Takes ownership of empty, compatible layer layouts.
Borrow a prepared request's state. Build every replacement before releasing the active group, so a failed admission leaves it intact.
Retain an immutable batch boundary without extracting or materializing individual rows. Layer layouts decide how to share their backing arrays.
Returns independent request state for a cache publisher or a new group.
The method calls this only between committed steps. A layer failure invalidates the group; its caller discards the state.
Prepare an initial cohort with one merge per layer.
Stage membership so a method can compose it with companion-state admission.
Stage several admissions with one membership operation per layer.
Layer layouts own tensor geometry. This collection only applies the same row membership operation to every layer. It has no scheduling, sampling, checkpoint-key or persistence policy.