fix(llm, cli): Stop OpenAI silently shortening summary requests

[JeanMertz]

Summarizing a turn range larger than an OpenAI model's context window now fails with the API's rejection instead of succeeding with a summary that only covers part of the range. jp conversation compact --turn=2..150 --summary against a GPT model previously returned a normal completed response built from whatever OpenAI kept, and that summary was stored as standing for all 149 turns; every later request projected it in place of turns the model never read.

The cause is truncation: auto, which JP sent on every OpenAI request. It lets the API drop input items from the middle of a conversation and answer anyway, so the response is indistinguishable from one that read everything. summarize_events rejects a response the model cut short, but a request the provider cut short arrives as an ordinary Completed and passed straight through.

ChatQuery gains a truncation field carrying whether the provider may drop input. Summarization asks for Forbidden and gets a 400 it can report; the main query loop and tool inquiries keep Allowed, since nothing fits the stream to the window on the query path and a long conversation stays answerable only if the provider may drop what it cannot hold. Only OpenAI offers the choice — the other providers either always reject an oversized request or truncate server-side, out of JP's reach.