Ask over the composer, one question at a time
The form-in-the-transcript was the wrong shape for this. A question the
run is waiting on is not part of the conversation, so it now leaves the
transcript entirely: a store hands it to the composer, which draws it as
a panel above the input. The sticky wrapper it replaces only moved the
form to the bottom of the scroller, where it still scrolled with the
conversation and showed through.
Options are rows, not a dropdown — one click on the text, with the
model's explanation under each label, and a radio or a checkbox marker
depending on whether more than one may be chosen. Questions come one at
a time with Back/Next rather than stacked into a form.
Answering settles it into the same collapsed row the transcript already
shows, which is the one part worth keeping in the stream.
