[BUG] Windows MSIX 1.32352.1 (latest): GPU process crash (exitCode 101457950) triggered by the app's own claude.ai auth webview — Code Integrity blocks vk_swiftshader.dll (Events 3033/3010, per-second correlation)

Preflight Checklist

• [x] I have searched existing issues — related: #81698 (tracking), #83835, #87204, #87329. Filing separately because this adds: reproduction on the latest desktop build 1.32352.1.0, a trigger that is the app's own claude.ai auth webview (not a user-opened external site), and per-second Code Integrity correlation on a desktop (non-hybrid-laptop) machine.
• [x] This is a single bug report
• [x] I am using the latest version of Claude Code

What's Wrong?

Claude Desktop (Windows MSIX) dies entirely — all running Claude Code sessions lost — when the Chromium GPU process crashes with the well-known signature. On this machine it happened 6 times between 2026-08-11 and 2026-08-18, four of them on 08-18 alone while running several concurrent coding sessions.

Death signature (main.log, byte-identical every time):

2026-08-18 20:51:29 [info] [oauth] using cached token for orgId=<org-id>
2026-08-18 20:51:31 [info] GPU process gone: {
  type: 'GPU',
  reason: 'crashed',
  exitCode: 101457950,
  serviceName: 'GPU'
}
-- log ends; whole app gone --

Trigger: every one of the 6 crashes is preceded 1–3 seconds earlier by the app opening its own claude.ai auth webview:

[info] using oauth config {
  apiHost: 'https://api.anthropic.com',
  redirectUri: 'https://claude.ai/desktop/callback',
  domain: '.claude.ai'
}
[info] [oauth] looking up token for orgId=<org-id>
[info] [oauth] using cached token for orgId=<org-id>
[info] GPU process gone: { ... exitCode: 101457950 ... }

So the Cloudflare-challenge/WebGPU-probe trigger reported in #87204 and #87329 fires here without the user browsing anywhere — the app's own login/auth surface is enough. Running multiple concurrent sessions makes it more frequent (more webview creations).

Root-cause correlation (Windows Code Integrity, per-second)

Microsoft-Windows-CodeIntegrity/Operational, at the exact second of each crash (verified for the 20:14:34 and 20:51:31 crashes on 08-18):

Event 3033:

Code Integrity determined that a process
(\Device\HarddiskVolume3\Program Files\WindowsApps\Claude_1.32352.1.0_x64__pzs8sxrjxfjjc\app\claude.exe)
attempted to load
(...\WindowsApps\Claude_1.32352.1.0_x64__pzs8sxrjxfjjc\app\vk_swiftshader.dll)
that did not meet the Microsoft signing level requirements.

Event 3010 (same second, ×3):

Code Integrity was unable to load the
\Device\HarddiskVolume3\Program Files\WindowsApps\Claude_1.32352.1.0_x64__pzs8sxrjxfjjc\AppxMetadata\CodeIntegrity.cat
catalog. Status 0xC000003A.

i.e. the 1.32352.1.0 MSIX still ships no CodeIntegrity catalog (AppxMetadata path does not resolve), so the GPU sandbox's signing-level check rejects the app's own bundled SwiftShader DLL — same mechanism documented in #87204 for 1.30096.5.

Control observation: Chrome 151 on the same machine hit the same Event-3033 block on its vk_swiftshader.dll minutes earlier (20:19:38) and degraded gracefully — no crash. Claude's GPU process dies and takes the whole app with it.

MSIX aftermath (matches #81875 / #85905)

After each crash, AppXDeploymentServer/Operational shows the package flipped to a bad state and Windows grinding through a repair loop before the app can start again, e.g. on 08-18: repeated Deployment Register 19:07:55–19:09:56 (×7) → Deployment Remove 19:10:02 → Deployment Add (fresh Claude-4061727268.msix) 19:12:47. Two full Remove→Add reinstall cycles that evening. To the user this presents as the app "crashing again" while reopening it.

Ruled out on this machine (all tested)

• NVIDIA driver update 591.86 (Jan 2026) → 610.88 (32.0.16.1088, 2026-07-22): crash reproduces identically on both
• Intel iGPU driver update → 32.0.101.8974 (2026-08-11): no change
• Deleted GPUCache / DawnGraphiteCache / DawnWebGPUCache: crashed 42 seconds after the caches were rebuilt
• No TDR (System log clean, no Event 4101), no nvlddmkm resets
• No Crashpad dump, no WER AppCrash, no Application Error 1000 for Claude.exe
• Not memory pressure: VRAM 2.5/16 GB, system RAM ~25 GB free at crash time (app's own process-memory log)
• --disable-gpu not attempted as a fix — #86265 already shows the signature reproduces with it, consistent with SwiftShader being the blocked module

Environment

| Item | Value |
|---|---|
| App | Claude Desktop 1.32352.1.0 (MSIX, Claude_1.32352.1.0_x64__pzs8sxrjxfjjc), Node 24.18.1 per startup log |
| Bundled CLI | Claude Code 2.1.229 |
| OS | Windows 11 Pro 10.0.26200 (desktop PC) |
| GPU 0 | NVIDIA GeForce RTX 5070 Ti, driver 32.0.16.1088 (610.88, 2026-07-22) |
| GPU 1 | Intel(R) Graphics (iGPU), driver 32.0.101.8974 (2026-08-11) |
| RAM | 48 GB |
| Security | Standard Windows 11 defaults; Smart App Control not the variable (Chrome control case above) |

What Should Happen?

1. The MSIX package should ship a valid AppxMetadata\CodeIntegrity.cat (or otherwise satisfy the GPU sandbox's signing-level policy for vk_swiftshader.dll), so the software-rendering fallback can load.
2. Failing that, a GPU-process crash should not take down the main process and every running session — Chromium/Chrome demonstrably survives this exact block on the same machine.
3. A failed relaunch should not corrupt the package into a Remove→Add reinstall loop.

Error Messages/Logs

Key excerpts included above. Full main.log excerpts, CodeIntegrity 3033/3010 event dumps, and the AppXDeployment timeline available on request.