Bound stored-dual reconstruction separately from actual returned-row error.
Stored Float64 inputs are exact here. Sequential multiply/add uses gamma(2m),
with an absolute subnormal term, before exact normalization by the weight sum.
All positive bounds and minimum-price endpoints are rounded outward.
Bound stored-dual reconstruction separately from actual returned-row error. Stored Float64 inputs are exact here. Sequential multiply/add uses gamma(2m), with an absolute subnormal term, before exact normalization by the weight sum. All positive bounds and minimum-price endpoints are rounded outward.