Sometimes input checks are missing (e.g. L989 of math_1 or L1023 of math_1a), sometimes exceptional cases are not tested (e.g. in fsum), some code is just inaccessible (like L3519 in the perm_comb_small() helper). See attached coverage report (for ./python -m test test_math).
I believe, everything should be tested (except for memory errors and similar special cases).
lcov-report.zip
Linked PRs