For context, I tried implementing this and it didn't work out well due to the way "fragment transformers" are currently used. We should get rid of those and then implement this.

Is this still the case?

Sadly, yes. Most of the AST transformation stuff should be ripped out and replaced with a fairly typical lexical scope mechanism. That would make elaboration several times faster, make local clock domains and module flattening actually reliable, and avoid needlessly creating Assign statements that are then deleted and create spurious warnings.

That is a lot of work though. :/ We could probably also add a workaround that ignores Assigns created by nMigen itself by looking at the path or something.

Okay, @BracketMaster relays it from the folks at Chip11 that this issue was the biggest pain point for them while learning. I think that's a solid indication that I should just add a workaround without waiting for the massive refactor to happen. Let's fix this in 0.4.