refactor: improve compaction to retain recent conversation context

[nexxeln]

Summary

• keep recent real user turns verbatim during compaction instead of summarizing the full remaining session
• add configurable compaction.tail_turns and compaction.tail_tokens limits with runtime defaults
• persist tail_start_id on compaction parts and update filtered session history to retain the tail across future compactions
• update the compaction prompt to summarize older context while relying on preserved recent turns
• add coverage for retained-tail selection, fallback behavior, and repeated compaction boundaries