This makes sense to be conceptually. The term "sibling" alerts me a little because the macro context is closer to a child than a sibling. I think the "main" context (we still don't have a good word for that) should be considered the root, and then the macro and core contexts are its children. I'm not sure how exactly to represent that though.

Yeah, I've been thinking about that too; maybe we could add them to the "main" context (as some new concept of child contexts?) upon creation directly?

I switched to a parent/child relationship between contexts, which seem much cleaner.

It's not included here, but I also ran a modified version with an assert in cache_module that made sure any module dependency to another context was to a child context (had to register macro context as child before checking the modules). All server tests passed without triggering that, so it seems this is not missing some edge case.

I think it doesn't crash for me too at #12846 (comment)
But class/field completion is much slower in this branch than in development for me.
Regression is not in first commit, but in second one: 76b59db
This is because staleContextMaxAge is 0 by default.

Right.. fixed that. Thanks!

@Simn anything against merging this?

I don't like the magic 600 appearing in two places, but other than that this looks good.

Yeah.. but they're also 5 lines apart in a simple define/reset pattern.
I could define a default value and use that for both current 600 if you prefer, but not sure it's needed here

I acknowledge that it seems pedantic, but it's the kind of thing that tends to cause surprises down the road. So yes let's make an effort to only have magic numbers in singular places!