SummaryCompressor

Writes the summary a SummarizationMiddleware keeps (design §4.34).

compress returns a summary of previous - the summary so far, if any - followed by messages, the messages after it, never those it already covers: a summary is extended, not recomputed. It should stay within maxTokens estimated tokens at four characters a token; the middleware cuts what it returns to that bound either way.

A compressor that calls a model must call it through context.accounted(client), so the call is counted in the thread's usage, also when the model step later fails. Its result is pure state: when the model step runs again (recover), it is called again.