Worst: Single generateDraftReply calls reach 264s in the worst-latency cohort the tool is 40-87% of end-to-end latency in 28/30 traces.
Problem
Evidence
Why-chain
AI text/layout recreation from video frame; verify against source image.