Anthropic provider: search_issues fails with JSON parse error

[clintonb]

Environment

• MCP Server Version: 0.29.0
• Transport: stdio via supergateway (streamable HTTP)
• LLM Provider: Anthropic (EMBEDDED_AGENT_PROVIDER=anthropic)

Steps to Reproduce

1. Configure @sentry/mcp-server with EMBEDDED_AGENT_PROVIDER=anthropic and a valid ANTHROPIC_API_KEY
2. Call search_issues with any natural language query (e.g., latest unresolved issues)
3. The embedded agent returns prose wrapping JSON instead of raw JSON, causing a parse failure

Expected Result

search_issues returns matching issues.

Actual Result

Error: No object generated: could not parse the response.

The root cause is visible in the AI_JSONParseError:

JSON parsing failed: Text: Based on the available fields, I can see that Sentry supports
filtering by various attributes. However, I don't see a specific "project" field...

For the query "latest unresolved issues for the API project", I'll construct a query...

```json
{
  "query": "is:unresolved",
  "sort": "date",
  "explanation": "..."
}

The Anthropic model returns explanatory prose around the JSON block instead of raw JSON, which fails `safeParseJSON`.

### Notes

- This is reproducible on every invocation with the Anthropic provider
- Related to #727, which was the same parse error but caused by GPT-5 returning malformed output
- The `list_issues` fallback (without LLM provider) works correctly
- All non-AI tools (`find_projects`, `get_issue_details`, etc.) work fine with Anthropic configured