Instead of re-nesting prior highlights under '- Previously compacted context:', flatten them directly into the top-level list with '- ' prefix. This prevents each compaction cycle from adding a nesting layer, which inflated the summary by ~depth * overhead per turn.
