Premium VM proof runbook

Use a disposable Windows clone for the initial same-console streaming proof. Keep the sealed source stopped, the existing Windows workspace on VNC, and Linux Selkies available. These are diagnostic steps, not an enabled premium profile.

Capacity and recovery gate

1. Select an eligible non-control-plane worker. Inspect kubectl top nodes, host MemAvailable, swap/load, pod reservations, quota and physical free disk. Reserve 4 vCPU / 8Gi guest RAM plus launcher and host/workload headroom for the software reference. A 2-vCPU/4Gi exploratory run must be explicitly recorded as below-reference; it cannot certify the reference performance target.
2. Verify both root and backend-state PV node affinity; local-path roots cannot be moved simply by changing a node selector. Reuse only a designated disposable clone with preserved root/EFI/TPM identity, or make a fresh isolated import.
3. Run scripts/windows-vm-proof.py preflight --help and its read-only preflight for the selected worker. Confirm current version-specific KubeVirt/CDI schema, import credentials/CA availability, artifact digest and certificate validity.
4. Record the initial VM spec, root/state UIDs and display topology. Establish direct VNC recovery before changing hardware or installing a driver. Bound each trial and stop on host pressure, security bypass or unrecoverable console.

P0 evidence sequence

Record requested and actual sizes, p50/p95/p99 timings, delivered vs rendered FPS, queue high-water marks, memory and audio discontinuities. WSS encoded GOP abandonment must flush to a fresh IDR; do not drop arbitrary reference frames. Keep input private to the disposable proof and do not expose an unauthenticated network agent. Enrollment and control fencing precede product remote-input access.

The proof decision record must freeze the language/media build, driver package, software workload envelope, native/browser codecs and transport, or explicitly record a no-go and its prerequisite. Only then start the agent product project.

Native Windows diagnostic

For a bounded actual draw/readback comparison after driver qualification, build the separate off-screen probe:

x86_64-w64-mingw32-g++ -std=c++17 -O2 -Wall -Wextra -Werror -static \
  scripts/windows-d3d11-render-probe.cpp -o windows-d3d11-render-probe.exe \
  -ld3d11 -ldxgi -ld3dcompiler -lole32 -luuid

Run windows-d3d11-render-probe.exe VENDOR_HEX DEVICE_HEX SECONDS_1_TO_10 under an external 60-second process watchdog. It explicitly selects a matching non-software PCI adapter, refuses fallback, and compares the same 1920×1080 pixel-shader workload against WARP. Three warmup frames are excluded. Every sample draws a fullscreen triangle, synchronizes via staging readback and checks nonuniform opaque content; no pixels are saved. JSON identifies each actual adapter/LUID and reports completed frames, wall-time p50/p95 and throughput. Readback/synchronization is included, so these are not application, capture or streaming FPS, GPU timestamp measurements or an encode proof. Session 0 is adequate for this off-screen workload, but interactive capture remains separate.

scripts/windows-premium-probe.cpp uses inbox Windows APIs, without a media runtime install. Build with MinGW-w64 (or a reviewed equivalent Windows toolchain):

x86_64-w64-mingw32-g++ -std=c++17 -O2 -Wall -Wextra -Werror -static \
  scripts/windows-premium-probe.cpp -o windows-premium-probe.exe \
  -ld3d11 -ldxgi -lmfplat -lmfuuid -lole32 -luuid -lpropsys -lversion

Transfer it to the disposable guest through the isolated proof tooling and run from its logged-in console session, saving JSON in a private evidence folder:

.\windows-premium-probe.exe | Set-Content -Encoding UTF8 inventory.json
.\windows-premium-probe.exe --capture-seconds 5 | Set-Content -Encoding UTF8 capture.json

Default inventory enumerates DXGI adapter identity/VRAM/output modes, D3D11 device creation, per-adapter D3D12 creation, Vulkan loader/instance/ physical-device enumeration, Intel OpenGL ICD file presence, separate WARP creation, actual active MMDevice render endpoints, WASAPI loopback capture trial (endpoint → mix format → 2s packet capture with silence accounting — silence is recorded, not hidden), and non-hardware Media Foundation H.264 candidates. Optional Desktop Duplication tries frames for a shared 1–10 second capture budget; it does not copy/save pixels, change modes, inject input, install drivers or open network sockets. Apply an external 30-second process watchdog for driver/API hangs. A run in Session 0 can diagnose enumeration but cannot certify console capture.

Acquired-frame counts are availability diagnostics, not delivered stream FPS. An idle screen legitimately yields few frames. Failed HRESULTs are retained in JSON even when process exit is zero: inspect each result. Encoder enumeration does not establish activation/encoding, and endpoint presence does not establish loopback/Opus/playback. WGC, encoding timing, audio loopback, mode changes and viewer/transport proof remain separate trials. A diagnostic written in C++ does not select the production agent’s implementation language.

G0 read-only physical inventory

On each candidate Linux worker, run locally or using an already trusted SSH host:

python3 scripts/vm-gpu-inventory.py
python3 scripts/vm-gpu-inventory.py --ssh [email protected]
kubectl get nodes -o custom-columns='NAME:.metadata.name,ALLOCATABLE:.status.allocatable'
kubectl get kubevirt -A -o yaml

The script reads graphics PCI IDs, current driver, IOMMU members, reset methods, boot-VGA/output state, mediated/SR-IOV exposure and capacity. It never changes device binding or resets hardware. Retain live output privately.

For a selected device additionally record vendor-supported host/guest driver matrix, VRAM, UEFI ROM, graphics API/encoder capabilities, cost/licensing and host console/workload ownership. Require a VM-capable advertised resource and operator-approved KubeVirt allowlist; a container /gpu resource is not proof of VFIO/vGPU compatibility. Coordinate worker maintenance before allocation.

First prove signed Windows hardware-adapter D3D11 rendering versus WARP with adapter LUID/vendor/device/VRAM and render-engine utilization. Probe D3D12, OpenGL, Vulkan and hardware H.264 separately. Verify capture of that output in native and browser viewers, retained VGA recovery and repeated reset/reallocation. If no usable spare device exists, record physical proof blocked and the required hardware/placement; continue the software streaming work independently.

Integrated Intel GPU candidates

An integrated GPU can be a research candidate even when no VM resource is advertised yet. Check privileged DRM client/host-console ownership, IOMMU groups, kernel GVT-g module and current i915.enable_gvt, mediated profile exposure and exact guest-driver support. No mediated directory while GVT-g is disabled is not a hardware impossibility verdict. Do not assume older Intel GPUs expose modern SR-IOV VFs.

Full IGD passthrough and GVT-g sharing have different firmware/device paths. Consult documentation for the running QEMU, not an older UPT recipe: Intel OpRegion/stolen-memory and guest firmware handling can require integration not provided by generic PCI allowlisting. Retain the current Secure Boot/OVMF and VGA recovery contract. GVT-g’s historical Windows validation is not certification of a current Windows 11 build; review its maintenance status before production. Host GVT enabling/reprobe/reboot or VFIO binding requires coordinated worker maintenance and a reversible host configuration. Do not change a live control-plane GPU or disable firmware/driver security to make the test pass.