #1682, which GitHub links to issues, and #1679, whose description references #1676, and flag #1675, #1516, #1510, and #1509, as reading their descriptions confirms. A 64 KiB description renders in about a second, and its text reads in at most 60 ms. test_issue_link.py covers the reading of GitHub's HTML, as gh api markdown renders it: keywords and targets, other repositories, code and <pre>, attributes, links, the ends of blocks and table cells, and inline elements; the verdicts and the writes through a fake gh api, the render request, gh's input and output in UTF-8, the bound on lookups, an API error at each read, a comment that quotes the marker or has no author, and the workflow's triggers, permissions, checkout, concurrency, and inputs. validate-nvx runs it on Linux and Windows. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
doc/contribute.md asks every pull request to reference an issue on the first line of its description, but nothing checks it: pull requests such as #1675, whose only references sit in the template's HTML comment, and #1516, whose Refs #1509 names a pull request, reference no issue, and a reviewer has to notice. Add the Issue link workflow, which pull_request_target runs, from the default branch, when a pull request is opened, reopened, pushed to, edited, labelled, or unlabelled. Its script, check_issue_link.py, reads the pull request through gh api, in UTF-8 whatever the locale, and passes it when GitHub links it to an issue of the repository (closingIssuesReferences) or when its description holds close, fix, resolve (in any form), Part of, or Refs, with an optional colon, then #N, microsoft/nvx#N, or the issue's URL, where N is an issue. A pull request's number, a missing number, a bare #N, and another repository's issue do not count. The script reads the description as GitHub renders it, through the Markdown API in the repository's context, and only the text that GitHub shows counts: not an HTML comment, an attribute such as a link's title, or code, inline, fenced, indented, or in <pre>, nor a keyword that code or the end of a block, such as a list item or a table cell, separates from its number. A link counts by where it goes, and of its text only a keyword that starts it: a link to an issue or a pull request, absolute or root-relative, with nothing after the number other than a slash, a query, or a fragment, as that keyword and OWNER/REPO#N, which no text after the link extends, and any other link as code does. An issue's URL that GitHub shows as text follows the same rule, but for the punctuation that ends a sentence. GitHub's own renderer keeps the check in step with what the pull request's page shows, which a parser of CommonMark, GFM tables, and raw HTML written here kept missing. It looks up at most the first 10 numbers, so a description cannot make it call the API without bound. Pull requests by dependabot[bot] or github-actions[bot], or labelled no-issue, are exempt, and closed ones are left alone. An unlinked pull request gets the needs-issue label and one marker comment that says how to reference an issue, and the check fails; editing the description reruns it, which then removes the label and rewrites the comment. Only a comment of github-actions[bot] that starts with the marker is the script's. The script does every read before its first write, so an API error fails the check and changes nothing, and its summary names the verdict and the numbers, never the description. The job has no condition, since a skipped run would replace the previous verdict, runs one check a pull request at a time, keeps no checkout credentials, and gets only the pull request's number. Its token reads contents and writes pull requests; reading an issue of a public repository needs no issues permission, so the nightly report and issue labelling stay the only workflows that write issues. Evidence: dry runs against the live repository pass #1687, #1686, and #1682, which GitHub links to issues, and #1679, whose description references #1676, and flag #1675, #1516, #1510, and #1509, as reading their descriptions confirms. A 64 KiB description renders in about a second, and its text reads in at most 60 ms. test_issue_link.py covers the reading of GitHub's HTML, as gh api markdown renders it: keywords and targets, other repositories, code and <pre>, attributes, links, the ends of blocks and table cells, and inline elements; the verdicts and the writes through a fake gh api, the render request, gh's input and output in UTF-8, the bound on lookups, an API error at each read, a comment that quotes the marker or has no author, and the workflow's triggers, permissions, checkout, concurrency, and inputs. validate-nvx runs it on Linux and Windows. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>


Refs #1509, #1510.
What and why
Current published Phase 3 slice and qualification (24bd)
Head
24bdced073a24418c1ebbf1d0775bb5ba42266ba, treeb9a15cd955dda2352d2cfe94b39fbdd752ef3685, sole parent2000aa078713467add77565acdae688a6d6ecfcf, is on Phase 2aacccd99e132233f4ae6e440b30aeadf941031f4and frozen NVX devc9817ae73263a0c32498dd43ebeb089164af8cbe. All nine layer commits / sixteen feature commits above those bases have one parent and zero new merges. Stack1517remains #1509 -> #1510 -> #1516, all actual Draft, published heads3103f24c / aacccd99 / 24bdced0, no queue/auto-merge. Later livedev1ad73759022f2ded66cfc93197288064cb02ad6aand its OpenVMM09cd771c061aa1d1b406b41eaf2eda851db87c15are observations only, not incorporated or qualified.Mergeability is not clean: fresh GraphQL/REST report CONFLICTING / DIRTY /
mergeable=false. PR3'sbaseRefOid,baseRef.target, RESTbase.sha, and actual Phase 2 branch all equalaacccd99; its head fields/actual branch equal24bd. That immediate base is an ancestor of24bd, not stale413586. The live-dev/full-stack comparison diverges fromc981(16 feature / 6 later-dev commits); pins09cdandb09diverge from73cf(4 upstream / 3 ARM commits). This is consistent with cumulative native final-base assessment, not proof of GitHub's internal cause or a complete conflict-path list. No hypothetical conflict resolution, retarget, merge, rebase or latest-dev intake is performed or authorized by this description.The selected OpenVMM pin is
b09c58843ae310dfff9ae68e9e822f9c5846b914, tree0ab5ac89b10bf83d0a129d0a328abe682bc97835, the published head of microsoft/openvmm#5045: three linear ARM/import commits on73cf403574a8907ee94dab0ab9d648152b832b05, roadmap positions 270-272 after inherited 269, mapped tip 124. This v6.1 baseline includes the reviewed upstream PIT/RTC/loader/state changes; the old C5/5d/6036/no-PIT descriptions below are historical, not current. The source signature remains unsigned, not verified. The actual tested native ARM64 VMM reportsopenvmm 0.5.0+gb09c58843, executable SHA256f171f776a76c7aeef5faae7221e7a508d4d3de48f7e62f227bc7fbc372eb7653, owner provenance SHA256fea523f0e1f3114a0920a79ec3679f16fd8c3ec45fe0254f6269c691b9eb6691. Its real build has opt-level 0, debug info 2, debug assertions/overflow checks enabled; its directory name is not an optimized-release claim. OpenVMM must still land before NVX; no landing or trusted-CI admission is implied.Runtime executed at
2000, tests at24bd: the sole ordinary test-only follow-up changes two Python test files / three methods, 6 insertions / 9 deletions, correcting the missing legitimate Azure Linux selector conjunct and restoring default-x86 progress/report expectations. It leaves all 1,032 other tracked mode/type/blob/path rows unchanged, digest73ae1805271658bcb25d876f5bd113afee42a70a05830713901ae56abdd0da28. Independent bridge SHA2567787029762c5560e2420c41a8986bc9c2839d27078cd5ab484c395d58cc4c3b0verifies product/source identity: the2000runtime results apply to24bdby that identity, not by claiming VM reexecution at24bd. Old f775/38016/6036/0d9460dc smokes and old matrices remain historical only.Current
aacccd99 -> 24bd: 218 modified existing test/wiring assets, 23 additions, zero unchanged relocations, not 218 methods.python scripts\check_test_changes.py --base aacccd99e132233f4ae6e440b30aeadf941031f4 --head 24bdced073a24418c1ebbf1d0775bb5ba42266ba --repository microsoft/nvxPASS. Follow-up-only2000 -> 24bdis 3 modified / 0 added, also PASS. The restored progress test is AST-identical to the lower base, explaining the prior 219-to-218 reduction; no test was removed/skipped.tests-modified, the original 29-file rationale and all historical receipts below remain; original 215/18, 217/20, 217/23 and 219/23 counts retain only their named old scopes.Legacy x86 NIC/admission/one-VP boot+managed/template assertions and explicit ARM expansion remain covered. Azure Linux selected-guest download, literal Windows/Linux native-Python binders, selected-ISA/debug-config admission, and ISB/MRS/ISB with no
nomemcounter ordering survive. The workflow remains blob7281163234a69bdbd13624622f7bf41137590456; the ordinary strict helper remains original blobec94c583a73e0a8b22aef3249ca198f1895e9ab1, raw SHA2565381eb15c7a1b0aafa03332ab6fe8abe3f48c1a12bbe8e3e103303a185e9df57. No active_TRIAL_allowance/tests/docs, double-revert or fork-proof substitution exists. All remaining qualification/proof limits are listed below; this is not overall green CI or host qualification.Historical first guarded rebase (f865 -> 34aa; superseded)
The historical first in-place rebase replayed only the same six Phase 3 commits onto the reviewed Phase 2 tip
f865578bdd7a4767e4b208f67d99792cb9215ac2, whose reviewed Phase 1 ancestor isd06bf47f89eddf2fa68edeb99caf3fd697d35f45. The historical head was34aa63aacad12bac4943727c0c465af33f949853, treec089a20a5977cde44816a6795cf8c5166889d033. No dev intake, merge, pin change, new implementation commit, unstack/recreate, or auto-merge is part of this rebase. Native stack1517stays #1509 -> #1510 -> #1516, all actual Ready; the literal[Draft]title is retained. Earlier Draft-state descriptions below are historical.The replay was conflict-free. All six
git range-diff cb017cf8..44d5cca8 f865578b..34aa63aaentries are equal (=). Ignoring hunk offsets/blob headers, the old-top-to-new-top semantic patch is exactly the old-base-to-new-base lower-review patch (10 files, 209 insertions/12 deletions): native source-archive selection and exact option spellings, ARM remembered-directory path admission, and the inert Windows PE fixture correction. All lower regression ASTs and byte-identical Rust/path-admission files remain in the final tree. Phase 3's explicit monitor architecture override/native fallback regression remains AST-identical. Local backup refs preserve the original top/base and reviewed new base underrefs/backup/nvx-phase3-rebase-20261112-2025/.f77577b4963b7b4ae8f8e73e5c0b6728cb8b19df587f5dba36346325435e1b3334cfd0260544a50319d29ebc7a6b4d825fb070317f4d534be5d59b3c1cb8daba8cbbb6d9cd1547bf508048bf55a16fa7934b9c2a85b2c8f7762688cba9d76b5c44fe26917442fa08fdc5206c15d293da77be8683e9dff9138e0016bb65e28e9b6baaec2af1fc4c1337b08cceae932b9e2b5660e0ed69dba69785e6cb955f750ea0f5e4e336cda405eacc1142b8267563d292c5c1d926833d60aa28fa801b1522605940b225330e2f44d5cca8e296286515a28ac7515820f42266e5f034aa63aacad12bac4943727c0c465af33f949853The temporary allowance remains reverted in the final tree. The ordinary OpenVMM CI helper is raw-byte/AST-identical to the original blob
ec94c583a73e0a8b22aef3249ca198f1895e9ab1(SHA2565381eb15c7a1b0aafa03332ab6fe8abe3f48c1a12bbe8e3e103303a185e9df57); no_TRIAL_allowance is active. The permanent Windows literal-Python binder workflow remains exact blobf93c8be4922720dd537750f4c107a80c25d2dcd1. The NVX OpenVMM gitlink is still38016f6637f0c61fb55c9ee31033ceec428e346c, not a newer OpenVMM PR head.Historical first-rebase committed slice
f865578b -> 34aa63aa: 217 existing test/wiring changes, 20 additions, zero unchanged relocations.python scripts\check_test_changes.py --base f865578bdd7a4767e4b208f67d99792cb9215ac2 --head 34aa63aacad12bac4943727c0c465af33f949853 --repository microsoft/nvx --report <comment-report> --summary <full-report>PASS, exit 0. Full maintained report SHA2560b9970b7f280a32c562e8f24c53c2e66a47a419a9f1088e31980224a37665c76; per-file rationale below remains applicable andtests-modifiedremains required.Historical first-rebase resource guards passed immediately before the scoped test/static/formatting stages (
pr3-20261112-rebased-static-prelaunch-01,-02, andpr3-20261112-rebased-format-prelaunch-03). Each checkedmin(available physical, available commit) >= 3072 MiB, found no matching compiler/VMM/Docker processes, and recorded Docker API unavailability as unknown running containers, not zero.Historical first-rebase validation used the maintained Python interpreter via
$Pythonas defined below:& $Python -m unittest scripts.test_guest_binaries.GuestBinaryArchitectureTests scripts.test_time_abi.MonitorTests scripts.test_nvx_tools.OpenVMMCiTests scripts.test_nvx_tools.CiConfigurationTests -qPASS, 140 selected/134 passed/six unchanged platform skips.& $Python -m ruff check scripts .github\speculaand& $Python -m ruff format --check scripts .github\speculaPASS (127 files formatted);& $Python -m pyright --pythonplatform Linux --pythonpath $Pythonand the Windows view PASS, zero errors/warnings, withPYTHONPATHunset.& $Python scripts\check_test_layout.py,& $Python scripts\check_markdown_links.py,& $Python scripts\nvx.py check-kernel-configand the same inventory check with--architecture aarch64PASS; six maintained PowerShell setup files parsed without errors.cargo +1.97.1 fmt --all --check --manifest-path nvx-rs\Cargo.tomland the same command fortests\runtime-adapter\Cargo.tomlPASS.git diff --check f865578bdd7a4767e4b208f67d99792cb9215ac2 HEADPASS.& $Python scripts\local_gates.py --base f865578bdd7a4767e4b208f67d99792cb9215ac2PASS as the applicable-gate inventory. The initial system-Python invocation lacked PyYAML; selecting the existing maintained dependency environment resolved that tooling prerequisite without a source change. This scope did not rerun the full Rust/Python action suites, unavailable submodule-coupled checks, native builds, heavy VM campaigns, or Linux-only shell/Specula/platform qualification. Rebase inspection, exact lower-fix retention, focused tests and static checks do not claim those gates passed.The bounded strict command
python scripts\nvx.py check-openvmm-ci --revision 38016f6637f0c61fb55c9ee31033ceec428e346c --architecture aarch64retains expected exit 1 / eight missing-evidence results under the restored trust policy. H4/H7/full UTC/host and actual native ARM runner admission remain unqualified. The old manual run43162791571checked the historicala0f5e4e3checkout only; neither its success nor the reviewed lower-layer evidence is relabeled as CI/native execution of34aa63aa, the replayed intermediated926833d, or final-base integration.Historical pre-rebase full Phase 3 committed test accounting (
cb017cf8aee13e5e7276236d1b0ab328e1b52e7f->44d5cca8e296286515a28ac7515820f42266e5f0): 217 existing test/wiring changes, 20 added items, zero relocated unchanged.python scripts\check_test_changes.py --base cb017cf8aee13e5e7276236d1b0ab328e1b52e7f --head 44d5cca8e296286515a28ac7515820f42266e5f0 --repository microsoft/nvx --report <comment-report> --summary <full-report>passed as the maintained committed inventory command. Full report SHA256:01509f794a010186d59b6dca163c9651d75be468fc52f4c6034ef0b28f6eec8f. The original 215/18 below is explicitly historicalCB017 -> f775accounting; rollback-only 8/0 isa0 -> 44, not this current full PR slice. All per-file rationale remains applicable andtests-modifiedremains required.TEMPORARY evidence-admission trial: completed and reverted
Rollback completed in
44d5cca8e296286515a28ac7515820f42266e5f0, a normal new commit whose sole parent isa0f5e4e336cda405eacc1142b8267563d292c5c1. The coordinator explicitly ended the bounded trial after its objectives passed. Only isolated temporary commit19d29ebc7a6b4d825fb070317f4d534be5d59b3cwas reverted; all permanent934b9c2a/8e0016bb/a0f5e4e3portability repairs remain.scripts/nvx_tools/openvmm_ci.pyis raw-byte/blob/AST-identical to pretrialf775, blobec94c583a73e0a8b22aef3249ca198f1895e9ab1(SHA2565381eb15c7a1b0aafa03332ab6fe8abe3f48c1a12bbe8e3e103303a185e9df57). No active trial allowance, trial policy documentation or temporary policy tests remain in the repository. The native-shell workflow remains exactly thea0blobf93c8be4922720dd537750f4c107a80c25d2dcd1; all permanent binder/configuration AST nodes and five documentation lines remain unchanged. The history below records the completed trial, not an active exception or pending rollback.Historical hosted trial success remains bound to immutable
a0, not the rollback head: workflow-dispatch run 43162791571, attempt 1, checked out exacta0f5e4e336cda405eacc1142b8267563d292c5c1. Its pin gate completed successfully with the exact temporary-proof warning and all eight actual passed result lines; raw log SHA256ca51bdafcc13c64fec01de3d66edd9e801ca591f363c096eaf6456ad22af282f. Actual Windows binder steps passed in guests, adapter, shard 1 and shard 2, with each Linux alternative skipped by its OS guard. These are real step outcomes, not full runtime-job, native ARM admission, UTC/host or final-base qualification. The existing dispatch was not cancelled or redispatched and may continue independently after the revert.One fresh post-revert snapshot at 2026-11-12 06:00:36 UTC still reports that immutable
a0dispatch queued overall: 35 successful jobs, six skipped, two queued, zero failed. All four Windows consumer jobs named above have now completed successfully, not just their binders. The native ARM CLI/VMM jobs remain queued without assigned runners. This newer snapshot does not establish a completed full workflow, native ARM admission or final-base integration.Rollback validation:
python -m unittest scripts.test_nvx_tools.OpenVMMCiTests scripts.test_nvx_tools.CiConfigurationTests -qPASS, 106 selected/100 passed/six unchanged platform skips; focused Ruff lint/format and strict Linux/Windows type views PASS, zero type errors/warnings; layout, Markdown links and staged whitespace checks PASS. The bounded actual normal commandpython scripts\nvx.py check-openvmm-ci --revision 38016f6637f0c61fb55c9ee31033ceec428e346c --architecture aarch64now returns expected exit 1 / eight missing-evidence results, without the trial warning: restored fork rejection, not a new regression.python scripts\check_test_changes.py --base a0f5e4e336cda405eacc1142b8267563d292c5c1passed at the committed revert, reporting eight existing test changes/no additions: restore the two original fake-API helpers, remove trial-onlygetand_trial_apihelpers, and retire the four trial-only admission/mismatch/latest-job tests with the allowance they exercised. Permanent binder tests are preserved;tests-modifiedremains.The user explicitly authorized a reversible trial after the original
f775publication. Temporary commit19d29ebc7a6b4d825fb070317f4d534be5d59b3c, parentf77577b4963b7b4ae8f8e73e5c0b6728cb8b19df, admits only upstream-hosted OpenVMM run43109699910, PR microsoft/openvmm#5045, pin38016f6637f0c61fb55c9ee31033ceec428e346c, forkahuja-gautam/openvmm, head refopenvmm-arm64-microvm-dev-20261111, eventpull_request, and workflow.github/workflows/openvmm-pr.yamlwith audited Git blob2cabb910cd6d8c974b809a30870502f69ac68e83(421776-byte regular file). The live PR must remain open/unmerged with that head and the expected upstream base repository/ref; no moving base SHA is trusted. Missing/malformed/mismatched proof fails closed. The ordinary trusted-upstream path, default six x86 requirements, and all eight selected latest-attempt job assessments remain unchanged: no failed, skipped or missing job waiver, no other fork/run/PR/pin admitted. The actual eight-job check passes locally with that exact proof; the overall upstream AMD-SNP failure remains retained/deferred, not green.Separate permanent repair
934b9c2a85b2c8f7762688cba9d76b5c44fe2691, parent19d29ebc7a6b4d825fb070317f4d534be5d59b3c, fixes the Phase 3 shared correctness binder's unconditional Bash dependency on self-hosted Windows. The binder now uses the native Python custom shell; its Python AST, environment, argument guards, JSON shape and step output are unchanged. Tests execute the actual script file with an empty PATH (no Bash or shell tools), preserving Linux/x86, Windows/x86 and ARM shapes and refusals. Dependent Windows ARM paths were inspected; remaining shared-workflow Bash steps are Linux/KVM-only. No Bash installation, runner/security/service/clock change, native rerun, build, pin change, lower-history rewrite or stack mutation was made.Permanent follow-up
8e0016bb65e28e9b6baaec2af1fc4c1337b08cce, parent934b9c2a85b2c8f7762688cba9d76b5c44fe2691, selects that shell through the declaredinputs.backend(whpselects Windows Python). The first automatic startup run at934b9c2afailed before creating any jobs because GitHub does not exposerunnercontext inshell; this failure is retained, not relabeled as a trial-gate pass. The follow-up restricts the test harness to the same allowed input context and keeps the Linux refusal fixture on its Linux backend. The 110 maintained policy/configuration tests passed again (104 passed, six platform skips), and both strict type views, lint and formatting passed.python scripts\check_test_changes.py --base 934b9c2a85b2c8f7762688cba9d76b5c44fe2691passed at the follow-up, reporting three existing test/wiring changes and no additions. Reverting the temporary policy commit must preserve both permanent portability commits.Final portability head
a0f5e4e336cda405eacc1142b8267563d292c5c1, parent8e0016bb65e28e9b6baaec2af1fc4c1337b08cce, replaces shell expressions entirely with mutually exclusive OS-guarded literalpython {0}/python3 {0}steps sharing one YAML-anchored script/environment; the existing artifact consumer reads the selected step's output. GitHub also rejectedinputsin the shell field in the retained second startup failure, before any jobs. At the final head, both binder Python ASTs and environments match the original exactly, the 110 targeted tests pass (104 passed/six platform skips), both strict type views and lint pass, and the final committed test/wiring inventory reports ten existing changes/no additions against8e0016bb(shared-script/env anchors, exactly-one-OS selection, and equivalent output coalescing). A read-only reverse-apply check proves that reverting only the temporary commit applies over all three permanent portability commits without undoing their repair; no actual revert was performed. No successful hosted trial-gate result is claimed from either failed startup run.Final-head CI admission blocker: bounded automatic-run observation at
a0f5e4e3found nopull_requestCI workflow run. PR #1516 is reported CONFLICTING / DIRTY (RESTmergeable=false), while its immediate base remains the exact Phase 2 commit. Only test-changepull_request_targetruns and the dynamic reviewer started at that head; neither is the NVX pin gate or runtime CI. There is no final-head hosted pin-gate/Windows-binder outcome to claim, and no new reusable-workflow startup failure was observed after the literal-shell correction. Stack/final-base compatibility remains for coordinator inspection; no pin choice, rebase, manual dispatch or rollback was performed. Final committed inventories passed: 14 existing/8 added follow-up changes versus preservedf775, and 611 existing/227 added for the separate all-stack comparison to localorigin/dev. These scopes do not replace the original Phase 3 inventory.The coordinator-directed rollback of only
19d29ebc7a6b4d825fb070317f4d534be5d59b3cis now complete in the new commit named above; it was not an automatic revert or cancellation of the immutablea0dispatch. The permanent Windows repair remains. The original source/product audit and qualification disclosures below describe the preservedf775snapshot, not fresh rollback-head native qualification. The exact temporary exception described above is historical and is no longer active.Follow-up validation using the isolated session Python environment:
python -m unittest scripts.test_nvx_tools.OpenVMMCiTests scripts.test_nvx_tools.CiConfigurationTests -qPASS, 110 selected/104 passed/6 unchanged platform skips;python -m ruff check scripts\nvx_tools\openvmm_ci.py scripts\test_nvx_tools.pyPASS;python -m ruff format --check scripts\nvx_tools\openvmm_ci.py scripts\test_nvx_tools.pyPASS;python -m pyright --pythonplatform Linux --pythonpath <isolated-python> scripts\nvx_tools\openvmm_ci.py scripts\test_nvx_tools.pyand the same command with--pythonplatform WindowsPASS, zero errors/warnings each;python scripts\check_test_layout.py,python scripts\check_markdown_links.pyandgit diff --checkPASS.python scripts\nvx.py check-openvmm-ci --revision 38016f6637f0c61fb55c9ee31033ceec428e346c --architecture aarch64changed from eight missing jobs before the trial to eight passed with the exact approved evidence; this is not an NVX CI success claim.Committed test-change inventories passed:
python scripts\check_test_changes.py --base f77577b4963b7b4ae8f8e73e5c0b6728cb8b19dfat temporary commit19d29ebcreports 2 existing helper changes/6 additions (API proof fixture support and fail-closed regression coverage);python scripts\check_test_changes.py --base 19d29ebc7a6b4d825fb070317f4d534be5d59b3cat permanent commit934b9c2areports 4 existing test/wiring changes/2 additions (execute the real custom-shell binder rather than stripping the Bash wrapper, remove the wrapper, preserve argument/output shapes and input refusal coverage).python scripts\check_test_changes.py --base origin/devalso passed as a committed inventory command. These follow-up counts are separate from the original Phase 3 215-existing/18-added inventory below;tests-modifiedremains required.Phase 3 of the authorized three-PR ARM64 stack, dependent on Phase 2 (#1510), which depends on Phase 1 (#1509). The lower PRs remain actual Drafts. This PR was initially Ready for review/CI, with a literal
[Draft]title at the user's request. It is now actual Draft with both lower PRs; neither state nor the title claims qualification is complete.Publish the existing audited commit
f77577b4963b7b4ae8f8e73e5c0b6728cb8b19dfunchanged: sole parentcb017cf8aee13e5e7276236d1b0ab328e1b52e7f, tree0339a9c7365fb8dd08ded08f5ff9d1ef8e4ff93e, exactly 64 changed paths against Phase 2. Publication adds no commits, source edits, rebases, merges, builds, or duplicate native runs.Extend the existing shared cold-runtime paths to native ARM64/WHP: architecture-bound guest/OpenVMM artifacts and native host/PE consumers; complete cold-case selection and CI argument, shard, duration, artifact and aggregate-status wiring; capability-scoped cold runtime-adapter tests; bounded same-provision scratch/RAM repetitions; and the actual ARM counter-warp probe. Correct DT CPU address-cell decoding, exact named MMIO/DT and GIC observations, layered Ubuntu control bootstrap, lifecycle time observations, and early quiet-console parity. Preserve x86 defaults/full snapshot tails, the existing 1000-iteration race, normal teardown, the exact required-status name, and the existing fork-headed workflow acceptance/rejection function. No trust exception or new every-job H5/H7 requirement is introduced.
Historical original f775 pin, provenance and landing constraints
The gitlink changes from
61cdd63978ea58ce045d159208ed4ca22f30cbb8to38016f6637f0c61fb55c9ee31033ceec428e346c, the published head of microsoft/openvmm#5045, available through upstream'srefs/pull/5045/headfromahuja-gautam/openvmm. The committed roadmap maps its ARM cold-boot foundationca913105dat position 252, following174da3a0dat position 251, and its shared CPU/time verifier38016f663at position 253. The verifier uses the sharedTimeAbiReport, preserves x86 output and rejects ARM fault hooks; it does not establish UTC discipline, snapshots, or a wider PA contract.The audit records source correspondence for the actual tested VMM
gca913105d.dirty, binary prefix2f776ab9, to the selected published38016source. This is not a clean38016rebuild or a cleanf775product rebuild; no provenance/product relabel is made. The source/product freeze, committed-blob inventory correspondence, and paired parent audit identify the exact tested inputs below.Required landing order remains OpenVMM first, with a reviewed fast-forward of
feature/microvm/devto the pin, before NVX integration. This publication does not perform that landing, grant a fork-CI trust exception, or assert it already happened. The audit reports all eight NVX-selected upstream jobs (six x86 plus native Windows ARM unit/VMM jobs) passed for38016; the overall upstream workflow is not claimed green. Its failed AMD-SNP receipt is retained and investigation is deferred per the user. Those eight jobs do not prove NVX runner/account admission.The historical publication snapshot of live NVX
devwasfc59c9065ceb2a939874123f9cdc856ad7784cf6, pinning OpenVMM90f87319d30d8c35bc3526f12fc7b6529c882cf0, divergent from the selected pin at174da3a0d236b4b01182a8d73a47598e4fcb1eca(four upstream versus two selected feature commits). The exact Phase 2 base does not prove final-base compatibility. That remains for coordinator inspection after publication, not permission to intake newer dev, replace the pin, or rewrite the stack.Validation
Current published-head execution and limits (24bd / b09)
Exact final descriptor:
C:\Users\gautamahuja\aci\phase3-refresh-20261113\evidence\final-phase3-refresh-qualification-handoff.json, SHA2564877476782c70b72978e7060b41f249429341c49246e32e1e5cb05d2c4d9f83a, preserving 568 raw command receipts and verified logs. Its status is requested execution complete; overall qualification NOT GREEN.$RefreshEvidencedenotes that evidence directory;$Pythonis the maintained interpreter already recorded below. These are recorded execution receipts, not new runs by this metadata session.2000NVX with cleanb09ARM VMM, bound bycurrent-native-runtime-inputs-refresh1.jsonSHA256f5223d14bb7060a4182f3a93f863256402148b2d7d14625d950591f3f6935a06b09guest_started=false) passed. These do not replace strict upstream GitHub CI or physical x64 execution qualification.Actual
24bdfull maintained 37-module Python suite:& $Python -m unittest scripts/test_performance.py scripts/test_nvx_tools.py scripts/test_kconfig.py scripts/test_kernel_config.py scripts/test_guest_binaries.py scripts/test_microvm_tests.py scripts/test_nightly_ci.py scripts/test_development_release.py scripts/test_crate_release.py scripts/test_host_telemetry.py scripts/test_time_abi.py scripts/test_tool_versions.py scripts/test_markdown_links.py scripts/test_module_diagrams.py scripts/test_nvx_rs_mutants.py scripts/test_nvx_rs_coverage.py scripts/test_nvx_rs_scenario_coverage.py scripts/test_setup_azure.py scripts/test_setup_azure_linux.py scripts/test_setup_azure_windows.py scripts/test_powershell_testing.py scripts/test_test_layout.py scripts/test_test_changes.py scripts/test_public_api.py scripts/test_scripts_coverage.py scripts/test_openvmm_git.py scripts/test_openvmm_pin.py scripts/test_openvmm_release.py scripts/test_local_gates.py scripts/test_pr_threads.py scripts/test_ci_triage.py scripts/test_issue_workspace.py scripts/test_host_worktree.py scripts/test_rebase_check.py scripts/test_nvx_hosts.py scripts/test_host_probe.py scripts/test_benchmark_series.py -vreturned exit 1, 2255 tests / 896.212 s / ZERO assertion failures. The precise outcome is 2147 complete-success methods, two partially successful methods, 3 Win1314 setup-error records in two unchanged Ubuntu methods, 107 skip records (104 whole-method + 3 subcases). Log SHA25628d3f3cbe761bc5c7924f333a9149a19d2a7813c168d1b2bbf2091e8782a68d4. Errors remain inBuildTests.test_ubuntu_customization_keeps_an_existing_wget_link(busybox/wget.real subcases) andBuildTests.test_ubuntu_customization_keeps_wget_installed_by_a_package; those bodies and failing_ensure_symlinkhelper retain baseline ASTs, not a claim that their entire files match baseline. No added skip/elevation/assertion weakening fixes those errors.The exact six-case regression command
& $Python -m unittest -v test_nvx_tools.CiConfigurationTests.test_ci_runs_the_steps_after_the_default_microvm_suite_beside_it test_microvm_tests.MicrovmTests.test_runner_flushes_each_progress_line_before_its_scenario test_microvm_tests.MicrovmTests.test_runner_reports_every_scenario_execution test_microvm_tests.MicrovmTests.test_only_arm_expands_boot_and_managed_processor_selection test_microvm_tests.MicrovmTests.test_runner_dispatches_selected_scenarios_once test_nvx_tools.CiConfigurationTests.test_ci_jobs_download_exactly_the_guest_artifacts_that_they_readPASS, covering all three corrected methods plus existing explicit ARM/default-x86 and artifact-binding fixtures. This owner's prior 58 selectors plus the three corrections also PASS: 61 / zero skips; raw red-at-2000 and green-at-corrected-source evidence is retained.& $Python -m compileall -q scripts,& $Python -m ruff check scripts .github\specula,& $Python -m ruff format --check scripts .github\specula, both& $Python -m pyright --pythonpath $Python --pythonplatform Linux/Windows,& $Python scripts\nvx.py verify, test-layout, links, kernel inventories and recorded format/PowerShell syntax checks PASS; full strict type views have zero errors/warnings. Exact commands/results are preserved in the final descriptor and linked static/source ledgers. No complete 37-module duplicate rerun occurred in this owner.Current host doctor is NOT green:
doctor --backend whp --architecture aarch64 --checks H1 H2 H3 H4 H5 H6 H7used the full 13 samples / 10 s schedule and returned exit 1. H1/H2/H3/H5/H6 PASS; H4 FAIL (host UTC free-running,GetSystemTimeAdjustmentPrecise disabled=1) and H7 FAIL (W32Time not started,0x80070426). Current H5: 66 pairs, 26 ns offset / 156 ns uncertainty, unchanged 1000 ns bound. Current H6: 128 MiB, 8 VP five rounds with idle gaps 0.1/1/5/1 s plus one VP; zero backward, 52 ns offset / 391 ns uncertainty. Older host 52/182, guest 52/313 and previous full matrices below remain their historical source-pair evidence. No host service, clock, security, threshold or CI waiver was applied.Explicit proof limits: normal-product ARM restore refusal is NOT PROVED (retained pre-manifest empty-fixture failure; source/unit guard is separate). Deterministic active-VP0/AP-overlap product proof is NOT PROVED. The 37-callback MMIO unit test is not raw-product concurrency proof. No ARM restore/snapshot/Phase4 qualification is claimed. Exact
b09trusted upstream CI and actual NVX native-ARM CI are not qualified; normal policy rejects untrusted/missing evidence with no active trial allowance. Old strict6036eight-MISSING results and failed upstream AMD-SNP receipts remain historical/unwaived, not a currentb09gate run or overall CI-green claim. Final source/conflict/architecture/consumer review found no required redesign or obsolete primitive, but that is not runtime/CI proof beyond the receipts above.Historical original-source validation (all results below retain their original scope)
Previously executed, audited evidence; not rerun by this publication session. The authoritative command definitions were inspected in
.github/actions/check-quality/action.yml,.github/actions/validate-nvx/action.yml, and.github/actions/check-nvx-rs/action.yml. This is not a claim that every action or the full suite passed.For the command spellings below,
$Pythonis the recorded interpreterC:\Users\gautamahuja\aci\phase1-20261106\nvx\target\phase1-python\Scripts\python.exe,$EvidenceisC:\Users\gautamahuja\aci\phase3-20261111\evidence, and the original source working directory isC:\Users\gautamahuja\aci\phase3-20261111\nvx. These are receipt locations, not paths expected to exist on a CI runner.Source and CLI checks
& $Python -m ruff check scripts .github\specula& $Python -m ruff format --check scripts .github\specula& $Python -m compileall -q .github\specula& $Python -m pyright --pythonplatform Linux --pythonpath $PythonPYTHONPATHunset.& $Python -m pyright --pythonplatform Windows --pythonpath $PythonPYTHONPATHunset. Earlier failed receipts retained.& $Python scripts\nvx.py verify& $Python scripts\check_markdown_links.py& $Python scripts\check_test_layout.py& $Python scripts\nvx.py check-kernel-config& $Python scripts\nvx.py check-kernel-config --architecture aarch64& $Python scripts\check_test_changes.py --base origin/dev& $Python scripts\check_test_changes.py --base cb017cf8aee13e5e7276236d1b0ab328e1b52e7f& $Python scripts\test_adversarial.py -vgit --no-pager diff --check HEADThe exact final maintained Python suite command was:
NOT GREEN, exit 1: 2230 selected methods, 2123 executed, 2121 passed, 107 skipped, three WinError 1314 error records in two AST-unchanged Ubuntu methods (
BuildTests.test_ubuntu_customization_keeps_an_existing_wget_linkandBuildTests.test_ubuntu_customization_keeps_wget_installed_by_a_package), zero assertion failures. No symlink-privilege workaround or test weakening was applied. Six maintained PowerShell setup files also parsed without errors; no publication-time rerun is claimed.Rust and native cold evidence
The initial maintained Rust batch was 22/24 gates passing, not fully green. The unchanged Windows host-adapter library outcome was 57/58 with a WinError 1314 error-kind mismatch (
UncategorizedversusPermissionDenied); the other nonzero gate was intentional ARM-host rejection of the x86 release-catalog bundle. Independent adapter integrations passed. Explicit x64 bundling withNVX_BUNDLEunset was compile-only, not fresh native x86 execution.cargo +1.97.1 fmt --all --checkcargo +1.97.1 test --locked --target aarch64-unknown-linux-musl --no-default-features --features guest --lib time_probe -- --nocapturecargo +1.97.1 clippy --locked --all-targets --target aarch64-unknown-linux-musl --no-default-features --features guest -- -D warningscargo +1.97.1 doc --locked --no-deps --target aarch64-unknown-linux-musl --all-featuresNVX_BUNDLE=skipandRUSTDOCFLAGS=-D warnings; the separate guest-only Rustdoc attempt failed unchanged host links and remains disclosed.cargo +1.89.0 check --locked --target aarch64-unknown-linux-musl --no-default-features --features guest --all-targetscargo +1.97.1 clippy --locked --target x86_64-unknown-linux-musl --no-default-features --features guest --all-targets -- -D warningspython3 scripts/nvx.py test-guest-binaries --architecture aarch64 --binaries /out/guest-binaries& $Python "$Evidence\native_guest_units.py" coldfix1-ubuntu ubuntu normal-devfs current-native-runtime-inputs-coldfix1.json& $Python "$Evidence\native_console_proof.py"& $Python "$Evidence\native_matrix_restart.py" cold coldfix1-full current-native-runtime-inputs-coldfix1.json& $Python "$Evidence\native_matrix_restart.py" adapter coldfix1 current-native-runtime-inputs-coldfix1.json& $Python "$Evidence\native_matrix_restart.py" repeat coldfix1 current-native-runtime-inputs-coldfix1.json& $Python "$Evidence\final_native_accounting.py"Historical original-source runtime results are bound to pair SHA256
d7d0cb48294230956b619903258f873532388a9be907d0ade6bff123c4709532, the new kernel prefixc0101427, native static-CRT CLI prefix00158fd2, and actual VMM prefix2f776ab9(gca913105d.dirty) with the disclosed source correspondence. Historical H6 passed 8 VP twice plus 1 VP once at the unchanged 1000 ns bound: maximum offset 52 ns, uncertainty 313 ns, backward step 0.Original Docker namespace/mount EPERM failures, minimal-Alpine userland failures, missing-helper H6 failures (guest exit 97/127), and the original cold outcomes (28 PASS/17 FAIL, then 14 PASS/3 FAIL) remain in the evidence. The corrected complete batch removes no applicable ARM cells and does not weaken assertions or bounds.
Not passed, not run, or left to CI
H4/H7 and full UTC/host qualification remain NOT PASS (published-correction predicate and service/source proof). H1/H2/H3/H5/H6 passed in the scoped admission evidence, but the complete host command returned exit 1.
disabled=TRUEalone is not independent proof of unsynchronized UTC. No service/security/clock change is authorized by publication.Actual NVX native ARM pool admission and NetworkService admission (both SIDs) remain unproved; inventory access returned 403. The existing fork-headed workflow rejection is unchanged. Upstream selected-job evidence does not establish NVX CI admission, and this PR does not approve any exception.
No ARM snapshots/restore, optional adversarial campaign, PA widening, fresh native x86/KVM/MSHV qualification, Linux-only public-API/macOS gate success, performance result, or full-suite/overall-upstream-CI success is claimed. Leave applicable cross-platform/action gates and real CI runner admission to CI. The original Ready publication was the intended PR-event trigger at that time; the subsequent immutable manual trial is disclosed above. All three PRs are now actual Draft. No new manual dispatch, merge, queue admission, or auto-merge is requested.
Existing-test and CI-wiring change explanations
Apply
tests-modified. The maintained committed comparison is against Phase 2cb017cf8aee13e5e7276236d1b0ab328e1b52e7f, not the dirty precommit tree or an all-stack comparison to dev: 215 existing test/wiring changes, 18 added items. The frozen per-file explanation inventory covers these 29 changed files. Enclosing ISA cfg changes also affect the unchanged x86time_probe/clock/tests.rsandtime_probe/latency/tests.rs; both remain x86-only. The shared parent now admits aarch64 without admitting those x86-only modules there. Counts include inherited enclosing-cfg effects, not 215 newly written test methods.nvx-rs/src/guest/time_probe/format/tests.rsnvx-rs/src/guest/time_probe/support/tests.rsnvx-rs/src/guest/time_probe/tests.rsnvx-rs/src/guest/time_probe/warp/tests.rsnvx-rs/tests/guest_tools.rsnvx-rs/tests/openvmm_cpu_profiles.rsscripts/test_guest_binaries.pyscripts/test_microvm_tests.pyscripts/test_nvx_tools.pyscripts/test_time_abi.pyscripts/test_tool_versions.pyscripts/test-layout-allowlist.jsonscripts/nvx_tools/benchmark_scripts/smp-probe.sh.inscripts/nvx_tools/microvm_test_scripts/filesystem-shares.shscripts/nvx_tools/microvm_test_scripts/interrupt-map.sh.inscripts/nvx_tools/microvm_test_scripts/lifecycle.shscripts/nvx_tools/microvm_test_scripts/sandbox-blocks.shscripts/nvx_tools/microvm_test_scripts/virtio-net.shscripts/nvx_tools/microvm_test_scripts/virtio-mmio.shscripts/nvx_tools/microvm_tests.pyscripts/nvx_tools/runtime_adapter_tests.pyscripts/nvx_tools/sandbox_lifecycle_tests.pyscripts/nvx_tools/sandbox_ram_upper_tests.py.github/actions/build-guest-artifacts/action.yml.github/actions/build-openvmm/action.yml.github/actions/validate-runner/action.yml.github/workflows/build-openvmm-binary.yml.github/workflows/run-nvx-microvm-tests.yml.github/workflows/ci.ymlProduction-file test-hook removals in guest status/clock/parse/scheduler support account for corresponding committed-report entries; they remove test-only gating from real shared support rather than adding inline tests or widening the allowlist.
Immutable evidence identity
The original local evidence is retained under
$Evidence; these hashes identify audit artifacts, not downloadable GitHub attachments:phase3-final-pushed-publication-handoff.json72b68e0d627306baadccd363d8259a0a70a0d0f73abec05cfe1d4f8fddc87a12phase3-committed-inventory-correspondence.json8e34a98a864d7eabb8bf05ce50140c9279261b708e86a0d7420a36759bfaa951phase3-committed-test-changes.md01a7eb19844aea9b8e4a221639d1114b333cd522f75f11ff5420c3d35ac8e9e7phase3-test-change-explanations.json4b0eebd4c95f264042b8d000b26db3a96fc5e0d5e280e4af8db920749f897630phase3-final-source-product-gate-handoff.jsone0c2401554ca69a159b585de430f1c6c9f7083b49b16017c98c3e40e07e2fbf85d9c179097ee7fc64b3fa0f58f5f9ca5f96805db5be1e5658e7b565ffae40f8e60a73b91ff6c9ba1318ab1e06930a773f8d4c635f96f8164c6819da0ff41e46cThe final committed test report, rather than an earlier precommit zero-change result, is authoritative. Original failures and protected logs remain preserved.