Add Exa MCP server + fix tool schema validation for Google AI Studio models

[Copilot]

Google AI Studio (Gemini/Gemma via OpenRouter) rejects tool schemas where required lists property names absent from properties. Many MCP servers ship schemas with this violation, making these models unusable when any such tool is loaded.

Schema sanitization (src/llm.rs)

Added sanitize_schema() — a recursive serde_json::Value mutator applied to every tool definition in chat_with_model() before the API call:

• Filters required[] to only retain keys that exist in properties
• Drops required entirely if the filtered array is empty
• Recurses into nested properties and items schemas
• No-op for already-valid schemas → all existing models unaffected

fn sanitize_schema(schema: &mut serde_json::Value) {
    let Some(obj) = schema.as_object_mut() else { return };
    let defined_props: Option<HashSet<String>> = obj
        .get("properties").and_then(|p| p.as_object())
        .map(|p| p.keys().cloned().collect());
    if let Some(ref props) = defined_props {
        if let Some(req) = obj.get_mut("required") {
            if let Some(arr) = req.as_array_mut() {
                arr.retain(|v| v.as_str().map(|s| props.contains(s)).unwrap_or(false));
            }
        }
    }
    // remove empty required, recurse into properties + items …
}

Exa MCP server

• setup/index.html: Added exa to MCP_CATALOG under Knowledge & Data — bridges https://mcp.exa.ai/mcp via npx mcp-remote, requires EXA_API_KEY
• config.example.toml: Added commented [[mcp_servers]] example block for Exa