Function run_tool 

pub async fn run_tool(
    state: &Arc<ServerState>,
    tracker: &Arc<ReadTracker>,
    caps: &Arc<CapabilitySet>,
    caller_id: Option<&str>,
    tool_id: String,
    args: Value,
    dry_run: bool,
) -> Response

Execute one RunTool request against the shared dispatch state.

Single-entry-point for both transports: UDS goes through dispatch_request which forwards Request::RunTool here; HTTP (atd-server-http::mcp::tools_call) calls this fn directly. The emitted Response is the same enum + same JSON encoding on either route — Response::ToolResultResponse on success / execution-failed, Response::Error on capability-denied / rate-limited / broker error / tool-not-found / invalid args / internal error.

The body is the verbatim move of the SP-streamable-http §6.3 atd-server::connection.rs::dispatch RunTool arm — every audit emission, capability check, semaphore acquire, broker call, binding dispatch, middleware step, and error map is preserved. The existing atd-server::connection::tests suite covers all branches and is re-routed through this fn unchanged after refactor.