re: type inference
Inferential type systems fall into a fuzzy zone between statically and dynamically typed. For example, many JITs for dynamically typed languages do type inference, and generate compiled code optimized to the types actually being used. That's a very obviously good idea. And many linters look at inferred types as well to do compile-time checks. That's also an obviously good idea. But I wouldn't call a JavaScript JIT a statically-typed system (or Python code that's been through Pylint).
re: Apples versus oranges
They're not unitless. In the first case, the units are apples. In the second case, they're oranges. You shouldn't compare or add those types. If I do have a type system, that's something I should be able to specify.
Not to mention implicit conversions, if I want them (12 inches + 1 foot = 24 inches).
Take a look at the Haskell units library. That can do the things you illustrate.
The second example with count of apples and oranges is bad. Those are unitless and a thus Just an natural should be fine for them. Numerical code can get messy but that's exactly were the power of static typing is the best. You can have dimensions and units in a type system and you get dimensional homogeneity correctness for free.