Preloading was counted in steps and then shared out between the panels,
so with four panels only the very next step was ready and each panel did
its resolving and buffering at the moment it was swiped to -- four
panels, four starts in a row. It is now counted per panel: every panel
keeps its next three videos ready (N * 3 for a step of N), the rendered
window reaches far enough forward to hold them, and the step right
behind the one being watched buffers for real rather than taking its
metadata alone.
Their formats are resolved up front and all at once, rather than one
panel at a time as each is swiped to.
The listing is fetched several pages at a time to keep up: a page number
is claimed before its request goes out, which is what lets pages
overlap, and finished pages wait their turn so the reader still gets
them in order. A page whose request fails puts its number back rather
than leaving a hole. How far ahead to run scales with the split, since a
step of N panels gets through a page N times as fast.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MPZFnNdHbPGDTqQUNiE4ZN
