In protoc, -nan is an allowed default value for a double or float field, but protocompile disallows it.

In protoc, float and double values accept inf, infinity, and nan with any kind of capitalization. But protocompile only allows inf and nan and only in lower-case.

Additionally, protoc disallows inf, -inf, nan, and -nan for other double or float options; they're only accepted for field default values. However, I think that is a protoc bug and not behavior we necessarily need to replicate (see protocolbuffers/protobuf#15010).