docs(overhaul): record cathedral GPU recurrence and isolated checks
This commit is contained in:
Erik
parent
f4bf5ffab8
commit
620540aba2
4 changed files with 217 additions and 1 deletions
165
docs/research/2026-09-01-overhaul/477-cathedral-recurrence.md
Normal file
165
docs/research/2026-09-01-overhaul/477-cathedral-recurrence.md
Normal file
|
|
@ -0,0 +1,165 @@
|
|||
# #477 — cathedral AMD recurrence and isolated diagnostics
|
||||
|
||||
2026-09-05. **Owner stability gate FAIL; no production repair in this record.**
|
||||
The owner first said the visible fixes looked good and withdrew confidence in
|
||||
the Rynthid placement report, then reported another AMD crash in the cathedral.
|
||||
Do not treat the earlier visual acceptance as crash/stability acceptance.
|
||||
FPS/#481/C2 performance work remains explicitly deferred; no main merge.
|
||||
|
||||
## Owner crash: confirmed, initiating GPU fault still unknown
|
||||
|
||||
Ordinary visible owner run `logs/owner-gate-20260905-162519-478-final/`,
|
||||
PID7852. At16:41:20+02 Windows .NET Runtime1026 recorded unhandled
|
||||
`ErrorDeviceLost` from `vkQueueSubmit2 (abandoned frame timeline signal)`.
|
||||
Stderr also records secondary failed teardown waiting for timeline178412.
|
||||
Windows Error Reporting1001 at16:41:28 and16:41:57 records LiveKernelEvent141,
|
||||
bucket `LKD_0x141_Tdr:C_AppFault_IMAGE_amdkmdag.sys-PF:1-HWS:1-SCH-PS:95d63b9d-CS:6708cc79-GS:62c333ab`,
|
||||
report b3e69f70-71fc-418c-b9e8-686a6bdaf552. WER names
|
||||
`C:/WINDOWS/LiveKernelReports/WATCHDOG/WATCHDOG-20260905-1641.dmp`; its content
|
||||
has NOT been opened or analyzed here. Two WER entries are one report, not two
|
||||
separate GPU resets. RX9070XT, Vulkan1.4.349/driver2.0.395, 1280x720,
|
||||
High/4xMSAA; ordinary profile without requested validation.
|
||||
|
||||
Generation12 at F4180104 completed and became visible before the crash.
|
||||
The observed arrival was (36.9,48.7,169.805); last input log includes turning,
|
||||
combat toggle and forward movement. The exact crash pose/trajectory is unknown.
|
||||
Subsequent test login was saved at F418010F; arrival is not an exact crash pose.
|
||||
Retail PID32728 also existed during the owner session; the owner closed it
|
||||
before the isolated tests. Its coexistence is context, not a demonstrated cause.
|
||||
|
||||
Same failure family as Neftet, not proven same initiating fault. Source still
|
||||
collapses fatal acquire and ordinary unavailability into false in
|
||||
`VulkanGraphicsContext.SwapchainBackbuffer.TryAcquire`; the empty timeline
|
||||
submit therefore may only expose an earlier device loss. No evidence connects
|
||||
this crash to C1a's three-versus-four-cell membership question, which remains
|
||||
unresolved. No geometry golden, AP159 repair, driver setting, TDR setting,
|
||||
shader, synchronization rule, frame pacing, or exception suppression changed.
|
||||
|
||||
Owner logs SHA256: stdout
|
||||
2B70E77CCDF9D4DFBB482143F49127216FF806A105984A30812A690B235C7A79;
|
||||
stderr BC3C98C31F0B95606BC717ADDD202BB095F57C7F26822C82AAEFF8400E742FD4.
|
||||
|
||||
## Exact unchanged binary and diagnostic scope
|
||||
|
||||
Both tests use Release built at b2bdb09ccbd515f180b4fb3b7159de4d3913e68d.
|
||||
Its locked restore/build is green0W/0E and literal-hermetic17028/0/0,
|
||||
`artifacts/s5-final-hermetic-expectations/release-gate-summary.json`.
|
||||
Source HEADf4bf5ffab was clean at launch and differs only in docs since that
|
||||
build. App/Core DLL hashes were checked before each launch:
|
||||
|
||||
- App5692DF7F1495D235351CF87C8741781260B4243E6922476F7A23B8F2B025F305.
|
||||
- CoreD4F15FEB44F99DBC257DF5911D302E82FC7FAB9BB549590B06FFE28E8685BF9B.
|
||||
|
||||
Same `logs/s5-c5-session-v10.json` and recipe10 package, High/MSAA4,
|
||||
1280x720. ACE UDP9000PID13340 checked up; both client names checked absent
|
||||
before each launch. One run per background command, no builds/tests beside
|
||||
graphics. Automation enables checkpoint/screenshot machinery and borderless
|
||||
exact-framebuffer mode; these are not identical to the ordinary owner's
|
||||
window/session, nor are validation runs FPS measurements.
|
||||
|
||||
The corrected81-command route is archived as `route.txt` in each successful
|
||||
run. SHA25692F5797C23455D569630C557A6237D8FFADC0636F26641D1B2FFE24BB7FEA9C7.
|
||||
Lead checked all verb and InputAction names against production source,
|
||||
balanced every held input, and checked the terminal close-client. Workload:
|
||||
saved cathedral login, F4180104 arrival, two forward/back and turning input
|
||||
sets/two combat toggles, F4180113 stairwell, eight -20 mouse-look steps,
|
||||
return tilt/right camera turn, and30seconds settled. Input pairs are workload,
|
||||
not proof of exact world displacement or a particular combat result.
|
||||
|
||||
## Discarded first attempt — test-script failure, not a GPU reproduction
|
||||
|
||||
`logs/selfgate-20260905-164528-477-cathedral-validation-r1-b2bdb09cc/`,
|
||||
PID7444, aborted at line30: unknown `MovementBackward` (correct enum is
|
||||
`MovementBackup`). Only login and arrival screenshots were captured. This is
|
||||
**FAIL/incomplete**, never a passing route. Original route is preserved there;
|
||||
SHA256A05CB6933ABD0C06467A639593AB6AA857753467F3A5106373B8BB5BB59D8882.
|
||||
CloseMainWindow returned false; computer-use found no targetable hidden window.
|
||||
After explicit owner authorization, only identity-checked PID7444 was forcibly
|
||||
stopped at16:50:43; client exit-1. Nothing else was killed. Waited over three
|
||||
minutes for ACE's stale session before retry. No production fix round consumed.
|
||||
|
||||
## Corrected core + synchronization validation — PASS, narrow/provisional
|
||||
|
||||
`logs/selfgate-20260905-165355-477-cathedral-validation-r2-b2bdb09cc/`,
|
||||
PID7976, start16:53:55.855+02, settled checkpoint16:55:41.434+02.
|
||||
Process-only VK_INSTANCE_LAYERS=VK_LAYER_KHRONOS_validation,
|
||||
VK_LAYER_VALIDATE_SYNC=1, VK_LAYER_GPUAV_ENABLE=0, ACDREAM_DEVTOOLS=1.
|
||||
Lead observed loaded C:/VulkanSDK/1.4.350.0/Bin/VkLayer_khronos_validation.dll.
|
||||
All81commands completed; client/runner exit0, graceful logout confirmed,
|
||||
both clients absent afterward. Zero matches across both logs for SYNC-HAZARD,
|
||||
Validation Error, VUID, ErrorDeviceLost, unhandled exception, timeout, invalid
|
||||
input or abandoned teardown. Stderr contains one known unsupported VFX emitter
|
||||
warning. Windows Application1000/1001/1026 matching AcDream/LiveKernelEvent/
|
||||
WATCHDOG in16:53:55–16:56:10: zero.
|
||||
|
||||
All six checkpoints: completed/visible true, wait cue false, invariant failures0,
|
||||
transit idle. Final staged uploads/live teardowns/landblock retirement/warmup0.
|
||||
All seven1280x720 PNGs were opened by the lead: narrow **PASS, provisional**
|
||||
for coherent scene/current screenshot and no stuck reveal. Not an independent
|
||||
retail comparison, complete cathedral visibility proof, or crash-recurrence
|
||||
guarantee. PNG paths below are relative to this run's `artifacts/screenshots/`:
|
||||
|
||||
| PNG | SHA256 |
|
||||
| --- | --- |
|
||||
|477-cathedral-00-login.png|D7A2677CE1683B9B77041C3EA05D12CE9D019DF5FB56D359034198C668C334E7|
|
||||
|477-cathedral-01-arrival.png|79F74A4228776ADB5BCD7389CD7EFE069E5A73AE8D099DF1C2DA58849EBB8097|
|
||||
|477-cathedral-02-movement.png|91021D332DE8C9CE774B00900EE2C66F65D0E713D45162E985C2A47BBB6BEB31|
|
||||
|477-cathedral-03-stairwell.png|BC32D524A699DE7055C1F3277C2862E47512A002C1845530BD81FB1535ACDBA2|
|
||||
|477-cathedral-04-stairwell-tilt.png|A8D1F31F94FA5D6BF9C80D13F5C8E29DCA1C4ABB83727F6466DB03DB4227BD4D|
|
||||
|477-cathedral-05-stairwell-turn.png|47510A14361801B3F1F6B255AAAC67E033312602C2D49D2B1CD7F96D33E6A83C|
|
||||
|477-cathedral-06-settled.png|FFA32856E18A21ECB01CA768A1F8EADEB737EBDE9E769F8A4BD5E84EBC16E8C6|
|
||||
|
||||
stdout C8428F1DAFA603626118EF1F5A7A78ECC35D452642F7115C3C60F7BB3B84507D;
|
||||
stderr4D40C9BEEA7C8FFDB7ED9C2AF70273714FF592E2BED2320E57D7FD87A56D2F87.
|
||||
|
||||
## GPU-assisted follow-up
|
||||
|
||||
Same corrected route, separate process, aimed at shader-memory errors rather
|
||||
than repeating core validation. Process-only GPUAV_ENABLE=1 with VALIDATE_CORE=0
|
||||
and VALIDATE_SYNC=0, using the full VK_LAYER_ prefixes, DEVTOOLS=1 and the same
|
||||
explicit layer. The environment mapping and separate-mode recommendation are
|
||||
from [Khronos](https://github.com/KhronosGroup/Vulkan-ValidationLayers/blob/main/docs/updating_from_VK_EXT_validation_features.md)
|
||||
and [LunarG](https://vulkan.lunarg.com/doc/view/latest/windows/gpu_validation.html).
|
||||
GPU-assisted instrumentation changes execution and overhead; it is neither
|
||||
a performance comparison nor a guarantee of every shader path being checked.
|
||||
`logs/selfgate-20260905-165627-477-cathedral-gpuav-r3-b2bdb09cc/`, PID2156,
|
||||
start16:56:27.991+02, settled checkpoint16:59:01.266+02. Lead observed the
|
||||
same loaded validation DLL. No reported GPU-AV initialization/disable warnings;
|
||||
this is not a negative-control proof of every instrumented shader path.
|
||||
All81commands completed, client/runner exit0, graceful logout confirmed,
|
||||
both clients absent afterward. Zero fault matches across both logs using the
|
||||
same checks as r2. Stderr contains one known unsupported VFX-emitter warning.
|
||||
Windows Application1000/1001/1026 matching AcDream/LiveKernelEvent/WATCHDOG
|
||||
in16:56:27–17:00:03+02: zero. System4101 across16:53:55–17:00:03: zero.
|
||||
|
||||
All six checkpoints completed/visible, invariant failures0 and transit idle.
|
||||
Login checkpoint has waitCueShown=true (reported, not concealed); subsequent
|
||||
five are false and its adjacent PNG has no portal caption. Final uploads,
|
||||
live teardown, landblock retirement and warmup0. All seven1280x720 PNGs below
|
||||
were opened: **PASS, narrow/provisional**, with the same scope limits as r2.
|
||||
Paths are relative to this run's `artifacts/screenshots/`:
|
||||
|
||||
| PNG | SHA256 |
|
||||
| --- | --- |
|
||||
|477-cathedral-00-login.png|E93C8495590D455D4EE72575F0B821FF15A823959D13988ABBF2722C5BCDBF44|
|
||||
|477-cathedral-01-arrival.png|0D74AE19EA868380D2C35400845D7AE03039DFE5325A282612BAF648429030C2|
|
||||
|477-cathedral-02-movement.png|CF6995FD988896A05CF33E589AA9079BA2ADDD5D9C83130BD6989E571734EC66|
|
||||
|477-cathedral-03-stairwell.png|E3378BC67AEB038EABF6FC98ED569BED165C2E8651E2F5D9B95933ECDE42E208|
|
||||
|477-cathedral-04-stairwell-tilt.png|B7CD111D3DCFD1421B52A5D1A023A61B467BC6909E8E963701ABDB7E56563A96|
|
||||
|477-cathedral-05-stairwell-turn.png|750388A582BE4841C53D39AFDFFC32091DB68F561304082DED135BDC58BEF534|
|
||||
|477-cathedral-06-settled.png|6B2844528F10EB0D31C52E5B3F74CF3F2582C15C5C35A3E2EC76FA44A491C202|
|
||||
|
||||
stdout BC5F97025E57FF886746C8F45F025AC5D5F517371D4FDBB634AE477CDAF3D38B;
|
||||
stderr BAF767100B9E5CEA9C34FED950F01A6D25F0B430C20D73F178A1D820D08AE293.
|
||||
|
||||
## Conclusion and next diagnostic
|
||||
|
||||
Neither bounded isolated mode reproduced the owner's confirmed AMD reset.
|
||||
The original roughly16-minute ordinary session and these roughly2–3-minute
|
||||
instrumented routes are not equivalent; do not close #477/G4 or blame the
|
||||
driver, retail coexistence, or cathedral membership without further evidence.
|
||||
Next useful diagnostic is an ordinary-profile longer reproduction with the
|
||||
first acquire failure retained, not another identical short validation loop.
|
||||
Any code change to preserve that result needs its own bounded contract; no
|
||||
speculative GPU repair is authorized by a clean test. C1a's exact retail
|
||||
membership comparison remains separate and incomplete. FPS remains deferred.
|
||||
|
|
@ -5029,3 +5029,29 @@ Original assertion/hash placeholder and scratch hashes remain unchanged.
|
|||
A production AP159 repair needs a separately expanded contract; narrowing
|
||||
the witness needs explicit owner approval. No option assumed. Scope stop,
|
||||
not static-impossibility or a proved Rynthid cause. FPS deferred, G4FAIL.
|
||||
|
||||
## 60. Cathedral AMD recurrence and isolated validation — 2026-09-05
|
||||
|
||||
Owner first reported the visible fixes looked good and questioned the Rynthid
|
||||
placement report, then crashed in the cathedral at16:41. Exact ordinary
|
||||
Releaseb2bdb09cc/PID7852: abandoned-frame ErrorDeviceLost and Windows141.
|
||||
Reveal had completed; this does not identify an initiating GPU fault or connect
|
||||
the crash to C1a/AP159. Owner closed retail, authorized isolated testing and,
|
||||
after the lead's initial script-verb error, explicitly authorized stopping the
|
||||
stuck diagnostic client. No unrelated app or ACE stopped.
|
||||
|
||||
Corrected81-command route preflight checks every verb/InputAction, balanced
|
||||
holds and terminal close-client. Core/sync run7976 and separate GPU-assisted
|
||||
run2156 use the unchanged green Release binary and v10 package. Both completed,
|
||||
exited0/gracefully, no reported validation/native failures or new Windows GPU
|
||||
events. Fourteen PNGs opened narrow PASS/provisional, not retail parity proof.
|
||||
GPU-assisted login checkpoint retained a wait-cue flag; later five false and
|
||||
all six completed/visible/invariants0. First attempt7444 is discarded/incomplete,
|
||||
not a passing route. Exact paths/hashes/profile and limits:
|
||||
`477-cathedral-recurrence.md`. Both clients now absent; ACE9000 stays up.
|
||||
|
||||
No source, shader, package, driver, TDR, quality or FPS changes. #477/G4 remain
|
||||
FAIL/open: roughly2–3-minute diagnostic routes cannot clear the observed
|
||||
roughly16-minute ordinary crash. Next useful evidence is longer ordinary
|
||||
reproduction plus first-acquire-failure preservation under a bounded contract,
|
||||
not repeating the same short clean route. C1a remains independently unclosed.
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue