Why not serialize NaN and Infinity as "NaN" and "Infinity" (or "-Infinity")? This is what C# does on an opt-in basis (link).

Two problems with the string representation:

• If a field is specified as a number in a JSON schema, "Infinity" is not of that type, so such output would not be valid for that schema, while 1e99999 is still compliant.
• If reading in data without any schema or other type known in advance, it will be interpreted as a string instead of a number, which means that the exporter would change the data type, while 1e99999 is always imported as a number.

For NaN, null also isn't a number, so either way breaks both of these cases, but it at least implies that the data is "invalid" while still being readable as JSON, unlike "NaN" which looks like a valid string.

Anyway, if we want this functionality, it must be opt-in (as already proposed above, so, good). However, this would mean we need to add a parameter to stringify, which would change the API, and we'd need to register a compatibility method. This could be done as a separate PR, or if it's really strongly wanted I guess I could put it in this PR. But I'm thinking that this would need a proposal, while this PR by itself is good as-is.