We already strengthen the type of the receiver at a call:

  a.foo();

After the call, a can be restricted to a those types which respond to 'foo'.

We can do the same for arguments.

Consider:

import 'dart:math';

get never => new Random().nextDouble() > 2;

main() {
  var a = 1;
  var b = 2;
  if (never) a = 'a'; // inferred type {int, String}
  if (never) b = 'b';

  var m = min(a, b); // min throws if either argument is not a number, returned value, m is number.
  var s = m + a;
  print('m = $m s = $s');
}

The generated code:

  main: function() {
    var a, m, line;
    a = C.C__JSRandom.nextDouble$0() > 2 ? "a" : 1;
    m = P.min(a, C.C__JSRandom.nextDouble$0() > 2 ? "b" : 2);
    if (typeof a !== "number") // argument check
      return H.iae(a);
    line = "m = " + m + " s = " + (m + a);
    H.printString(line);
  }

We could avoid the argument check if the model for 'min' included the strengthened argument types.

This might be an expensive suggestion since the analysis would need fresh constraint variables for argument values after each call (but only those that are used later).