Make expand-selection smarter

Suppose we have this:
(|...| delimits the selection)

let r = test(0, 2).saturating_|a|dd(3).saturating_sub(4);

This is what RA expand selection currently does:

let r = test(1, 2).|saturating_add|(3).saturating_sub(4);
let r = |test(1, 2).saturating_add|(3).saturating_sub(4);
let r = |test(1, 2).saturating_add(3)|.saturating_sub(4);
let r = |test(1, 2).saturating_add(3).saturating_sub|(4);
let r = |test(1, 2).saturating_add(3).saturating_sub(4)|;

That's not very useful.
This is what I would expect:

let r = test(1, 2).|saturating_add|(3).saturating_sub(4);
let r = test(1, 2).|saturating_add(3)|.saturating_sub(4);
let r = test(1, 2).|saturating_add(3).saturating_sub|(4);
let r = test(1, 2).|saturating_add(3).saturating_sub(4)|;
let r = test|(1, 2).saturating_add(3).saturating_sub(4)|;

// Or
let r = test(1, 2).|saturating_add|(3).saturating_sub(4);
let r = test(1, 2).|saturating_add(3)|.saturating_sub(4);
let r = test(1, 2).|saturating_add(3).saturating_sub(4)|;
let r = |test(1, 2).saturating_add(3).saturating_sub(4)|;

RA (as many TreeSitter based editors) follows strictly the grammar structure (left-associativity in the example above) which is easy to implement but often expands the selection the "wrong" way.
IntelliJ does a better job expanding the section for Kotlin, it's very smart and in the example above it goes to the right. (If I recall correctly)

As of today this kind of "blind" expand-selection is of little use to me, I wonder what other people here think.

Locate rust-analyzer's expand-selection entry point and any existing tests, then compare their behavior with the chained-call examples in the issue. The work is done when selection expansion follows useful expression boundaries rather than only grammar left-associativity, with tests covering the expected sequences.