Reimplemented “to_nearest” rounding and it now passes all 100 test cases in cases.p plus the failed case reported. Any additional test data is welcome.
This is a completely separate implementation from rational number conversion and requires no additional bigints to be created. But it does involve 4-5 calls to GMP library functions. Execution time on my MacOS Intel config is 1.7 sec for 50000 reps (5,000,000 calls to =\=).