The injected block for a compound question: one labeled line per criterion, plus a single
retrieval floor. Each line names the sub-claim number and the question's own wording, so the model
matches a window to a criterion by reading rather than by inferring.
At most one cadence criterion enumerates its sub-periods — a second enumeration would repeat up to
MAX_BUCKETS lines for little gain, so later ones collapse to their span and a period count.
The injected block for a compound question: one labeled line per criterion, plus a single retrieval floor. Each line names the sub-claim number and the question's own wording, so the model matches a window to a criterion by reading rather than by inferring.
At most one cadence criterion enumerates its sub-periods — a second enumeration would repeat up to
MAX_BUCKETSlines for little gain, so later ones collapse to their span and a period count.