Build audit evidence packs for agent runs from objectives, tool calls, approvals, outputs, and exceptions