[Windows] GPU-process crash in Cowork preview bricks MSIX install (NeedsRemediation loop) → forced reinstall wipes local sessions

Environment

• Claude Desktop 1.24012.9.0 (MSIX/sideload package Claude_pzs8sxrjxfjjc, signature valid, files verified intact against AppxBlockMap)
• Windows 11 Pro build 26200, 64 GB RAM
• GPU: AMD Radeon RX 7800 XT, driver 32.0.31035.1003 (2026-07-24)

Summary

Rendering a page that initializes MediaPipe tasks-vision (FaceLandmarker, WebGL/WASM + WebGPU probe) inside the Cowork/local-agent preview crashes the Electron GPU process. After that crash, the app can never be activated again: Windows fails to start the packaged service and flags the package Modified, NeedsRemediation; remediation "succeeds" but the flags return on every activation → infinite loop. The only user-visible fix is uninstall+reinstall of the MSIX, which deletes %LOCALAPPDATA%\Packages\...\LocalCache\Roaming\Claude — all local-agent sessions/history are lost. Happened repeatedly on two different machines.

Crash evidence (window log, 1 s before GPU death)

13:36:47 [warn] WebGL: INVALID_ENUM: getInternalformatParameter: invalid internalformat   (x dozens)
13:36:47 [warn] ... when EXT_color_buffer_float is not enabled
13:36:47 [warn] The powerPreference option is currently ignored when calling requestAdapter() on Windows. See https://crbug.com/369219127
13:36:48 [warn] A valid external Instance reference no longer exists.
13:36:48 [warn] WebGL: CONTEXT_LOST_WEBGL: loseContext: context lost

main.log:

13:36:48 [info] GPU process gone: { type: 'GPU', reason: 'crashed', exitCode: 101457950, serviceName: 'GPU' }

(also GPU crashes on 2026-06-23 exitCode 34 and 2026-07-01; WER RADAR_PRE_LEAK_64 for claude.exe 1.18286.0.0 on 2026-07-04; crash-loop clusters: 5 restarts/6 min on 06-28, 7 restarts/4 h on 07-11)

The bricking loop (the worse half of the bug)

After the GPU crash, every activation attempt:

1. SCM cannot start the packaged CoworkVMService — manual sc start returns error 87 "The parameter is incorrect"; the service exe is never launched (its own log shows no start attempts).
2. AppXDeploymentServer sets package status Set=0x20 (NeedsRemediation), runs a Register remediation that "finishes successfully" and clears 0x300 / 0x10000C00 / 0x20…
3. …the next activation sets 0x20 again. Get-AppxPackage permanently shows Modified, NeedsRemediation.

Package integrity was verified OK: all 2180 files present with exact AppxBlockMap sizes, Authenticode/AppxSignature valid (Anthropic PBC cert, expires 2026-10-21, timestamped). The "Modified" flag is false — the registration state is what's wedged.

Recovery that worked without data loss: installing the Squirrel build (Claude-...exe from downloads.claude.ai) alongside, then migrating LocalCache\Roaming\Claude → %APPDATA%\Claude (cached OAuth token and sessions came back cleanly), then Remove-AppxPackage.

Suggestions

1. Don't let a packaged-service start failure brick activation — after N failed remediations, fall back to launching without the VM service (degraded mode) instead of refusing to start.
2. Treat LocalCache\Roaming\Claude as user data — local-agent sessions should survive uninstall (migrate out of the MSIX container, or export/restore on reinstall). Losing all conversations on every forced reinstall is the painful part.
3. GPU-process crash from embedded WebGL/WebGPU content (MediaPipe on AMD) is reproducible for us — happy to test builds.

A 34 MB Crashpad minidump of the GPU crash + full logs are preserved and available on request (not attached publicly: may contain memory contents).