Dealing with re-entrant updates about fusion

[dphfox]

The new execution model should help solve for this. Invalidation is a completely controlled process; no state objects are invoked during that stage. Eager objects are only revalidated after all invalidation is complete.

Revalidation is a little more interesting. It's conceivable that, during revalidation, an object can cause state objects to become invalidated. In particular, an object might invalidate itself, or it might be transitively invalidated if it invalidates a state object it transitively depends on.

For example:

local foo = scope:Value(2)

local bar = scope:Computed(function(use)
   return use(foo) * 2
end)

-- imagine this ran synchronously
local baz = scope:Observer(bar):onBind(function()
    foo:set(peek(foo) + 1)
end)

The specific question we should ask here is: what should happen when a state object becomes invalid during revalidation?. Seemingly, the answer would be "revalidate itself again", but this could easily lead to infinite loops. Really, what we're dealing with here is a "hidden cycle" in the reactive graph.

So, I would propose we add an error for this case:

Detected an infinite loop. Consider adding an explicit breakpoint to your code to prevent a cyclic dependency.

Notably, this can happen during invalidation too, if the graph has an explicitly-modelled cycle in it. However, we should be careful - unlike revalidation, objects being invalid is not itself an invalid state for the invalidation process, because objects can be invalidated multiple times by design. We'd likely need to track state inside the invalidation process to catch this.

However - if you noticed the 'synchronous' stipulation above - this would only really catch errors in the implementations of state objects. We'd need something more sophisticated to solve for user-introduced loops from asynchronous sources like Observers.

Overall, I think that tracking infinite loops during invalidation is a bit of a distraction, since the process can be implemented in a way that trivially supports cyclic graphs by merely terminating when encountering an already-transitively-invalidated graph object, which the invalidation process is marking out anyway. So revalidation is the only part I think we need to care about, which is good because it makes all the scary questions go away.

[dphfox]

Solved in #323