Latest user Vulkan crash evidence — 2026-09-16

Evidence collection was read-only apart from new evidence copies; no player launch, GPU stress, settings changes, or privilege changes. A subsequent offline QA log-scanner correction is described below. Evidence copies were captured during the 22:45–22:52 CEST investigation. The 2249- prefix is a snapshot label; original timestamps and hashes are in snapshot-index.json.

Latest incident

• User reports the Vulkan build stopped after 1:45, just before the stage 1-2 finish.
• The captured production Player.log is 40,534 bytes, last written 22:38:02.1632004 CEST, SHA256 13FCB74C47E590066765DA50A0600BED230CE4F54543F55AC0F0656652AB05D1. It identifies the actual Vulkan build, engine 6000.3.24f1, alps-02, Rain weather and all eight riders. It ends after race ready, camera restoration and grass/habitat draw diagnostics, with no device-lost or crash footer. The first directory metadata observation reported 0 bytes /22:37:41; the subsequent copy and repeat read establish that the preserved log is populated, so the initial observation must not be used as the final evidence state. The previous player log remains the already-known 17:49 D3D11 session. No new Unity native crash directory exists after the 17:20 crash.
• Windows System records nvlddmkm 14 and 153 at 22:39:29.7629489 CEST. Event 153 XML explicitly contains UCodeReset TDR occurred on GPUID:100. nvidia-event-xml.json preserves record IDs, full XML and event properties, including messages which Windows did not render through the usual Message property.
• WER records a LiveKernelEvent 117 referring to C:\WINDOWS\LiveKernelReports\WATCHDOG\WATCHDOG-20260916-2239.dmp, first seen here at 22:39:36.9469409; a LiveKernelEvent 193 refers to WATCHDOG-20260916-2240.dmp, recorded 22:40:05.1696195. Later WER entries repeat these reports; they are not additional independent crashes. See latest-watchdog-reports.json.
• Latest boot is 22:40:30 CEST. Kernel-Power event 41 records an unclean restart with BugcheckCode=0; its complete XML is preserved in kernel-power-xml.json. Event 6008 records an unexpected shutdown time of 22:39:18. That retrospective time is not a precise GPU-fault timestamp.
• These establish a new NVIDIA GPU timeout/reset followed by an unclean system restart in the time window of the user report. They do not identify the originating draw, shader, race position or driver defect. The 1:45 race time and finish location remain the user's direct observation because no continuous race telemetry or crash details persisted.

Distinguish the older full-system freeze

The repeated WER BlueScreen / 133 messages refer to the same older 091626-13375-01.dmp and report ID 907bd142-1fd8-43c4-b94f-fd23cf9c6022, already recorded after the 22:26 reboot. They are not proof of a second bugcheck 0x133 during the latest 22:39 incident.

The earlier automated Vulkan test ../../20260916-200814-graphics-stability was started at 20:08 and froze after its last saved report at 142.120224 seconds, 5,593 rendered frames, 5 completed visits, phase alps-03 / countdown. It has no managed or native graphics error in the final player-log lines. The recovered outcome in ../../20260916-crash-followup/post-reboot/recovered-outcome.json marks it failed-system-freeze; its original invocation status remains stale running because cleanup never executed.

Engine, API, GPU and memory evidence

• Existing Vulkan build, its automated player log and the captured newest manual player log: Unity 6000.3.24f1, changeset 4e7b9b5b6244, threaded Vulkan, RTX 4060 Laptop GPU, Vulkan driver string Vulkan 1.1.0 [616.368.0 (0x9a170000)] in the automated report. The actual newest manual log independently confirms this engine, build path, Vulkan device selection and NVIDIA driver.
• Fresh read-only Win32_VideoController inventory: NVIDIA 32.0.16.1692 / 616.92, date 2026-09-04; Intel UHD 31.0.101.4255, date 2023-03-16.
• The production profile snapshot selects alps-02, quality=0, motionBlur=false, weatherDisabled=false. This is useful context, but saved preferences cannot substitute for actual runtime telemetry.
• The older 20:08 Vulkan freeze's last GPU sample was 2,119 MiB / 8,188 MiB, 69 °C, 46.95 W, and the last system sample still had about 43.9 GB physically available. Those observations do not support ordinary memory exhaustion for that run. No corresponding telemetry exists for the newest manual run.

Existing test limits and hypotheses

The graphics stability probe repeatedly loads stages but drives the road for only about five seconds per visit before roadside/reset/pause/home checks. It does not exercise 1:45 continuous driving or the finish transition. Its total elapsed time must not be presented as uninterrupted race coverage.

Earlier D3D11/D3D12/Vulkan failures and the new Vulkan TDR mean switching API alone is not a demonstrated fix. The saved 6000.6/HDRP 17.6 comparison also failed its full forest run, while the minimal HDRP scene passed; an engine upgrade alone is not a demonstrated fix. Prior isolated passes with scenery disabled or the billboard renderer suppressed remain useful clues, but the native HDRP/Unlit tree replacement also reset and full-production effects differed in some comparisons. Neither the custom billboard shader alone nor hardware failure has been established as the cause. Other games being stable is consistent with a workload-specific fault.

Dump availability

The two newest WATCHDOG paths returned PathNotFound. Their exact WER queue directories exist but enumerating their contents returned Access denied. The older 0x133 minidump also returned Access denied. Exact attempts are in dump-access-checks.json and latest-wer-access-checks.json. No access-control change, escalation, dump copy or debugger launch was attempted. A readable dump remains the direct path to deeper kernel attribution; existing event records already establish the new timeout.

Offline log guard correction

After the read-only evidence collection, Tools/qa-log-check.ps1 was extended to reject Vulkan device-lost and host/device allocation failure symbols, explicit Vulkan validation errors, VUID/UNASSIGNED validation-error records, and explicit VK_LAYER error severity. Informational layer/extension availability, enablement, warning severity and VUID identifiers without error severity remain nonfatal. This closes a test-result interpretation gap; it does not fix GPU execution or establish that a silently truncated log is healthy overall. The wrapper must still require its completed report and normal process outcome.

Tools/qa-log-check-test.ps1 passed 45/45 offline cases with the six original historical log fixtures unchanged. The saved real DirectX GBV fixture still contains exactly 83 ID 961 errors. The result is in ../qa-log-check-regression.json. No Unity/player/GPU execution was performed for this regression.