[Windows] MSIX package gets flagged 'tampered' after CodeIntegrity catalog load failure — app unlaunchable until full reinstall (twice in one morning)

Environment

• Claude Desktop 1.26832.0 (MSIX package Claude_1.26832.0.0_x64__pzs8sxrjxfjjc, installed under C:\Program Files\WindowsApps)
• Bundled Claude Code 2.1.222
• Windows 11 Home, build 10.0.26200
• AMD dual-GPU system (RX 6900 XT + RX 9070 XT), 64 GB RAM, NVMe system drive (healthy)
• Windows Defender only; Smart App Control OFF; no third-party AV; no Defender detections

What happened

The desktop app hard-crashed and became permanently unlaunchable twice in one morning (2026-08-11, ~7:28 AM and ~7:55 AM local). Each time, the only fix was a full uninstall + reinstall, which also wipes %APPDATA%\Claude (session index, logs, Crashpad) — the app then shows no session history even though ~/.claude/projects transcripts are intact.

Root cause evidence (from Windows event logs)

Microsoft-Windows-CodeIntegrity/Operational, at both failure times (7:28:52 and 7:55:06):

• Event 3033 (Error): Code Integrity determined that a process (...\WindowsApps\Claude_1.26832.0.0_x64__pzs8sxrjxfjjc\app\claude.exe) attempted to load ...\app\vk_swiftshader.dll that did not meet the Microsoft signing level requirements.
• Event 3010 (Warning, 3x each time): Code Integrity was unable to load the ...\WindowsApps\Claude_1.26832.0.0_x64__pzs8sxrjxfjjc\AppxMetadata\CodeIntegrity.cat catalog. Status 0xC000003A.

Microsoft-Windows-AppModel-Runtime/Admin: immediately after each CI failure, the package status flips to 0x2 (tampered/needs-remediation). Once flagged, the app cannot launch. Windows attempted automatic re-register/remediation six times (7:55:19–7:57:31), each "finished successfully" per AppXDeploymentServer, but the status returned to 0x2 every time. Only a full Remove + Add cleared it.

AppXDeploymentServer/Operational also shows the first recovery attempt partially failing: Register attempts failed with 0x80073D02 (package in use), and three Add attempts failed with 0x80073CF6 before a reboot allowed reinstall #2 to succeed at 7:35:47. The app then ran ~20 minutes and the identical CI failure recurred at 7:55 on the freshly installed package.

Likely trigger

vk_swiftshader.dll is only loaded when Chromium falls back from hardware GPU / re-probes the Vulkan stack. The same CI 3033 Vulkan-probe events (vulkan-1.dll / vk_swiftshader.dll) fired for Chrome on 08-05 and Brave on 08-10 on this machine — harmless for them. Only Claude, being MSIX-packaged with a per-package signing catalog, had the catalog read fail (0xC000003A) during that validation, which flags the entire package tampered and bricks it. The pattern began the morning after a GPU change in this machine (RX 9070 XT installed 08-04), which presumably invalidated Chromium's GPU info cache and triggered Vulkan re-probing.

Impact

• App unlaunchable until manual reinstall (twice in one morning).
• Each reinstall wipes %APPDATA%\Claude, so the desktop app loses its entire session index/history. Transcripts survive in ~/.claude/projects but the app UI shows nothing, which reads as total data loss to the user.

Suggestions

• Investigate why AppxMetadata\CodeIntegrity.cat fails to load (0xC000003A) during vk_swiftshader signature validation on an otherwise healthy install.
• Consider hardening: if the package gets flagged, an in-place repair path that preserves %APPDATA%\Claude; or store the session index under ~/.claude so reinstalls don't orphan session history.