Express follow_recommendations in terms of account_summaries rather than
accounts, integrate filters that are unconditionally used, and materialize
the resulting view.
This should result in the bulk of the computation being performed only once
instead of **once per recommendation language**.
