[BUG] Native Windows: startup waits ~4.5 s on three embedded-ripgrep calls that each re-launch claude.exe; agent-view sessions take ~10 s to open (USE_BUILTIN_RIPGREP=0 halves it)

### Preflight Checklist
- [x] Searched existing issues: #10979 and #66395 cover ripgrep at startup, neither the per-call re-launch cost on the native Windows install
- [x] Reproduced on the latest version (2.1.263)

### What's Wrong?
On the native Windows install, the embedded ripgrep is invoked by launching `claude.exe` again with `rg` as the program name. A bare launch of the binary takes about 1.6 s on this machine (`claude --version` alone: 1.6 s, twice). Startup makes three of these calls in series before the first API request and waits on them:

| time after launch | child process (sampled with `Win32_Process` on the parent pid every 150 ms) |
|---|---|
| 1.0 s | `claude.exe` as `rg --version` |
| 2.8 s | `claude.exe` as `rg --no-config --files --hidden --no-ignore --max-depth 4 --glob .orphaned_at C:\Users\\.claude\plugins\cache` |
| 4.3 s | `claude.exe` as `rg --no-config --files --hidden ""` |

In a `--debug-file` log this is a silent 4.3–4.7 s gap right after `Initialized versioned plugins system with 0 plugins`, while `[STARTUP] setup()` itself reports about 200 ms.

Timings for the same headless call (`-p "Reply with the single word ok" --model claude-haiku-4-5-20251001 --setting-sources "" --strict-mcp-config --no-session-persistence --tools "" --debug-file`):

| | embedded rg | `USE_BUILTIN_RIPGREP=0` (system rg 15.2.0 on PATH) |
|---|---|---|
| wall time of the call | 9.4 s | 4.7 s |
| startup to first `[API REQUEST]` | 5.7 s | 1.3 s |
| `Ripgrep first use test` | `mode=embedded` | `mode=system` |

The user-visible effect is that sessions opened from `claude agents` take about 10 s to become ready. Jobs created from the agent view on 2.1.263 took 8.8–9.1 s from `state.json` `createdAt` to the first transcript record, against 1.9–3.9 s for this machine's jobs from June–August 2026 (versions 2.1.178–2.1.222). With `USE_BUILTIN_RIPGREP=0` in the settings `env` block, a new background session took 4.8 s.

2.1.260, 2.1.261 and 2.1.263 all behave the same. The network is not a factor: `curl` to api.anthropic.com completes in 40 ms and the API request's first byte arrives in about 700 ms.

### What Should Happen?
Startup should not pay a full runtime boot per ripgrep call. Any of these would do: keep one rg child alive for the session, move the plugin-cache sweep and the `rg --version` probe off the startup path, or ship a small `rg.exe` next to the binary on Windows instead of re-executing `claude.exe`.

### Error Messages/Logs
```
12:45:36.415Z [DEBUG] Initialized versioned plugins system with 0 plugins
12:45:40.891Z [DEBUG] [servedCatalog] off (flag)
...
12:45:41.186Z [DEBUG] Ripgrep first use test: PASSED (mode=embedded, path=C:\Users\\.local\bin\claude.exe)
```

### Steps to Reproduce
1. Native Windows install, `USE_BUILTIN_RIPGREP` unset.
2. Run the headless call above with `--debug-file x.log`.
3. Note the gap after `Initialized versioned plugins system` and the time to `[API REQUEST]`.
4. Repeat with `USE_BUILTIN_RIPGREP=0` and `rg` on PATH.

### Claude Model
Not model-related (Haiku 4.5 used for the timing).

### Is this a regression?
Yes, by the job records above: 2–4 s to first transcript record on versions up to 2.1.222, 9 s on 2.1.260 and later. No binaries in between are cached here to narrow it down.

### Last Working Version
2.1.222 (from job records, not a controlled comparison)

### Claude Code Version
2.1.263 (same on 2.1.260 and 2.1.261)

### Platform
Native Windows install (`~/.local/bin/claude.exe`, 218 MB)

### Operating System
Windows 11 Enterprise 10.0.26100

### Terminal/Shell
Git Bash and PowerShell 7; also the agent view (`claude agents`)

Reproduce the native Windows startup path with the shown headless command and compare embedded ripgrep against USE_BUILTIN_RIPGREP=0. Start around [STARTUP] setup(), the plugin-cache sweep, and the “Ripgrep first use test” entry in the debug log. Done means startup no longer serially pays for three claude.exe launches while preserving ripgrep behavior.

Do mô hình lập chỉ mục viết ra từ nội dung của issue.