Offline continuation replayed each slot's newest cached answer. A
provider-source slot is answered with the provider's next item on every
play, so its newest answer is the image already on screen: on a timeline
of one source slot, every offline boundary "advanced" to the same
picture. Stage 0's real timeline is exactly that, which is why the wall
froze whether the network or only the API went away, and resumed the
moment the server answered.
Each slot's cached answers are now its observed rotation (one per item,
oldest-played first), and the k-th play of the slot since the anchor
shows the k-th entry. The play count is arithmetic, skipping whole
passes wherever the eligible set is stable, so a week offline is still
one evaluation. Where a play starts is unchanged and still decided by
advanceCursor alone; a pinned slot is a rotation of one and behaves as
before. Nothing is invented: only answers the server already resolved
are replayed, and an item whose newest answer is unusable is skipped.
Regression tests: the resolver on a one-slot rotation of two and three
images, across a week, and with an unusable image; and the controller
on the real shape for six slide intervals with the network down and
with only the API down.
Co-Authored-By: Claude Opus 5.5 <[email protected]>