ReadonlyprefillWhether prompts leave their last token pending for the first verify window
(the composition's prefillTailSplit).
One new position per sequence through the graph's decode phase, capturing tapLayers: a round without
drafts, the step plain decode runs.
OptionaltapLayers: readonly number[]Optionalwork: MlxForwardWorkThe graph's composed caches for one row.
One prompt step in the phase its planner named, capturing tapLayers.
OptionaltapLayers: readonly number[]Optionalwork: MlxForwardWorkWhether the forwards can capture these hidden layers (a provider's declared taps).
A verify window, capturing tapLayers.
OptionaltapLayers: readonly number[]Optionalwork: MlxForwardWork
The target operations the speculative group and the fill group call, bound once from the graph's phases. A replacement graph supplies this entire port, including any hidden taps. The verifier never needs a MlxTokenGraph or an artifact-family check. Each forward names its phase;
work, when given, is the mixed iteration's hook and runs the same rows inside a shared forward.