Intermittent ACCESS_VIOLATION (0xC0000005) crash on Windows during AI model requests

### Describe the bug

Copilot CLI crashes intermittently with exit code `-1073741819` (`0xC0000005` — Windows ACCESS_VIOLATION / segfault). The process runs normally for a variable duration (4 minutes to 100+ minutes), then dies abruptly during "Sending request to the AI model" with no graceful shutdown, no Node.js diagnostic report, and no Windows Error Report.

Out of ~13 sessions in a single day, 6 crashed with this exit code and 7 completed successfully. Both short and long sessions are affected. The crash is non-deterministic.

### Affected version

GitHub Copilot CLI 1.0.13-1

### Environment

- **OS:** Windows 11 Insider (10.0.26200), x64
- **Node.js:** v24.11.1 (bundled with copilot.exe)
- **RAM:** 64 GB
- **Terminal:** Windows Terminal
- **MCP servers:** 3 configured (fabric-rti-mcp via uvx, bluebird, workiq)
- **Copilot plan:** Enterprise

### Crash pattern

All crashes share the same signature in the process log:

```
[INFO] --- Start of group: Sending request to the AI model ---
[DEBUG] Flushed N events to session

```

The Agency wrapper reports:
```
Error: Agent execution failed: launch_engine - copilot.exe exited with non-zero status (exit code: -1073741819)
```

### Session data

| Session | Duration | Outcome |
|---------|----------|---------|
| 104415 | 428s | CRASHED |
| 105203 | 1503s | CRASHED |
| 105556 | 3313s | CRASHED |
| 111741 | 44s | OK |
| 111944 | 241s | CRASHED |
| 115130 | 145s | OK |
| 115415 | 4630s | CRASHED |
| 131258 | 54s | OK |
| 131358 | 6047s | CRASHED |
| 132055 | 5503s | OK |
| 145914 | 1836s | OK |
| 150029 | 1460s | OK |

### What I've ruled out

- **Not OOM:** V8 heap peaks at ~664 MB vs 4.5 GB limit; 64 GB system RAM
- **Not command-line args:** Both single and double `--resume` flags appear in OK and CRASHED sessions
- **Not duration-dependent:** 91-min session succeeded, 4-min session crashed
- **No crash dump or WER report generated**
- **Latest version:** `copilot update` confirms 1.0.13-1 is current

### Steps to reproduce

Not reliably reproducible on demand. Happens during normal interactive usage, more likely during extended sessions. The crash always occurs during the network request phase to the AI model.

### Expected behavior

The process should not segfault. If a native error occurs, it should be caught and reported gracefully, or at minimum generate a Node.js diagnostic report (`--report-on-faultaddr`).

If there is more diagnostics I can provide let me know.

No source file, test, or reproducible entry point is identified. Start by comparing the provided successful and crashed session logs around “Sending request to the AI model,” then investigate the bundled Node.js v24 process on Windows and the configured MCP servers. Done means identifying the native fault and preventing the crash or producing a graceful diagnostic report.

索引モデルが issue の本文から書いたものです。