We run video analysis via generateContent with YouTube fileData on gemini-2.5-pro (paid, Tier 1). Since Aug 19 we get persistent 503 UNAVAILABLE on every native-video call — over 9 continuous hours on Aug 19 (11:25–20:30 UTC) and again most of Aug 20, across different videos. The evening before we received 429 RESOURCE_EXHAUSTED (generic body, no QuotaFailure details, no retry-after header). Calls succeed only in narrow windows (e.g. 07:50 UTC). Text-only calls work normally throughout — only video input is refused. Is there a per-project throttle on video ingestion we may have triggered (same videos re-ingested repeatedly during testing)? What signal distinguishes model capacity from project-level throttling, given the 503 body carries no details?