Function keysetPredicate

Build the "row after this one" keyset comparison as a portable FilterExpression.

The comparison is the lexicographic expansion of (resolved sort) > (cursor values): one or leg per position of the resolved sort, where each leg pins every earlier field to its cursor value (eq) and compares the leg's own field strictly in its own direction. For a sort (a asc, b desc, k asc) over cursor values (a0, b0, k0) the tree is or(gt(a, a0), and(eq(a, a0), lt(b, b0)), and(eq(a, a0), eq(b, b0), gt(k, k0))) — for the plan §3.8 case, createdAt desc with an id asc tiebreaker, exactly or(lt(createdAt), and(eq(createdAt), gt(id))). A flat or of per-field strict comparisons is NOT this tree: it matches rows before the cursor whenever any single field sorts beyond its cursor value.

The resolved sort is the caller's orderBy followed by every key column not already ordered (each as asc): the primary-key tiebreaker is a correctness requirement, not a refinement — over rows carrying only two distinct sort values, a walk that omitted it returned 4 of 6 and reported success. A key column already present in orderBy is a sort position already and is never appended twice.

Cursor values layout: orderedValues carries the value of each orderBy field, in order, from the row the cursor was minted against; keyValues carries the primary-key column values in key-column order. An orderBy field is indexed by its orderBy position; a pure-tiebreaker key column by its own key-column position — never a shared orderedValues[0], which may belong to a non-key field.