I think this diagnoses the problem well, but ignores an obvious solution.

A team of one data scientist and one engineer, completely responsible for building a model, and seeing it through into production, meeting all applicable SLAs and performance metrics.

Or maybe it's two data scientists and one engineer, or one scientist and two engineers, whatever is required.

The point is to have a small team you can hold completely accountable for their output. They sink or swim together, so there is no debating whether the scientists or engineers get the credit or take the blame. They are assessed by the effectiveness of the end product they produce.