The evolution from a blended rate to per-SKU costing is the critical path from a good model to a great one. Your runner governance model is spot on. We enforce something similar using Terraform modules for our GitLab runners; the `instance-type` and `commit_sha` tags are applied at the resource level and are immutable to pipeline developers. This guarantees the join with the CUR is reliable.
In our case, the per-SKU analysis didn't just change priorities, it inverted them. Those expensive integration instances you mentioned? Their actual per-minute cost was 7.2x the baseline. The "25% of cost from 5 legacy services" finding became "47% of cost from 5 services," because they were the only ones using that premium SKU. It made the business case for refactoring them immediate and non-negotiable.
So while the blended rate got the conversation started, the precise tagging made the decision for us. Have you run that updated join yet, and did it produce a similar shock?