fix: under concurrency pressure, oldVNode may be empty, and preact may crash. Fix for that scenario

[joel-solymosi]

This is a highly tricky concurrency issue. Under some mix of preact signals, and preact suspense, oldVNode._children can evaluate to null, leading to a full preact page crash. I have limited time in digging deeper for a repro, but attached PR fixes the issue.

[github-actions]

📊 Tachometer Benchmark Results

Summary

A summary of the benchmark results will show here once they finish.

Results

The full results of your benchmarks will show here once they finish.

_{tachometer-reporter-action v2 for CI}

[JoviDeCroock]

We should have a test for this, I actually think an issue like this could also be related to our changing of the raf timing in signals/hooks. The main reason I want a test is to see whether this is the actual root-cause rather than a band-aid, there is little concurrency going on in Preact as the diff happens synchronously.

It's entirely possible that we somehow succeed at initiating a state-update while a component is suspended which, when the component is suspended upon mount, could lead to this. I've admittedly seen this issue occur once as well d55f62d and it started somewhere in the 10.26 line so my assumptions are the change in RAF timing or this one.

Another thought of something that got introduced in the 10.26.x release is that top-level Fragments are now cloned in the diff, which if the Suspense boundary captures this node could actually lead to this issue. as we'd rely on this being a diff and erroneously assume that this boundary has _children defined. This is in all scenarios prevent by this line but maybe there's a way to bypass it. It recursively removes the _vnodeId (_original) property of all of the suspended vnodes so it should never bail for the _original case which kind of disproves the #4724 assertion but maybe as you say there's a way to hit this case in a concurrent scenario.

I don't mean to be difficult about this, I would just love to get to the root of this problem and if you have a good way of reproducing then that is a great opportunity for that. I want to avoid us defensively setting _children and then realising that the DOM is clobbered because of it, or it duplicates vnodes.