When we use AI Reply with Approval Mode, we need audit traceability: show who approved the reply (and ideally who edited it) and keep the label that it was AI-generated even if we tweak a word. This should also be available in export/reporting so we can prove who did what if an issue happens later.