Intermittent SDK crash on PR review - exits with code 1 after ~150ms with $0 cost

## Description

The Claude Code Review workflow intermittently crashes almost immediately after initialization, before making any API calls.

## Error Pattern

```
{
"type": "result",
"subtype": "success",
"is_error": true,
"duration_ms": 158,
"num_turns": 1,
"total_cost_usd": 0,
"permission_denials": []
}

error: Claude Code process exited with code 1
at $ (.../node_modules/@anthropic-ai/claude-agent-sdk/sdk.mjs:19:7668)
```

Key observations:
- **duration_ms: ~150-180ms** - crashes almost instantly
- **total_cost_usd: 0** - no API calls were made
- **permission_denials: []** - not a permissions issue
- Model initializes successfully: `"model": "claude-sonnet-4-5-20250929"`

## Intermittent Nature

Same workflow, same configuration - some PRs succeed, others fail:

| Time | PR Type | Result | Duration |
|------|---------|--------|----------|
| 09:35 | Human PR | ✅ Success | 1m52s |
| 07:46 | Human PR | ✅ Success | 1m43s |
| 10:57 | Dependabot PR | ❌ Fail | 21s |
| 12:11 | Human PR | ❌ Fail | 33s |

Re-running the failed workflow produces the same failure.

## Workflow Configuration

```yaml
- name: Run Claude Code Review
uses: anthropics/claude-code-action@v1
with:
anthropic_api_key: ${{ secrets.ANTHROPIC_API_KEY }}
prompt: |
Please review this PR...
claude_args: |
--allowed-tools "mcp__github__get_pull_request,mcp__github__get_pull_request_files,mcp__github__create_pending_pull_request_review,mcp__github__add_comment_to_pending_review,mcp__github__submit_pending_pull_request_review,mcp__github__get_pull_request_diff"
```

## Environment

- Runner: `ubuntu-latest`
- Action version: `anthropics/claude-code-action@v1`
- API key: Confirmed valid (works on other PRs in same repo)

## Expected Behavior

Workflow should either succeed or fail with a meaningful error message.

## Actual Behavior

SDK crashes with minified JavaScript stack trace, no clear error message about root cause.

## Questions

1. Is there a known issue with certain PR content triggering this?
2. Would pinning to a specific version help?
3. Is there additional debug logging we can enable?

Thanks!