ReadonlynamespaceReadonlyrowReadonlytapOptionalconsumeConsume externally chosen target inputs at a committed round boundary. Context and tokens are borrowed rectangular windows; lengths retain each row's consumed prefix, excluding right padding and the pending output.
Each row returns 0..depth proposals. The method verifies the returned lengths; an empty row still consumes its pending token.
OptionaldraftOptional: the same proposals left on the device as [B, depth] ids, resolved from one later host readback. Lets the round build its verify graph while the GPU is still drafting. Null when unavailable for this call.
OptionalmaterializeDiagnostic barrier for retained lazy state; never called by normal rounds.
OptionalpreviewOptional: the proposals draftDevice would make from pending [B, depth], owned and left on the device,
without changing the rows' state (a recurrent head drafts under a round of its own state that it rolls back), so
a method can compare them with the target's own next samples instead of verifying them. A drafter may propose
fewer positions than depth (the first ones): a recurrent head proposes one.
Method-owned state membership, independent of scheduling and cache tiers. Open/append borrow checkpoints; capture returns owned immutable state. Membership changes and capture occur only between committed rounds. The caller retains the provider's residency lease for the group's lifetime.