Compacts a full Responses API input window into a shorter canonical window.

The client sends its current input (the same items that would be passed to POST /v1/responses) and receives a compacted output window. The output should be used verbatim as the input of the next /v1/responses call (appending only the new user turn).

This generalizes the compaction approach used by the coding-agent harness (generate_session_compact in xai-grok-shell):

1. Strip tool-result noise from the history.
2. Ask the model to produce a structured <summary> of the conversation.
3. Rebuild a compact window: system message → summary → last user query.
4. Return that window to the client.