Thank you for adding the regression test! Looks like it's asserting the checked legs, so we actually found something interesting!

Hmm, that is weird, I believe I tested the test locally in debug mode and I didn't see any assertions although this one looks deterministic. I'll take a first look to get more info regarding how we should proceed here. Thanks Michal for reviewing!

I have managed to repro the failure locally - it was a bit tricky as it only reproes when tiered compilation is turned off. According to my understanding the problem is that CanCastTo uses multiple code paths some of which (most notably the Array-related code paths) require cooperative GC mode; I have fixed this by changing ResolveVirtualStaticMethod to use the targeted version CanCastToInterface which doesn't require the GC cooperative mode.

I have blocked out the failing Mono version of the test and the PR pipeline is green now. @davidwrighton, could you please review the runtime change to double-check that it's functionally equivalent?

Considering David is on vacation this week I'm inclined to merge this in before today Preview 7 fork as it fixes an important CoreCLR runtime bug; I'll address any feedback David might have in a follow-up change once he's back. Assuming there's no pushback I'll merge this in after 2PM today.