Claude Desktop 1.40609.0 (Windows, MSIX) hard-crashes with 'GPU process isn't usable. Goodbye' when the built-in browser pane opens a Cloudflare Turnstile login page

Environment

• Claude Desktop 1.40609.0, Microsoft Store/MSIX package (Claude_pzs8sxrjxfjjc), Windows 11 Pro 26200
• NVIDIA GeForce RTX 3080 Ti; reproduced on drivers 610.88 and 616.56
• 4K display, 10-bit color (P3)

Repro (5/5 tonight: 01:01:24, ~01:50, ~02:20, 02:45:17, 06:27:26 local)

1. In a Claude Code session, open the built-in browser pane at https://dash.cloudflare.com/login
2. The page renderer process spawns; the GPU process dies the same second; ~1 s later the entire app exits. Crash #1 left a dump with LOG_FATAL gpu_data_manager_impl_private.cc:418: "GPU process isn't usable. Goodbye."
3. A leftover chrome-native-host.exe keeps running after the fatal, so Windows thinks the app is still open and blocks relaunch until it is killed.

Ruled out experimentally

• GPU driver: reproduced on both 610.88 and 616.56.
• Hardware acceleration: reproduced with isHardwareAccelerationDisabled: true; the GPU process was confirmed running software (--use-angle=d3d11-warp-webgl).
• Profile corruption: reproduced after clearing the pane partition's Cache, Code Cache, GPUCache, Dawn caches, Service Worker storage and the site's IndexedDB.
• The pane renders other sites fine (example.com, YouTube).
• Control: the same URL in regular Chrome on the same machine — no crash, no WebGL/WebGPU/GPU console errors; only Cloudflare Turnstile client error 300010.

Additional observations

• Crashes 3–5 left no Crashpad dump, no WER entry, no Event Log or kernel driver (nvlddmkm) events — the GPU process appears to exit instantly and repeatedly rather than fault, until Chromium's fatal fires.
• Process timeline from an external 1 Hz watchdog: pane renderer born → +4 s page renderer born and gpu-process dies the same second → +1 s all app processes gone.
• The bundled crashpad handler runs with --url=https://f.a.k/e, so no crash telemetry reached Anthropic; reporting manually.

Suspected: Turnstile's GPU-capability probing hits an initialization path in this build's GPU process that exits immediately, and the relaunch loop escalates to the app-wide fatal. Happy to provide the watchdog log and timestamps.

Filed via Claude Code on the user's behalf.