The scroll position was a pixel offset into a document that changes while
it is being read. Growing an event above the reader left scrollTop where it
was, so the text under the reading line slid: measured going from "Lucas
Bennett" to "Lunch" on an unchanged offset of 900.
Hold the position as a place in the script instead — the block under the
reading line plus an offset into it — and re-resolve it to pixels whenever
the document is measured. This follows how professional prompters cue a
story plus an offset rather than a scroll offset, so a reorder carries the
reader with the block and an edit above them changes nothing. A deleted
block falls back to the end of the nearest surviving event before it.
Add event-to-event stepping on Shift with the vertical arrows, matching
Shift being the coarser step on the horizontal ones, and take the step from
where the scroll is headed so pressing again mid-ease moves on rather than
re-aiming at the same event.
Free scrolling and the existing nudge and page keys are unchanged, but they
now hand over the scroll by one rule: how far the reader moved the script
themselves, accumulated. Distance from the follow target could not tell a
catch-up still running from a reader who had moved, so touching the wheel
while the prompter eased towards a newly loaded event stopped it following.
Accumulating also measures a wheel gesture the browser spreads over many
frames, which the old per-event check sampled only once.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Cb8RVPNQ2ETPJxdy4b8CHf
Two problems were feeding the confusion:
- Any wheel, touch, or pointerdown event broke the follow, with no tolerance.
A stray touch or trackpad momentum after a deliberate gesture disengaged it
as readily as a real scroll away from the read position, with no way to
tell the two apart from the outside. The operator view solved the same
problem with a distance check against the loaded item's expected position,
which this now mirrors: it takes a real scroll (past 1.5 lines) to count as
taking over, in a new hasBrokenFollow(), unit tested like the rest of the
scroll maths.
- The runtime flag and the view option shared adjacent, opposite-valence
names (followLocked / followLoaded) and were ANDed together in a third
place to get the value the button actually needed. Renamed the runtime
flag to autoScrollLocked, matching the operator's own lockAutoScroll for
the same concept, and moved the AND into the hook so it exposes one signal
the view no longer has to assemble itself.
Also fixes the threshold check reading a stale position: a burst of wheel
events can fire faster than the animation frame that keeps the scroll ref in
sync, so it now reads the element's scrollTop directly rather than the ref a
frame behind it. Caught by testing the tolerance in a live browser rather
than trusting the unit tests alone.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Cb8RVPNQ2ETPJxdy4b8CHf
Claude
Carlos Valente