Fast alternative to double-precision floating point which guarantees “a - b + b == a”

Question

Due to the issue described here (where a shortest-path algorithm fails because -7.430337482745286 - 4.0 + 4.0 does not equal -7.430337482745286 for