1.0.61 regression: repeated 0xC0000005 ACCESS_VIOLATION on Windows (clean on 1.0.60, survives cache re-extract)

### Describe the bug

Copilot CLI **1.0.61** crashes repeatedly on Windows with exit code `-1073741819` (`0xC0000005` — ACCESS_VIOLATION). This is a **version regression**: 1.0.60 ran clean for 3 days, then the update to 1.0.61 produced **19 crashes in a single day**.

Critically, the crash **survives a clean re-extract** of the native package cache — so it is **not** binary corruption. Deleting `%LOCALAPPDATA%\copilot\pkg\win32-x64\1.0.61` forced a fresh re-extract of clean bits, and the process still crashed 5 more times immediately after. A `winget uninstall`/`install` of 1.0.61 also did not help (winget does not touch that cache). This points to a genuine code defect in the 1.0.61 build.

### Affected version

- **Broken:** 1.0.61 (WinGet)
- **Last stable:** 1.0.60 (0 crashes over 3 days of use)

### Crash frequency by version (Windows Application Event Log, Event ID 1000, Exception code `0xc0000005`)

| Version | Crashes | Window |
|---------|---------|--------|
| 1.0.57 | 5 | 06-04 → 06-07 (occasional) |
| 1.0.58 | 0 | clean |
| 1.0.59 | 0 | clean |
| 1.0.60 | 0 | clean (3 days) |
| **1.0.61** | **19** | **06-10 (single day)** |

### Fault signature

The faulting module is `copilot.exe` itself, with fault offsets **clustering in one code region**, plus several crashes where the faulting module is `unknown` and the offset is a near-null / wild address (indicative of a corrupted call/jump target, e.g. a bad callback or JIT path):

```
14:50:38 mod=copilot.exe ec=0xc0000005 off=0x0000000001aa41dc
14:36:27 mod=unknown ec=0xc0000005 off=0x0000019b95607d79
14:36:14 mod=copilot.exe ec=0xc0000005 off=0x0000000001aa41b0
14:35:11 mod=unknown ec=0xc0000005 off=0x0000000000000001
14:23:31 mod=copilot.exe ec=0xc0000005 off=0x0000000001aa41cc
12:03:24 mod=copilot.exe ec=0xc0000005 off=0x0000000001aa423c
11:59:40 mod=copilot.exe ec=0xc0000005 off=0x0000000001aa412f
```

The consistent `copilot.exe + 0x1aa41xx` cluster suggests a single defective code path; the `mod=unknown` jumps to `0x1` / `0x19b95607d79` suggest control-flow corruption (bad function pointer / callback).

### What I ruled out

- **Not binary corruption** — deleted `%LOCALAPPDATA%\copilot\pkg\win32-x64\1.0.61` to force a clean re-extract (re-extracted 14:16:51); **5 more crashes on the fresh bits** (14:23 → 14:50).
- **Not a stale winget install** — `winget uninstall` + reinstall of 1.0.61 had no effect (winget does not refresh the `%LOCALAPPDATA%\copilot\pkg` runtime cache).
- **Not a stale multi-version cache** — pruned to current + 1 rollback; no change.

### Environment

- **OS:** Windows 11 (10.0.26200), x64
- **Node.js:** v24.11.1 (bundled with copilot.exe)
- **Install method:** WinGet (`GitHub.Copilot_Microsoft.Winget.Source_8wekyb3d8bbwe`)
- **Launcher:** Agency wrapper — `agency.exe copilot -- --agent general-frodo --yolo`
- **MCP servers:** ~20 configured (teams, kusto, icm, sharepoint, planner, mail, word, workiq, m365-*, enghub, es-chat, msft-learn, bluebird, etc.)
- **Copilot plan:** Enterprise
- **Native module:** `pkg\win32-x64\1.0.61\prebuilds\win32-x64\cli-native.node` (~0.73 MB)

### Workaround

Roll back to 1.0.60 and pin:

```powershell
winget install GitHub.Copilot --version 1.0.60 --uninstall-previous --silent --accept-package-agreements --accept-source-agreements
winget pin add GitHub.Copilot --version 1.0.60
```

### Possibly related (but distinct)

- #2387 — intermittent 0xC0000005 during AI model requests (v1.0.13); non-deterministic, runs fine for a while. This report differs: it is **version-correlated** (clean 1.0.60 → broken 1.0.61) and far more frequent.
- #2702 — 0xC0000005 on **first launch after auto-update** (race condition; retry succeeds). This report differs: crashes **persist across relaunches and a clean re-extract**, not a one-time launch race.
- #2608 (closed) — 0xC0000005 from `win_error_mode.node` in v1.0.22 (same native component family). Possibly the same area regressed in 1.0.61.

### Ask

Please confirm whether a known change landed in 1.0.61 affecting the native engine / a callback path on Windows x64, and whether a 1.0.62 fix or 1.0.60 re-promotion is planned.