…hy support

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

…tore

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

- tool_search now matches on tags and path fields (present on skill-backed tools)
- Include tags/path in search result JSON when available
- Add path field to AvailableSkill type for prompt rendering
- Orchestrator includes path in skill list when present: (id: `...`, path: ...)

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

… get LLM-simulated responses

- simulator.rs: calls gpt-5-nano (Azure) to generate plausible tool responses
- Whitelist of safe tools (tool_search, load_skill, write_todos, search, final)
- Unsafe tools (api_request, connection_request, execute_shell, etc.) are simulated
- ExecutorContext gets dry_run: bool flag
- execute_tool_calls_with_timeout checks flag before each tool execution
- Simulated responses marked with _simulated: true

…ntext

Add dry_run field to ExecutorContextMetadata so clients can pass
{"dry_run": true} in the JSON-RPC message metadata. The A2A handler
reads this field and sets ExecutorContext.dry_run, enabling the eval
simulator to trigger simulated tool execution without side effects.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

…e before handle_external_tool_inline

External tools were bypassing dry-run check and going straight to
handle_external_tool_inline which waits for a client response that never
comes in eval mode. Now dry-run check happens BEFORE the external tool check.