Personally, even though I haven't really thought it through deeply, I feel like Scala would be better if everything were curried.
> Subtyping also has many complex consequences
But also substantial benefits!
> There is limited type inference
I wish Scala would improve in this regard. Maybe paradoxically, inference works better with code tuned for subtyping and variance.
> tail call optimisation
This will be fundamentally solved by Project Loom. At least on the JVM (there's also Scala on JS and on LLVM).
> F# has units of measure
I dearly miss those. Type Providers even more so.
> OCaml has polymorphic variants
Could be nice in Scala. Or maybe not, can't tell now. And there's always the risk of too many features, which makes a language worse, not better.
> It was not designed as a functional programming language first and foremost
This is misleading at best. It was designed to show that the distinction between OOP and FP is arbitrary, not fundamental. In other words, Scala erases the distinction, it's fully OOP and fully FP. Scala has higher order functions, ADTs, pattern matching, a module system, like SML/OCaml. But it uses keywords `class` and calls it's instances "object" like Java. And it's interoperable with Java. But theoretically there's no fundamental distinction.
More languages are becoming more and more like Scala, btw.
https://www.lihaoyi.com/post/FromFirstPrinciplesWhyScala.htm...
And the concept of a programming language paradigm is fake from the start anyway ;)
https://www.cambridgeblog.org/2017/05/what-if-anything-is-a-...