#964 TokenGate client: per-process concurrency cap below TG's 24 per tenant

EARS: The sum of in-flight TokenGate calls across ALL Highway processes (api, worker x2, activity-worker, internal-worker x2, CLI) shall stay at or below 20, because TokenGate's 24-slot gate is per TokenGate tenant and all of Highway is one tenant (highway). Default per-process cap = floor(20 / number of calling processes), configurable via [tokengate] max_in_flight. Context: TG answers 2026-10-17 (01M53P27901QY3SX0PQAWJ3KAK): bursts above 24 wait up to 10 s then 429 tenant_concurrency_limit; at ~15 ms/call 20 in flight carries ~1,300 calls/s.