3.5.0 async uncancelable changes are not enforced by laws

[neko-kai]

In 3.5.0 there was a breaking semantic change to async/async_:

As of 3.5.0, the following is now considered to be uncancelable:

IO {
  // ...
  None    // we aren't returning a finalizer
  }
}

Previously, the above was cancelable without any caveats. Notably, this applies to all uses of the async_ constructor!

There are no new laws to ensure this behavior, a build updating to 3.5.1 passes without having to add uncancelability to the implemenation of async.

I think this may be a potential footgun for any third-party implementors of CE typeclasses if the laws don't specify this behavior.

[djspiewak]

Oh this is a really good catch. We should add a new law here. Unfortunately, the obvious law (async(_ => pure(None)) <-> uncancelable(_ => never)), while accurate, won't actually catch the failure. There are two options here.

The unambitious and ugly option is to add a law which just, makes a fiber, fusses around with latches to line things up, and then races some cancels. It'll probably mostly work. It'll suck.

The more ambitious option would be to enhance the Eq[IO[A]]. Specifically, if you get into a situation where both the LHS and RHS are deadlocked, rather than assuming they're equivalent, try canceling both. Then run forward some more… then see if they're still equivalent. That would be sufficient (I think) to illustrate the change here, and it moves the Eq more in line with how people conceptualize equivalence, I think.

[armanbilge]

The more ambitious option would be to enhance the Eq[IO[A]].

Note, this doesn't actually address the issue, because third-party effect types won't use this equivalence. Basically this would be kicking the can down the curb, since we are more strictly specifying what equivalence means in the context of laws.