In my experience, #4 has held up extremely well over time and is the most compelling reason for choosing an ML over another language. Algebraic data types are an absolute joy to work with when dealing with abstract syntax trees. The other points are obviously great to have, but strong language support for ADTs is key, in my opinion.
I'm not certain I agree with #3. It seems to defeat the purpose of a strong type system. Either way, it can be very nice to express meta-level constructs in a matching object-level type. For example, if the language you are writing a compiler for has an int32 datatype, but you use an int64 in the language you're writing the compiler in, you'll need to simulate overflow. It would just be easier and safer to use an int32 in both places.
These days, I'd recommend Menhir over ocamlyacc unless you have a very specific use case that the former breaks on but the latter works on.
I'm not certain I agree with #3. It seems to defeat the purpose of a strong type system. Either way, it can be very nice to express meta-level constructs in a matching object-level type. For example, if the language you are writing a compiler for has an int32 datatype, but you use an int64 in the language you're writing the compiler in, you'll need to simulate overflow. It would just be easier and safer to use an int32 in both places.
These days, I'd recommend Menhir over ocamlyacc unless you have a very specific use case that the former breaks on but the latter works on.