[BUG] Claude Desktop crashes under memory pressure during heavy Claude-in-Chrome browser automation (Windows)

## Description

Claude Desktop crashed on Windows 11 during a heavy browser-automation session (Claude in Chrome MCP tools driving many sequential `computer`/screenshot calls against Upwork). No graceful shutdown was logged. This has happened more than once under similar conditions.

## Environment

- OS: Windows 11 Pro 10.0.26200
- Total system RAM: 15.3 GB
- Free RAM at time of crash (from the app's own periodic memory log): declining steadily from ~2.8 GB to ~2.1 GB free over a few minutes, then the log stops mid-session
- Claude Desktop's own process tree was using ~4.3 GB RSS in a separate, freshly-started instance shortly after (the `electron_renderer` process alone was ~1.8 GB), suggesting the app is a significant contributor to memory pressure on lower-RAM machines

## What I found in the logs

`%LOCALAPPDATA%\Claude\Logs\main1.log` (the rolled-over previous-session log) ends abruptly at `2026-08-30 05:51:20` with no shutdown/exit log line — consistent with the process being killed (e.g. by the OS under memory pressure) rather than exiting cleanly.

In the ~2 minutes before the log cuts off, there are two warnings:
```
[warn] [claude-in-chrome] Tool call error: computer (24de56a6) after 30334ms
[warn] [claude-in-chrome] Tool call error: computer (756d8e23) after 30335ms
```
i.e. `computer` tool calls (screenshots) were timing out at ~30s right before the crash — this matches CDP `Page.captureScreenshot` timeouts I was independently seeing during a similar heavy-automation session.

`%APPDATA%\Claude\Crashpad\reports` is empty — no crash dump was captured, which points toward a hard OS-level kill (e.g. OOM) rather than an in-process exception Crashpad could catch.

## Steps to reproduce (best guess)

1. Run a long Claude Desktop session with Claude-in-Chrome doing sustained browser automation (many sequential screenshot/computer tool calls) against a real, complex webpage (in my case, Upwork search/job pages).
2. Leave other memory-hungry apps running concurrently (Docker, IDE, multiple browser tabs).
3. On a machine with ~15 GB total RAM, free memory drops toward ~2 GB.
4. App becomes unresponsive to `computer` tool calls (30s timeouts) and is then killed/crashes with no graceful shutdown and no crash dump captured.

## Expected behavior

Either the app should handle memory pressure more gracefully (release renderer memory, degrade instead of crashing), or at minimum a crash dump should be captured in Crashpad so the failure is diagnosable, and/or a clear warning should surface before tool calls start silently timing out and the app dies.

## Additional notes

I wasn't able to determine the exact installed app version from the local install directory in this session, so I'm not including one — happy to provide it if pointed at where it's recorded on disk.

Start by reproducing the sustained Claude-in-Chrome automation described, while monitoring free RAM and the app process tree. Inspect %LOCALAPPDATA%\Claude\Logs\main1.log around the 30-second computer-tool timeouts and check %APPDATA%\Claude\Crashpad\reports for diagnostic evidence. Done means the failure mode is confirmed and the app either avoids the crash or records enough information to diagnose it.

索引モデルが issue の本文から書いたものです。