Ah, I see your point, but I don't think it's related. It was happening before 4.5 actually. And I'm sure OH does condensing usually long before context is anywhere near full.
My best guess is some threading/async issue. FWIW I've never seen it locally, only on the cloud.
Oh! and I'm using GPT-5. 😅
That Anthropic context thing is so concerning IMO though. I don't know yet how it behaves, I am still to spend enough time with 4.5, but the fact that it inserts "system warning" in context over API is ... I have no words