Counts rebuilt from trigger events — the same sum TriggerEngine._recount does server-side.

Every stretch here fired at least once. A rule already past its warning cap goes on being true silently, so a late stretch of one can be missing until the call ends — hatched tails mark a stretch still open when it was last reported.

Lines arrive as bot_say / user_said, and each warning as a row of its own where it fired — on a call opened after it ended too, from its trigger timeline. A line carries, as dots, the colour of every warning that stood while it was being said (the server's warning_marks, the same as the stored line); a warning that stood over no line goes on the next one. On a hand-off the faint italic rows at the bottom are what is being said right now, one per speaker: RECRUITER and CANDIDATE off their own legs, and LINE off the bridge, which hears both of them and can name neither. Only the two named rows become transcript lines; LINE is there so you can still read the side of the conversation a leg is mishearing. A BOT line can disappear again: a reply composed on a pause is taken back if the candidate goes on before a word of it plays, and a grey “took back” row marks where it was.