Load test a fleet slice against a mock

Three processes run side by side: the mock (port 8000) plays the provider, the service (port 8100) is the fleet slice under test, and Locust plays the fee earners. The service writes one line per turn to out/turns.jsonl, so the report comes from the system's own view (class, status, queue time, attempts, 429s and 529s) rather than from Locust's HTTP timings. The decision that matters most is to scale the mock down (60,000 OTPM, 400 tokens a second) so a sweep takes minutes, and to compare shapes and ratios, not absolute seconds, with the real fleet.