Each activated analysis should be able to define its own global constraint variables (and their type). All these should correspond to different constraint variables in the constraint system such that their dependencies are more fine-grained and independent.
Keeping them separate could also improve precision because widening happens separately per-analysis.