Settle for good on a sequence type inference model

[cstuartroe]

Currently, sequences with no expected type use the first element to set the expected type for the rest of the elements. Is that the best way to do it?

The main alternatives are doing union/greatest common ancestor types.