A finished answer could be left with its tail, usually the citations,
below the fold, and nothing would bring it back: the view stayed put with
the question still on screen above an answer whose end you had to scroll
to by hand.

The chat log followed the answer by hand: a tick every 150ms measured the
log and scrolled it to the bottom, with a pair of flags deciding whether
the user had scrolled away and wanted to be left alone. Measuring is the
problem. While the model streams, the answer often still fits, so there
is nothing to scroll; it only outgrows the pane when the answer finishes
and renders the tail it had buffered plus the citations. Markdown mounts
those asynchronously, so the last scroll of the turn ran against a widget
that had not grown yet, did nothing, and left no tick behind to fix it.

Textual has this: an anchored widget stays scrolled to the bottom as
content arrives, until the user scrolls away. It re-pins the scroll while
it arranges the layout, off the size it just computed, so there is no
moment where the view is behind the content and no measurement of ours to
get wrong. It also handles the scrollbar and the wheel, which our version
never did.

So the follower goes, and a turn anchors the log instead. This is the
second bug in that hand-written follower, after the auto-follow disabling
itself in #294.

Tests: the scroll ones drove _scroll_to_bottom directly, so they are now
written against what the reader sees, and all of them set up by sending a
question rather than anchoring the log themselves, which is what lets
them fail when the anchoring goes missing.

The scroll tests read scroll_y and max_scroll_y back after a single
pilot.pause(), which settles none of what they depend on. The scroll
bindings animate, so scroll_y samples somewhere on the easing curve at
whatever point wall-clock allowed: on a loaded runner it had not moved
at all yet and "the reader scrolled up" asserted 135 < 135. A bubble
also only buffers appended content until compose hands it a content
widget, so an answer grown before that lands leaves max_scroll_y at 0
and the preconditions compare 0 to 0.

Wait on Textual's own wait_for_scheduled_animations, which drains
pending messages, runs the animations out, and drains again. The turn
helper now also asserts the bubble composed, so a test that would
otherwise pass against an empty bubble fails instead.