How a graph whose composition qualifies committed known-token spans takes
them (MlxTokenPhases.appendKnown over its committed read): how many
positions one call takes. Every cache lego's committed read keeps the
one-position arithmetic (the cache contract holds it to the same positions
read one at a time), whatever the composition's KV scheme. A graph without
it takes known tokens one position at a time.
Positions one sequence's appendKnown call takes at the caches' current
length. Recheck after each call: native arithmetic can change at a
cache-length boundary.
How a graph whose composition qualifies committed known-token spans takes them (
MlxTokenPhases.appendKnownover its committed read): how many positions one call takes. Every cache lego's committed read keeps the one-position arithmetic (the cache contract holds it to the same positions read one at a time), whatever the composition's KV scheme. A graph without it takes known tokens one position at a time.