AI Studio silently routing gemini-3.1-pro-preview requests to Flash Lite

Update: After extensive testing, the routing trigger is input token count, not content. Requests with ~12,600 prompt tokens are served by Pro; adding ~73 more tokens causes routing to Flash Lite. The threshold appears to be approximately 12,600–12,700 input tokens. Content, safety settings, API endpoint, and API key are all ruled out as factors.