I reviewed the field implementation in fp25519_x64 and either I'm missing something or both the add_EltFp25519_1w_x64 and sub_EltFp25519_1w_x64 have incomplete carry handling.
e.g. I can't see why this line would never carry, since RAX could hold any value, and this line is (possibly) adding 38 to it:
I think a similar logic also applies for:
I don't see any tests specifically for add_/sub_. They are covered e.g. by the "mul/sqr" test, but random inputs with 64-bit limbs are very unlikely to catch carry handling errors, especially since random_EltFp25519_1w_x64 is only generating 255-bit values, while add_/sub_ apparently support 256-bit values (?).
I only briefly checked fp448_x64 for comparison, but I believe add_EltFp448_1w_x64 and sub_EltFp448_1w_x64 also have similar errors, probably even less likely to be detected by random inputs.