Tool calls and briefings stay intact

Fixed

• Tool calls keep their arguments when the model makes several at once. For a multi-call response the ChatGPT-plan route sends each call's arguments only on its final stream event, and langchain-openai builds arguments from incremental deltas alone, so every call arrived empty. In the v0.2.0 live test the dispatcher's first four flight searches failed this way before a single search succeeded. The model adapter now passes those final arguments through.
• A finished leg's briefing is no longer held back. When the dispatcher finished a leg and went straight to the next leg's approval, the first briefing only appeared at the very end. An approval prompt now carries the briefings finished earlier in the same turn (a briefings field in the MCP response; the terminal chat prints them first).
• The approval prompt's takeoff thrust note checks the departure airport. It said "estimate available" for any supported aircraft, even at Denver, where the thrust model has no data. It now uses the airport's elevation: available, "may be unavailable" near the edge of the modeled band, or unavailable, for example A21N at KDEN (5,434 ft).
• Tests cannot reach the internet. A local test run had made one real FlightAware airport lookup through a missed stub; tests now fail on any request that leaves the machine.