3.5 regression: return type forced more restrictive than needed

<!-- 🚨 STOP 🚨 𝗦𝗧𝗢𝗣 🚨 𝑺𝑻𝑶𝑷 🚨 Half of all issues filed here are duplicates, answered in the FAQ, or not appropriate for the bug tracker. Even if you think you've found a *bug*, please read the FAQ first, especially the Common "Bugs" That Aren't Bugs section! Please help us by doing the following steps before logging an issue: * Search: https://github.com/Microsoft/TypeScript/search?type=Issues * Read the FAQ: https://github.com/Microsoft/TypeScript/wiki/FAQ Please fill in the *entire* template below. --> <!-- Please try to reproduce the issue with `typescript@next`. It may have already been fixed. --> **TypeScript Version:** 3.5.0-dev.20190508 <!-- Search terms you tried before logging this (so others can find this issue more easily) --> **Search Terms:** restrictive return type **Code** ```ts class Box<A> { constructor(public val: A) {} } interface ContainerFor<A> { array: Array<A>; box: Box<A>; } function single<T extends keyof ContainerFor<number>>(t: T, x: number): ContainerFor<number>[T] { switch (t) { case "array": return new Array<number>(x); case "box": return new Box<number>(x); } return new Box<number>(0); // this should not be needed } function test() { let t1 = single("array", 42); let t2 = single("box", 42); //let t3 = single("boo", 42); } ``` **Expected behavior:** This code worked with 3.4, but had an unsoundness bug in that `single` was allowed to return the wrong type (swap array and box). It was clear that switching over `t` did not increase knowledge about the expected output type and I was forced to have a fallback return which is of course unreachable. But the code itself worked. The `test` function also shows things work as expected. The types for `t1` and `t2` are inferred correctly and asking for an unknown thing ("boo") is reported as an error. I would expect this to still work in 3.5, perhaps fixing one of the issues I'm having in the internals of `si…