Subscription-wide LLM lockout from workflow model inheritance + same-quota-pool fallback chain

[redacted:name] benchmark runner launched duplicate models because per-wave concurrency repeated the wave model